title
stringlengths 14
154
| paper_url
stringlengths 42
42
| authors
sequencelengths 1
21
| type
stringclasses 3
values | abstract
stringlengths 413
2.52k
| keywords
stringlengths 4
397
| TL;DR
stringlengths 5
250
⌀ | submission_number
int64 2
14.3k
| arxiv_id
stringlengths 10
10
⌀ | embedding
sequencelengths 384
384
| github_url
stringlengths 0
126
| github_stars
int64 0
55k
| num_models
int64 0
82
| num_datasets
int64 0
7
| num_spaces
int64 0
100
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Artificial Kuramoto Oscillatory Neurons | https://openreview.net/forum?id=nwDRD4AMoN | [
"Takeru Miyato",
"Sindy Löwe",
"Andreas Geiger",
"Max Welling"
] | Oral | It has long been known in both neuroscience and AI that ``binding'' between neurons leads to a form of competitive learning where representations are compressed in order to represent more abstract concepts in deeper layers of the network. More recently, it was also hypothesized that dynamic (spatiotemporal) representations play an important role in both neuroscience and AI. Building on these ideas, we introduce Artificial Kuramoto Oscillatory Neurons (*AKOrN*) as a dynamical alternative to threshold units, which can be combined with arbitrary connectivity designs such as fully connected, convolutional, or attentive mechanisms. Our generalized Kuramoto updates bind neurons together through their synchronization dynamics. We show that this idea provides performance improvements across a wide spectrum of tasks such as unsupervised object discovery, adversarial robustness, calibrated uncertainty quantification, and reasoning. We believe that these empirical results show the importance of rethinking our assumptions at the most basic neuronal level of neural representation, and in particular show the importance of dynamical representations. | Oscillatory neurons, Feature binding, Object-centric learning, Reasoning, Adversarial robustness | Oscillatory neurons strongly bind object features, can reason, and are robust to adversarial and natural perturbations | 923 | 2410.13821 | [
-0.08344820141792297,
-0.15643464028835297,
0.053497038781642914,
0.035306066274642944,
-0.0779576525092125,
0.005842019338160753,
0.0072928136214613914,
-0.09602993726730347,
0.06897777318954468,
-0.05165919288992882,
-0.03817019611597061,
-0.07478953897953033,
-0.042372848838567734,
0.017878927290439606,
-0.0717426985502243,
0.04281820356845856,
0.08423939347267151,
0.10536587238311768,
-0.1016731709241867,
-0.0348779559135437,
0.0005104854353703558,
-0.07161498814821243,
-0.010077201761305332,
0.01062325295060873,
0.039412759244441986,
0.10987573862075806,
-0.012035330757498741,
-0.06452777981758118,
0.07821116596460342,
-0.0206192247569561,
0.04679189994931221,
-0.1151597648859024,
0.027754174545407295,
-0.04014981910586357,
-0.0723891630768776,
0.05064518749713898,
-0.04822646453976631,
-0.01588226482272148,
0.023126212880015373,
-0.045649170875549316,
0.07506495714187622,
0.005691458471119404,
-0.027694519609212875,
0.07534778118133545,
0.04159003123641014,
0.0802508145570755,
0.023448778316378593,
-0.02437691204249859,
-0.06274452060461044,
-0.06506723165512085,
-0.06313106417655945,
-0.052776020020246506,
-0.003856493392959237,
0.03030036948621273,
0.0075271655805408955,
0.0661967471241951,
-0.04235748201608658,
0.06345794349908829,
0.011577746830880642,
0.014312751591205597,
0.06059360131621361,
-0.006607667077332735,
0.07844256609678268,
0.0004341381136327982,
0.04592231661081314,
-0.011096695438027382,
-0.0013188050361350179,
0.057943422347307205,
0.009149602614343166,
0.053319480270147324,
0.12004128843545914,
0.10830013453960419,
0.01177211944013834,
0.04425320401787758,
0.07109016180038452,
0.015524810180068016,
-0.02566833607852459,
-0.03135513886809349,
0.016124485060572624,
-0.04960033297538757,
0.06002366915345192,
-0.021477963775396347,
-0.0006834620726294816,
0.020817380398511887,
0.10974938422441483,
-0.02944740280508995,
-0.0324041023850441,
0.059825196862220764,
0.03829643130302429,
-0.023509856313467026,
-0.01759585551917553,
-0.033781327307224274,
-0.020460177212953568,
-0.05831087753176689,
0.04170633852481842,
0.014959259890019894,
0.012751389294862747,
-0.0016465391963720322,
-0.017644234001636505,
0.05495845153927803,
-0.056369535624980927,
0.0676279291510582,
-0.015954691916704178,
0.010127055458724499,
0.02355136163532734,
-0.030963845551013947,
0.07340636849403381,
0.006088804919272661,
-0.010215259157121181,
0.01352827437222004,
0.02041483111679554,
-0.006776116322726011,
0.03336786851286888,
0.0822451040148735,
0.04635763540863991,
0.0014846617123112082,
0.00856840331107378,
0.06621038168668747,
0.05871449038386345,
0.036730214953422546,
0.02012615092098713,
-0.04951299726963043,
-0.0955454409122467,
0.024361617863178253,
0.006935251876711845,
-0.021102499216794968,
-0.1016496941447258,
2.819360752309869e-33,
0.043227408081293106,
-0.02532799169421196,
0.06272876262664795,
-0.06685701012611389,
0.0030558863654732704,
-0.07989609241485596,
-0.07991431653499603,
-0.037290316075086594,
-0.023796681314706802,
0.05550440028309822,
-0.08532536029815674,
0.03752897307276726,
-0.04447315260767937,
0.10994013398885727,
0.08980754762887955,
-0.02681361511349678,
-0.04654982313513756,
-0.047792546451091766,
0.07051380723714828,
-0.10709557682275772,
0.052850205451250076,
-0.02090686932206154,
-0.057775095105171204,
-0.019934386014938354,
-0.052944865077733994,
0.05052059888839722,
0.03824238106608391,
-0.0017692651599645615,
-0.044989921152591705,
0.017050039023160934,
-0.08468092978000641,
0.028421703726053238,
-0.05649496614933014,
0.026272231712937355,
-0.04425101354718208,
-0.06285670399665833,
-0.02455752156674862,
-0.042906418442726135,
-0.0638136938214302,
0.0076119801960885525,
0.018572719767689705,
-0.03560522571206093,
-0.06280070543289185,
0.002404163358733058,
-0.02162078022956848,
-0.0035788503009825945,
0.04599551483988762,
0.03318227827548981,
-0.051786378026008606,
-0.06844484061002731,
-0.07329880446195602,
-0.010137383826076984,
0.00012500665616244078,
-0.08228116482496262,
0.02844766527414322,
0.03897855803370476,
0.01656544767320156,
0.023616909980773926,
-0.023701557889580727,
0.03265213966369629,
-0.03143998980522156,
-0.035038143396377563,
0.06611334532499313,
0.03764549642801285,
0.008344014175236225,
0.10873688012361526,
-0.005349040497094393,
0.03837955370545387,
0.017368406057357788,
0.012608055025339127,
-0.018264412879943848,
0.09408227354288101,
-0.07386896759271622,
-0.07493078708648682,
0.0005414146580733359,
0.020812131464481354,
0.062474410980939865,
-0.034358151257038116,
-0.11596617102622986,
-0.015327082946896553,
0.019321395084261894,
-0.14950011670589447,
-0.0870184376835823,
0.09422340244054794,
-0.0008583164308220148,
-0.002214316977187991,
0.10212626308202744,
0.0366540364921093,
-0.09639111161231995,
0.03298506513237953,
-0.04109521582722664,
-0.05273625999689102,
0.034232158213853836,
0.0016154665499925613,
-0.05967399850487709,
-4.0716813731400704e-33,
0.009877754375338554,
0.017596960067749023,
0.019390525296330452,
0.041908953338861465,
0.008078940212726593,
0.04784560576081276,
-0.08709736168384552,
0.041417207568883896,
-0.0806950032711029,
-0.01778382994234562,
-0.05770073086023331,
-0.01355752069503069,
-0.008857657201588154,
-0.008358881808817387,
-0.02597765065729618,
0.03388745337724686,
-0.0192270427942276,
0.012848795391619205,
0.00022673877538181841,
-0.059052079916000366,
-0.025844687595963478,
0.0865941122174263,
-0.09310638159513474,
-0.001131435506977141,
0.004473552573472261,
0.08083920180797577,
-0.003865861799567938,
0.03490552306175232,
-0.06518691033124924,
0.08735679090023041,
-0.06629244983196259,
-0.015104974620044231,
-0.06628480553627014,
-0.08213707059621811,
-0.017814137041568756,
0.11118673533201218,
0.059982284903526306,
-0.0495726652443409,
0.0012091025710105896,
-0.03836462274193764,
0.023056400939822197,
0.010988006368279457,
-0.021360183134675026,
0.0023938945960253477,
0.045128822326660156,
-0.02104325033724308,
-0.02066439762711525,
0.015060442499816418,
-0.07501228153705597,
0.009946141391992569,
0.07147134840488434,
0.053440991789102554,
-0.07380762696266174,
-0.04162876680493355,
-0.0627509206533432,
0.06951267272233963,
-0.03062421642243862,
-0.00774779450148344,
0.07785912603139877,
0.03451245278120041,
-0.060229986906051636,
-0.14787420630455017,
0.033019501715898514,
-0.015902239829301834,
-0.055423714220523834,
0.0011408586287871003,
0.033795129507780075,
0.048101380467414856,
0.04410269111394882,
0.0569596029818058,
-0.039485570043325424,
-0.03592846170067787,
0.05387762933969498,
0.038120515644550323,
0.000559446751140058,
-0.015164812095463276,
-0.005415117833763361,
-0.013026823289692402,
0.021655021235346794,
0.03307351469993591,
-0.012379529885947704,
-0.02345995418727398,
0.009467385709285736,
0.026748914271593094,
0.06334846466779709,
0.08522040396928787,
-0.030418071895837784,
0.023594073951244354,
0.09264862537384033,
0.04854847490787506,
0.023487722501158714,
0.04092063382267952,
-0.042196016758680344,
0.06796915829181671,
-0.04739665240049362,
-4.8652470496790556e-8,
-0.029108183458447456,
0.00610083295032382,
0.030183641240000725,
0.005189701449126005,
0.10447954386472702,
-0.013929320499300957,
-0.01658073253929615,
-0.03009023331105709,
-0.06511076539754868,
-0.059497103095054626,
0.11895512789487839,
-0.0320453979074955,
-0.0032291447278112173,
0.005519373808056116,
0.024427689611911774,
0.05309924855828285,
0.02497515268623829,
-0.02805924229323864,
0.02602527104318142,
-0.030935442075133324,
0.05280466377735138,
-0.01596660353243351,
0.0013357760617509484,
-0.01997217908501625,
0.04848852381110191,
-0.04408543184399605,
-0.09217576682567596,
0.04607801139354706,
-0.015546835958957672,
0.05504336580634117,
-0.0011481205001473427,
0.050490230321884155,
0.0713423416018486,
-0.008373035117983818,
-0.03796772286295891,
0.08112393319606781,
-0.028032127767801285,
-0.05347567796707153,
-0.07289907336235046,
-0.038048937916755676,
0.014647545292973518,
0.025122683495283127,
-0.014599582180380821,
0.05361252278089523,
-0.008060378953814507,
0.041286010295152664,
0.05596073716878891,
-0.05685724318027496,
0.019296100363135338,
-0.011036855168640614,
-0.015843549743294716,
0.006567091215401888,
0.0028499839827418327,
0.04614702984690666,
-0.005923747085034847,
0.022345855832099915,
-0.0033925268799066544,
-0.047887999564409256,
-0.04968804866075516,
0.015932146459817886,
-0.024448761716485023,
0.042826101183891296,
-0.07397086918354034,
-0.04167835786938667
] | https://github.com/autonomousvision/akorn | 77 | 0 | 0 | 0 |
Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation | https://openreview.net/forum?id=tTPHgb0EtV | [
"Tiansheng Huang",
"Sihao Hu",
"Fatih Ilhan",
"Selim Furkan Tekin",
"Ling Liu"
] | Oral | Harmful fine-tuning attack poses serious safety concerns for large language models' fine-tuning-as-a-service. While existing defenses have been proposed to mitigate the issue, their performances are still far away from satisfactory, and the root cause of the problem has not been fully recovered. To this end, we in this paper show that \textit{harmful perturbation} over the model weights could be a probable cause of alignment-broken. In order to attenuate the negative impact of harmful perturbation, we propose an alignment-stage solution, dubbed Booster. Technically, along with the original alignment loss, we append a loss regularizer in the alignment stage's optimization. The regularizer ensures that the model's harmful loss reduction after the simulated harmful perturbation is attenuated, thereby mitigating the subsequent fine-tuning risk. Empirical results show that Booster can effectively reduce the harmful score of the fine-tuned models while maintaining the performance of downstream tasks. Our code is available at https://github.com/git-disl/Booster | Harmful fine-tuning, LLM, safety alignment | This paper proposes Booster, an alignment stage solution against harmful fine-tuning issues for LLMs | 680 | 2409.01586 | [
-0.09731985628604889,
-0.0226671751588583,
0.011634123511612415,
-0.03275473043322563,
0.004676955286413431,
0.0355997234582901,
-0.010103822685778141,
-0.003660190384835005,
0.044295184314250946,
-0.034605782479047775,
-0.027282699942588806,
0.008895333856344223,
0.07892608642578125,
-0.019769098609685898,
-0.020148947834968567,
-0.01973012089729309,
0.03118765726685524,
0.1267128586769104,
-0.13852789998054504,
-0.03334718942642212,
-0.01717374287545681,
0.04239119589328766,
-0.019721616059541702,
0.04825912415981293,
-0.030058130621910095,
0.018119243904948235,
-0.07749585807323456,
-0.03374981880187988,
0.02194194868206978,
-0.09621268510818481,
-0.0009479414438828826,
-0.05666248872876167,
-0.006684564054012299,
0.06732872128486633,
-0.049293987452983856,
0.0837048590183258,
-0.04636022448539734,
-0.050837934017181396,
0.009073414839804173,
-0.04982490465044975,
-0.015473276376724243,
0.025957779958844185,
-0.0023334340658038855,
0.015510081313550472,
0.06928690522909164,
-0.03454717993736267,
0.008156688883900642,
-0.031649526208639145,
0.0029320609755814075,
-0.04122422635555267,
-0.1250881403684616,
-0.047048553824424744,
0.05054940655827522,
0.05155232176184654,
0.007208484690636396,
0.0015370995970442891,
0.05062243714928627,
0.012488250620663166,
-0.04064599797129631,
0.00813691783696413,
-0.021213198080658913,
-0.045987050980329514,
-0.03093738667666912,
-0.004560656845569611,
-0.040507204830646515,
-0.04958551749587059,
0.05414692312479019,
0.05816023424267769,
-0.030116567388176918,
0.1307428628206253,
-0.06413006782531738,
0.06309816241264343,
0.025995446369051933,
0.06396285444498062,
0.04541265591979027,
0.11624513566493988,
-0.0035994923673570156,
-0.010496866889297962,
0.08414924144744873,
-0.05295443534851074,
0.042982958257198334,
-0.023831326514482498,
0.0778331607580185,
-0.02976943366229534,
0.08694041520357132,
0.002007133327424526,
0.003879502648487687,
-0.04958968609571457,
0.09259861707687378,
0.02634357288479805,
-0.045577552169561386,
-0.05050366744399071,
0.09092524647712708,
0.07371377944946289,
-0.007353521417826414,
0.059591781347990036,
0.007064082194119692,
-0.08978493511676788,
-0.018618356436491013,
0.08834978938102722,
0.03300844877958298,
0.0552276074886322,
0.047744184732437134,
-0.010622254572808743,
0.00891395565122366,
-0.09097263216972351,
0.00006525853677885607,
0.03540535271167755,
-0.033646512776613235,
-0.09634153544902802,
0.046142224222421646,
0.02577563002705574,
0.0023905006237328053,
0.04903696849942207,
-0.023431623354554176,
0.017652172595262527,
-0.004210328217595816,
-0.06645224988460541,
-0.02361513115465641,
0.004336329642683268,
-0.04180992767214775,
-0.010933485813438892,
-0.027730632573366165,
0.018260221928358078,
0.011826532892882824,
0.014885363169014454,
-0.020096467807888985,
2.8797657335787802e-33,
0.044443417340517044,
0.1191558688879013,
-0.014602449722588062,
0.0018153614364564419,
-0.033612847328186035,
-0.03956186771392822,
-0.04259595274925232,
0.011010419577360153,
-0.034582529217004776,
-0.01768866367638111,
-0.01347032468765974,
-0.04487825557589531,
0.0010375408455729485,
0.06123489513993263,
0.029958417639136314,
-0.06714127957820892,
-0.018710089847445488,
0.07557789236307144,
-0.07235345244407654,
-0.042182840406894684,
0.060408517718315125,
-0.010009617544710636,
-0.04721871390938759,
-0.09448542445898056,
0.010059146210551262,
-0.003636165987700224,
0.06077302247285843,
-0.08549197018146515,
-0.0964924693107605,
0.017621148377656937,
-0.09959054738283157,
-0.08597683906555176,
0.05591948330402374,
0.043280068784952164,
0.03828088566660881,
-0.0721542164683342,
-0.05522187426686287,
-0.06540296226739883,
-0.006380320526659489,
0.019925348460674286,
0.0011288051027804613,
0.06252720952033997,
0.00047189497854560614,
-0.008761120028793812,
0.044976767152547836,
-0.04905860126018524,
0.008758547715842724,
0.02560305781662464,
-0.0030420389957726,
-0.14381921291351318,
0.03960573300719261,
0.0654696673154831,
0.0017237627180293202,
0.046802062541246414,
-0.0897095575928688,
0.023251518607139587,
0.039469845592975616,
0.00016760925063863397,
0.004392349626868963,
0.08825802057981491,
-0.05122653767466545,
-0.04458703100681305,
-0.010857105255126953,
0.03689631074666977,
0.08960670232772827,
-0.06438553333282471,
0.020068135112524033,
-0.0012091398239135742,
-0.011926711536943913,
0.01896967925131321,
-0.0216993298381567,
-0.06687159836292267,
-0.01699327491223812,
-0.0004017219180241227,
-0.0014616437256336212,
-0.04009934514760971,
0.06939973682165146,
-0.024911371991038322,
0.034678224474191666,
-0.05355712026357651,
0.025182390585541725,
0.027705147862434387,
0.038491301238536835,
-0.0201252531260252,
-0.06420086324214935,
-0.09711167216300964,
0.05476301163434982,
-0.09696362912654877,
-0.026787959039211273,
0.06631753593683243,
-0.004836737178266048,
0.02314777672290802,
0.0021842666901648045,
-0.013994692824780941,
-0.02742714248597622,
-2.2296141344763736e-33,
-0.024405594915151596,
-0.04563524201512337,
-0.027680110186338425,
0.09331229329109192,
-0.04348285496234894,
-0.061551451683044434,
0.028844673186540604,
0.04889276251196861,
-0.052815891802310944,
-0.04239800199866295,
-0.03097912296652794,
-0.05560078099370003,
0.042596880346536636,
0.017559660598635674,
0.04882488772273064,
-0.000657925324048847,
-0.01886824332177639,
0.07155154645442963,
0.044864967465400696,
0.04310647398233414,
0.06554210931062698,
0.07336194813251495,
-0.0663047507405281,
0.0748230442404747,
0.025943730026483536,
-0.0007630786276422441,
-0.05559006333351135,
0.08422741293907166,
0.03534192219376564,
-0.0004076478653587401,
-0.014154009521007538,
0.06368329375982285,
-0.0335511676967144,
-0.030893873423337936,
-0.037844087928533554,
0.04805891215801239,
-0.003262678626924753,
-0.0010563511168584228,
0.04873540624976158,
0.06328043341636658,
0.0734536349773407,
0.024344323202967644,
-0.10012160986661911,
-0.028124352917075157,
0.033943984657526016,
-0.016432248055934906,
-0.07687855511903763,
-0.007738214451819658,
-0.02487468719482422,
-0.01162787340581417,
-0.02237880788743496,
-0.0015855764504522085,
0.005064812954515219,
0.05683493986725807,
-0.035480234771966934,
-0.06622225046157837,
-0.022665906697511673,
-0.04130832850933075,
-0.01774960197508335,
0.03889397904276848,
-0.014913402497768402,
-0.0073838625103235245,
0.05147413909435272,
-0.06597838550806046,
0.05788694694638252,
-0.007971047423779964,
0.022193346172571182,
-0.015664467588067055,
0.05586271360516548,
-0.0168113112449646,
0.07399504631757736,
-0.03155796229839325,
-0.054651420563459396,
0.082061268389225,
-0.06269273161888123,
-0.0533619299530983,
-0.02814493514597416,
-0.03437057510018349,
0.008322212845087051,
0.0038592505734413862,
-0.0019941532518714666,
0.06886965781450272,
0.03506284952163696,
0.015020303428173065,
-0.021887509152293205,
0.04415900632739067,
0.07170867174863815,
0.10882791131734848,
-0.00863558892160654,
-0.010438818484544754,
-0.029359962791204453,
-0.0026016237679868937,
-0.009312334470450878,
0.044359367340803146,
-0.05251770466566086,
-4.820878629629988e-8,
-0.06654396653175354,
0.012901903130114079,
-0.06897418200969696,
0.12289781123399734,
-0.05348311364650726,
-0.0790366530418396,
-0.03391476348042488,
0.02511739917099476,
0.017737288028001785,
-0.06050077825784683,
0.040594086050987244,
-0.029891015961766243,
-0.0641700029373169,
-0.046772271394729614,
-0.04682842269539833,
0.1254066377878189,
-0.046783696860075,
0.051516756415367126,
-0.07372027635574341,
0.042204275727272034,
-0.0248949583619833,
0.11812999844551086,
-0.010219358839094639,
0.018570533022284508,
0.04293527826666832,
-0.010366787202656269,
0.017997656017541885,
0.09158418327569962,
0.12077021598815918,
-0.027073604986071587,
-0.012703999876976013,
-0.015410167165100574,
-0.006956080906093121,
-0.00837323721498251,
-0.022005630657076836,
0.16445349156856537,
-0.03658482804894447,
-0.030465222895145416,
0.047484368085861206,
0.09345711767673492,
0.08609554171562195,
0.08909109979867935,
-0.028963712975382805,
0.025481682270765305,
0.028508415445685387,
-0.019546596333384514,
-0.06978317350149155,
-0.0682484358549118,
0.04130946472287178,
-0.1026773527264595,
0.05494339391589165,
-0.014235133305191994,
0.00955287553369999,
0.03244508057832718,
-0.025989552959799767,
0.005946711171418428,
-0.0034904200583696365,
-0.03402913361787796,
0.06731800734996796,
0.10490227490663528,
0.038330044597387314,
-0.0438864529132843,
0.031875815242528915,
-0.022772083058953285
] | https://github.com/git-disl/booster | 25 | 0 | 0 | 0 |
Unlearning-based Neural Interpretations | https://openreview.net/forum?id=PBjCTeDL6o | [
"Ching Lam Choi",
"Alexandre Duplessis",
"Serge Belongie"
] | Oral | Gradient-based interpretations often require an anchor point of comparison to avoid saturation in computing feature importance. We show that current baselines defined using static functions—constant mapping, averaging or blurring—inject harmful colour, texture or frequency assumptions that deviate from model behaviour. This leads to accumulation of irregular gradients, resulting in attribution maps that are biased, fragile and manipulable. Departing from the static approach, we propose $\texttt{UNI}$ to compute an (un)learnable, debiased and adaptive baseline by perturbing the input towards an $\textit{unlearning direction}$ of steepest ascent. Our method discovers reliable baselines and succeeds in erasing salient features, which in turn locally smooths the high-curvature decision boundaries. Our analyses point to unlearning as a promising avenue for generating faithful, efficient and robust interpretations. | Explainability, Attribution, Debiasing, Bias | UNI computes a debiased, adaptive baseline for gradient-based interpretations by perturbing the input towards an unlearning direction of steepest ascent. | 604 | 2410.08069 | [
-0.015766695141792297,
-0.08597414195537567,
0.046696778386831284,
0.04580254480242729,
0.06393763422966003,
0.021412337198853493,
-0.038466934114694595,
-0.04367058724164963,
-0.023338770493865013,
-0.026838282123208046,
-0.08837014436721802,
-0.017809737473726273,
0.03874638304114342,
0.04651135578751564,
-0.08399657905101776,
0.07123594731092453,
0.057938672602176666,
0.14886784553527832,
-0.1064600944519043,
-0.036547742784023285,
-0.0027212374843657017,
0.007718597538769245,
-0.017269305884838104,
0.013785597868263721,
0.0726412832736969,
-0.007795161567628384,
-0.014034398831427097,
0.020084112882614136,
0.0876423567533493,
-0.06948939710855484,
0.02083050273358822,
-0.0018492980161681771,
-0.011121533811092377,
-0.020337622612714767,
-0.02101350948214531,
0.06696397066116333,
-0.006553320679813623,
0.00000269994666268758,
0.029560793191194534,
-0.016930146142840385,
0.03388076275587082,
0.047208476811647415,
-0.03795619681477547,
0.053752101957798004,
0.12730509042739868,
-0.006710042245686054,
0.06338796764612198,
-0.04013131558895111,
-0.04071087762713432,
0.03617444261908531,
0.012025127187371254,
-0.0340447835624218,
-0.028943317010998726,
-0.04149729385972023,
-0.03202473744750023,
0.08218341320753098,
0.00829994771629572,
0.0029375737067312002,
0.022949643433094025,
-0.014086630195379257,
-0.023709267377853394,
-0.07953645288944244,
0.01624881848692894,
0.02092328667640686,
0.02990553341805935,
-0.012749459594488144,
0.05226469412446022,
0.021712856367230415,
0.027141910046339035,
0.06810200959444046,
-0.0013965519610792398,
0.04606272280216217,
-0.007778891362249851,
-0.07129210233688354,
0.03963407501578331,
0.07995477318763733,
0.0909680426120758,
0.009407117031514645,
0.018179910257458687,
-0.09452598541975021,
-0.020806249231100082,
-0.04160996153950691,
0.023389378562569618,
0.008636042475700378,
0.08771835267543793,
0.03515883907675743,
-0.01221930980682373,
-0.020768439397215843,
-0.0035070038866251707,
0.11454895883798599,
-0.03177458420395851,
-0.05893179401755333,
-0.04591524973511696,
-0.03311067447066307,
-0.019636042416095734,
0.004892259836196899,
0.0024796451907604933,
-0.10814506560564041,
-0.09041707217693329,
0.1355293095111847,
-0.041754208505153656,
0.012617641128599644,
-0.010694247670471668,
-0.019922075793147087,
0.07527457922697067,
-0.094926618039608,
0.08578046411275864,
0.00033365137642249465,
0.03379451110959053,
-0.042850933969020844,
0.05979924276471138,
0.009666504338383675,
-0.011474918574094772,
0.0451730415225029,
-0.021424414590001106,
0.01919165998697281,
0.006989584304392338,
-0.027667777612805367,
0.0027770695742219687,
0.0071799056604504585,
-0.03499819338321686,
-0.016805043444037437,
-0.04316972568631172,
0.07226996123790741,
-0.009395336732268333,
-0.03875244781374931,
-0.025219056755304337,
3.04301049116831e-33,
0.010327283293008804,
0.020289001986384392,
0.05704334005713463,
-0.004676577635109425,
0.04933701455593109,
-0.015772344544529915,
-0.03985683247447014,
-0.023725371807813644,
-0.03775976970791817,
0.11047284305095673,
-0.05465317517518997,
0.06640620529651642,
-0.05814358964562416,
0.12609697878360748,
-0.012237065471708775,
-0.023766377940773964,
-0.008516931906342506,
0.013270116411149502,
-0.07301076501607895,
-0.027503808960318565,
-0.027529435232281685,
-0.022463632747530937,
-0.04749163240194321,
-0.1283530592918396,
-0.06869242340326309,
-0.036502718925476074,
-0.007235967554152012,
0.041538305580616,
-0.08154871314764023,
0.02396046370267868,
-0.09652528911828995,
-0.022879814729094505,
0.0524938590824604,
0.027680309489369392,
-0.013953899033367634,
-0.058270130306482315,
-0.0025508860126137733,
-0.020212609320878983,
0.021877184510231018,
0.08482285588979721,
-0.011856579221785069,
0.0762757807970047,
-0.0020799627527594566,
-0.02451496012508869,
-0.041676051914691925,
0.005178669933229685,
0.0041799056343734264,
0.011287182569503784,
-0.09134138375520706,
0.007489972747862339,
-0.029172832146286964,
0.020619599148631096,
0.02055545710027218,
-0.06733380258083344,
-0.04440997540950775,
-0.010937764309346676,
0.016877107322216034,
-0.0035985622089356184,
0.016844218596816063,
0.04479454085230827,
-0.005219181999564171,
0.06637606769800186,
0.0027568198274821043,
0.04334750771522522,
-0.02885494753718376,
0.08068490028381348,
-0.027486417442560196,
0.06743013858795166,
0.04873066022992134,
0.005073655396699905,
-0.07020558416843414,
0.08037593960762024,
-0.07628161460161209,
-0.10788711160421371,
0.06105174869298935,
0.03479444980621338,
0.04959642142057419,
-0.05880929157137871,
0.1021399274468422,
0.011153889819979668,
0.002294995356351137,
0.010886436328291893,
0.008207166567444801,
-0.059345684945583344,
-0.10736154019832611,
-0.0018057957058772445,
-0.0037102459464222193,
-0.06824701279401779,
-0.03613865375518799,
0.005281549878418446,
-0.10954587161540985,
-0.017985930666327477,
0.05035359039902687,
-0.070860855281353,
0.01870463602244854,
-2.914898137317989e-33,
-0.046212971210479736,
0.04474814981222153,
-0.020210718736052513,
0.02972107008099556,
-0.09107593446969986,
-0.024449104443192482,
-0.06144086644053459,
0.011487212963402271,
-0.05204344540834427,
-0.030422590672969818,
-0.03443986922502518,
0.00562684703618288,
0.014214941300451756,
-0.025617392733693123,
0.027066927403211594,
0.004019453190267086,
-0.05551903694868088,
0.02847888320684433,
0.013987072743475437,
0.028117945417761803,
0.015304659493267536,
0.09258195757865906,
-0.10466674715280533,
0.009471002966165543,
-0.04677261412143707,
0.10141562670469284,
-0.033831119537353516,
0.05120110511779785,
-0.005585318431258202,
-0.05476263538002968,
-0.046252913773059845,
0.002279696287587285,
-0.020275218412280083,
-0.07538674026727676,
-0.06260883063077927,
0.03387165814638138,
0.007590264081954956,
-0.022147150710225105,
-0.06851383298635483,
0.08152683079242706,
0.053620003163814545,
-0.019564464688301086,
-0.031331948935985565,
-0.0300271138548851,
0.012061243876814842,
-0.010007303208112717,
-0.022968482226133347,
0.007382062263786793,
0.06638522446155548,
0.05139874294400215,
-0.012555532157421112,
0.038264378905296326,
-0.06437896192073822,
0.10091545432806015,
-0.086757130920887,
-0.05145629122853279,
-0.049863122403621674,
0.027604877948760986,
0.11454949527978897,
0.005850552581250668,
-0.05398174375295639,
-0.03207129240036011,
-0.08120688050985336,
-0.015873752534389496,
-0.0034809804055839777,
-0.03205530345439911,
0.013908380642533302,
-0.009292603470385075,
0.042495761066675186,
0.010846208781003952,
0.023927096277475357,
-0.02275446616113186,
-0.011902122758328915,
0.001129036652855575,
0.016583720222115517,
-0.015314172022044659,
0.06553009897470474,
-0.038043271750211716,
-0.014472451992332935,
-0.05926109850406647,
0.08917642384767532,
-0.017130739986896515,
0.0710383728146553,
0.07256288081407547,
0.07615920156240463,
0.03641436621546745,
-0.013746259734034538,
-0.010862281545996666,
0.06450573354959488,
0.008236080408096313,
0.0012802050914615393,
0.031937334686517715,
-0.10711705684661865,
0.07441098988056183,
-0.06606503576040268,
-5.1902134146075696e-8,
-0.06159484386444092,
0.019889378920197487,
0.00439648749306798,
0.019383806735277176,
-0.0010334954131394625,
-0.063125379383564,
0.026880184188485146,
0.04344713315367699,
-0.06783216446638107,
-0.03728498890995979,
0.0355847105383873,
0.01209377869963646,
-0.03080197051167488,
-0.10104669630527496,
-0.0011553962249308825,
0.12777310609817505,
0.004733772948384285,
0.07954743504524231,
-0.01178052369505167,
0.015917286276817322,
-0.020417770370841026,
-0.009036652743816376,
0.09174977242946625,
-0.053876277059316635,
0.028567109256982803,
-0.06878746300935745,
-0.07062080502510071,
0.09672187268733978,
0.10801757127046585,
-0.03324243426322937,
0.03891470283269882,
0.03919446840882301,
0.07706636935472488,
-0.011169571429491043,
0.014041776768863201,
0.0938333049416542,
-0.023136727511882782,
-0.008500153198838234,
-0.005737127270549536,
0.0015161925693973899,
-0.0734439343214035,
0.07891014218330383,
0.038406241685152054,
0.0033097006380558014,
-0.03783329203724861,
0.0058396365493535995,
0.10141663253307343,
-0.0740666389465332,
0.011296332813799381,
-0.002240015659481287,
-0.02572205290198326,
-0.016944503411650658,
0.002653290517628193,
0.10080850124359131,
0.04805458337068558,
-0.00746020395308733,
-0.03444318845868111,
-0.031220221891999245,
-0.07853088527917862,
0.10704589635133743,
0.032406363636255264,
0.04622531682252884,
-0.0317964144051075,
-0.09887731075286865
] | 0 | 0 | 0 | 0 |
|
ChartMoE: Mixture of Diversely Aligned Expert Connector for Chart Understanding | https://openreview.net/forum?id=o5TsWTUSeF | [
"Zhengzhuo Xu",
"Bowen Qu",
"Yiyan Qi",
"SiNan Du",
"Chengjin Xu",
"Chun Yuan",
"Jian Guo"
] | Oral | Automatic chart understanding is crucial for content comprehension and document parsing. Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in chart understanding through domain-specific alignment and fine-tuning. However, current MLLMs still struggle to provide faithful data and reliable analysis only based on charts. To address it, we propose ChartMoE, which employs the Mixture of Expert (MoE) architecture to replace the traditional linear projector to bridge the modality gap. Specifically, we train several linear connectors through distinct alignment tasks, which are utilized as the foundational initialization parameters for different experts. Additionally, we introduce ChartMoE-Align, a dataset with nearly 1 million chart-table-JSON-code quadruples to conduct three alignment tasks (chart-table/JSON/code). Combined with the vanilla connector, we initialize different experts diversely and adopt high-quality knowledge learning to further refine the MoE connector and LLM parameters. Extensive experiments demonstrate the effectiveness of the MoE connector and our initialization strategy, e.g., ChartMoE improves the accuracy of the previous state-of-the-art from 80.48% to 84.64% on the ChartQA benchmark. | Multimodal Large Language Models, Chart Reasoning, Mixture of Expert | null | 526 | 2409.03277 | [
-0.010915442369878292,
-0.12346382439136505,
-0.02362985350191593,
-0.026096560060977936,
0.01134492177516222,
-0.02766699343919754,
-0.09694834053516388,
0.03702382743358612,
-0.021678203716874123,
0.026878762990236282,
0.03643794730305672,
-0.030565427616238594,
-0.013458234257996082,
0.09444762766361237,
-0.01753056049346924,
0.047023460268974304,
-0.05003698170185089,
0.029119696468114853,
-0.08507613837718964,
-0.050573911517858505,
-0.0405416265130043,
0.016717579215765,
0.0042077405378222466,
-0.0025316826067864895,
0.06610257923603058,
0.0945369228720665,
0.018337372690439224,
-0.024844124913215637,
0.0449494831264019,
-0.05164697766304016,
0.0014992636861279607,
-0.025872429832816124,
0.027246546000242233,
0.09622769057750702,
0.0019678364042192698,
0.09830264747142792,
0.025630338117480278,
-0.0041898093186318874,
0.03977053612470627,
-0.011621545068919659,
-0.023453205823898315,
-0.00598656153306365,
0.03392625227570534,
-0.05555201321840286,
0.06595830619335175,
0.017180105671286583,
-0.06893883645534515,
-0.06902986764907837,
0.06750360876321793,
0.08684694021940231,
-0.10483238846063614,
-0.013875994831323624,
0.0301813967525959,
0.05310763046145439,
0.01531611941754818,
0.0865008533000946,
-0.006251432467252016,
-0.01748020574450493,
0.06665355712175369,
-0.013103163801133633,
-0.062012627720832825,
-0.019816119223833084,
-0.03238639980554581,
0.018749983981251717,
0.036307185888290405,
-0.015635987743735313,
0.024230705574154854,
0.03586733713746071,
-0.022101718932390213,
0.03031996265053749,
-0.07138491421937943,
0.023377910256385803,
-0.011944146826863289,
0.04013945162296295,
-0.009179897606372833,
0.08797630667686462,
0.03661902993917465,
-0.04012644663453102,
0.009943090379238129,
-0.058692749589681625,
0.021370554342865944,
0.05177571251988411,
0.03020244464278221,
0.07075434923171997,
0.057670630514621735,
-0.030229007825255394,
-0.037914782762527466,
0.021634478121995926,
0.019633738324046135,
0.016612399369478226,
-0.07445316016674042,
-0.07642356306314468,
0.0634060800075531,
0.008318307809531689,
0.0431525856256485,
0.16631574928760529,
0.0789598673582077,
-0.07992289215326309,
-0.04564737156033516,
0.07383272051811218,
-0.012127491645514965,
0.03906051442027092,
0.005473042838275433,
-0.0822107344865799,
-0.03700753673911095,
-0.0922737866640091,
0.00870504043996334,
0.0019357061246410012,
-0.00926067866384983,
-0.10272430628538132,
0.02806219458580017,
0.01946725882589817,
0.02362961322069168,
-0.0014328351244330406,
-0.011038532480597496,
-0.025124244391918182,
-0.0587162971496582,
-0.04156802222132683,
0.05265601724386215,
0.031853895634412766,
-0.04780173674225807,
0.08747588843107224,
-0.036942243576049805,
0.013323421590030193,
0.02711479552090168,
0.006598262581974268,
-0.03788589686155319,
1.1197205417891278e-33,
0.053511444479227066,
0.10033604502677917,
0.08839412033557892,
0.0882878229022026,
0.03378058224916458,
-0.013433870859444141,
-0.08172646909952164,
0.0034389079082757235,
-0.03643675521016121,
0.005732161458581686,
-0.013800480403006077,
0.0356961190700531,
-0.06845303624868393,
0.09104157984256744,
-0.0315965861082077,
0.014438889920711517,
-0.003990821074694395,
0.023168684914708138,
-0.1433718055486679,
0.02388150803744793,
-0.04980449378490448,
-0.026742927730083466,
0.02499104104936123,
0.013784744776785374,
0.023391777649521828,
0.04014410451054573,
0.07390271872282028,
-0.033000681549310684,
-0.09123533964157104,
0.002980930497869849,
-0.13201498985290527,
-0.04743124172091484,
0.008922328241169453,
0.005576078314334154,
0.05863622575998306,
-0.08048602938652039,
-0.03755149617791176,
-0.08597610145807266,
-0.02528991363942623,
0.022383850067853928,
-0.05259506404399872,
0.000953346723690629,
-0.0700707957148552,
-0.006662196479737759,
-0.10560661554336548,
-0.022284341976046562,
0.008823656477034092,
0.00008638168219476938,
0.04410823434591293,
-0.06637345254421234,
-0.004823902156203985,
0.052359335124492645,
-0.04100112244486809,
0.0034610149450600147,
0.010877197608351707,
-0.0345006063580513,
0.026502253487706184,
0.047140415757894516,
0.0795622169971466,
-0.0045205275528132915,
-0.010145291686058044,
0.05043574050068855,
-0.011590821668505669,
0.08056517690420151,
0.060587815940380096,
0.020219335332512856,
-0.07159938663244247,
0.05790010839700699,
0.0345195010304451,
0.036691632121801376,
-0.07488463819026947,
0.040869537740945816,
0.0012023607268929482,
-0.04042154550552368,
0.021116554737091064,
0.038476087152957916,
-0.029605954885482788,
-0.05914141237735748,
0.060999032109975815,
-0.018257303163409233,
-0.006933433935046196,
-0.0037166511174291372,
0.03795783221721649,
-0.07236228138208389,
-0.07677042484283447,
-0.10633809119462967,
0.07678426057100296,
-0.029564516618847847,
-0.00343401194550097,
-0.058382321149110794,
-0.001104707014746964,
-0.0024707880802452564,
-0.009621070697903633,
-0.00001878700641100295,
0.05382861942052841,
-1.0066714133606253e-33,
0.0140322120860219,
0.05348346382379532,
0.03949088975787163,
0.048557791858911514,
-0.00179798633325845,
-0.06572353094816208,
-0.0005701312911696732,
0.04467395320534706,
0.008661537431180477,
0.02315792255103588,
0.022255582734942436,
-0.09997241199016571,
-0.008176569826900959,
-0.04970237612724304,
0.07386364787817001,
0.07564353197813034,
-0.0626840814948082,
0.053805381059646606,
0.027792127802968025,
0.002497252309694886,
-0.013947988860309124,
0.044550031423568726,
-0.06912383437156677,
0.04731643572449684,
0.018111029639840126,
0.029328593984246254,
-0.06050824746489525,
-0.03637504205107689,
-0.0635768324136734,
0.0316024087369442,
-0.018307503312826157,
-0.028970230370759964,
-0.05078926682472229,
-0.09342169761657715,
-0.0011695370776578784,
0.09295447915792465,
0.06546559184789658,
-0.031979452818632126,
0.02189233899116516,
0.03681814670562744,
0.04794061556458473,
0.0029639550484716892,
-0.06908237934112549,
-0.02633938565850258,
-0.0002553184749558568,
-0.00022435479331761599,
-0.0718969851732254,
-0.02535499818623066,
0.0018636856693774462,
0.013975704088807106,
-0.04738577455282211,
0.0876966342329979,
0.023709582164883614,
-0.09333129227161407,
0.015454640612006187,
-0.09205637872219086,
-0.014473216608166695,
-0.0022391441743820906,
-0.030590424314141273,
-0.03722953051328659,
-0.05672679841518402,
-0.05713281407952309,
0.08302957564592361,
-0.04780576378107071,
0.025197889655828476,
-0.02711622603237629,
0.0162337776273489,
-0.04435903951525688,
0.003055471694096923,
-0.0013559529324993491,
0.02375638298690319,
-0.11972109973430634,
0.014845089986920357,
0.020018471404910088,
0.04354627802968025,
-0.06315289437770844,
0.01002635806798935,
-0.028991596773266792,
0.012056012637913227,
-0.04990092292428017,
0.08118438720703125,
-0.02902837283909321,
0.04747186228632927,
0.07393147051334381,
0.07727012783288956,
0.08102504909038544,
0.05419393256306648,
0.06459689885377884,
0.01013205386698246,
0.07938195765018463,
-0.03906593844294548,
0.015012743882834911,
-0.0053773559629917145,
0.10627023875713348,
-0.01764214225113392,
-4.065278247367132e-8,
-0.12250185757875443,
-0.00038259560824371874,
-0.07958190143108368,
0.04564298316836357,
0.0022033744025975466,
-0.055733710527420044,
-0.04030459001660347,
0.10355614125728607,
0.03346100449562073,
0.01889665052294731,
0.12307246029376984,
-0.0005899168318137527,
-0.0760611966252327,
-0.06123973801732063,
-0.003268313594162464,
0.044927362352609634,
-0.018916213884949684,
0.04715229198336601,
-0.007638026028871536,
0.05242308974266052,
0.08378119021654129,
0.03734260052442551,
0.004087080713361502,
0.009808724746108055,
0.0033813510090112686,
-0.10717057436704636,
-0.0341501459479332,
0.1360258013010025,
-0.006536724045872688,
-0.03818126395344734,
-0.057282887399196625,
-0.030446039512753487,
0.01306634210050106,
-0.023531902581453323,
-0.012815270572900772,
0.05965486541390419,
0.03766987845301628,
-0.014832851476967335,
-0.0025367802008986473,
-0.00803133100271225,
0.04298137128353119,
0.08553528785705566,
-0.11077173054218292,
0.03749702870845795,
0.0304249320179224,
-0.011498825624585152,
-0.045141901820898056,
-0.014980480074882507,
0.023632057011127472,
-0.08592354506254196,
-0.045503806322813034,
-0.042287327349185944,
0.06815455108880997,
0.012905222363770008,
0.05840640887618065,
-0.006150006782263517,
0.017974385991692543,
0.05177094042301178,
0.06438329070806503,
-0.006870879326015711,
-0.014803132973611355,
0.01515320036560297,
-0.021260181441903114,
-0.021113160997629166
] | 0 | 0 | 0 | 0 |
|
Probabilistic Learning to Defer: Handling Missing Expert Annotations and Controlling Workload Distribution | https://openreview.net/forum?id=zl0HLZOJC9 | [
"Cuong C. Nguyen",
"Thanh-Toan Do",
"Gustavo Carneiro"
] | Oral | Recent progress in machine learning research is gradually shifting its focus towards *human-AI cooperation* due to the advantages of exploiting the reliability of human experts and the efficiency of AI models. One of the promising approaches in human-AI cooperation is *learning to defer* (L2D), where the system analyses the input data and decides to make its own decision or defer to human experts. Although L2D has demonstrated state-of-the-art performance, in its standard setting, L2D entails a severe limitation: all human experts must annotate the whole training dataset of interest, resulting in a time-consuming and expensive annotation process that can subsequently influence the size and diversity of the training set. Moreover, the current L2D does not have a principled way to control workload distribution among human experts and the AI classifier, which is critical to optimise resource allocation. We, therefore, propose a new probabilistic modelling approach inspired by the mixture-of-experts, where the Expectation - Maximisation algorithm is leverage to address the issue of missing expert's annotations. Furthermore, we introduce a constraint, which can be solved efficiently during the E-step, to control the workload distribution among human experts and the AI classifier. Empirical evaluation on synthetic and real-world datasets shows that our proposed probabilistic approach performs competitively, or surpasses previously proposed methods assessed on the same benchmarks. | learning to defer, expectation - maximisation | null | 451 | null | [
-0.06566299498081207,
-0.05455765500664711,
0.0017818930791690946,
0.05760018527507782,
0.09522025287151337,
0.013323705643415451,
0.09160412847995758,
-0.01999843679368496,
-0.010556435212492943,
-0.03881591185927391,
-0.05196361616253853,
-0.05010175704956055,
0.04193131625652313,
0.029407480731606483,
-0.019512442871928215,
0.024376682937145233,
0.09734925627708435,
-0.0021600236650556326,
-0.08477158844470978,
-0.04739048331975937,
0.01926681399345398,
-0.0145759591832757,
0.007800672203302383,
0.011922307312488556,
0.0068076616153120995,
-0.007848949171602726,
0.032032910734415054,
-0.04858986288309097,
-0.03313993290066719,
-0.010599866509437561,
0.011395434848964214,
-0.04403075575828552,
0.009126025252044201,
-0.05691380053758621,
-0.06770146638154984,
0.022182948887348175,
-0.08093202114105225,
0.0439157597720623,
0.04132567346096039,
0.04186655580997467,
-0.04627150297164917,
0.03353477269411087,
-0.007804505992680788,
0.05758985877037048,
0.046254441142082214,
0.08120693266391754,
-0.028712399303913116,
-0.03736952692270279,
0.04981044679880142,
0.09405209124088287,
-0.01529886294156313,
-0.05452346056699753,
0.030752578750252724,
0.04104965180158615,
0.014254631474614143,
0.018860381096601486,
0.02711394429206848,
-0.0018860644195228815,
-0.034444257616996765,
0.005916775204241276,
-0.039476074278354645,
-0.09811000525951385,
0.01857685297727585,
-0.007772541604936123,
0.03599151596426964,
-0.07597947120666504,
0.03721541538834572,
0.058188896626234055,
-0.03744663670659065,
0.004514553118497133,
-0.014175111427903175,
0.09597014635801315,
-0.05322539061307907,
0.015276186168193817,
0.04314829036593437,
0.028928279876708984,
0.0023763279896229506,
-0.0009778111707419157,
0.01277493592351675,
-0.09372352808713913,
-0.025743667036294937,
0.031564418226480484,
0.04484141245484352,
-0.0018743568798527122,
0.08077391237020493,
-0.037209637463092804,
-0.023272456601262093,
0.06090690195560455,
0.06883794814348221,
-0.027413995936512947,
-0.11126326769590378,
0.030596347525715828,
0.03132111579179764,
-0.012069182470440865,
0.02963964082300663,
0.13422998785972595,
-0.027345536276698112,
-0.030206432566046715,
-0.07480545341968536,
0.1054057627916336,
-0.06037595868110657,
0.017834972590208054,
0.041727080941200256,
-0.08598721027374268,
0.02951979637145996,
0.013735843822360039,
-0.0030652410350739956,
0.04124867171049118,
0.031695861369371414,
-0.07179928570985794,
0.018076080828905106,
0.018958061933517456,
-0.002358716679736972,
-0.043132953345775604,
0.005054504610598087,
0.05076618492603302,
-0.003491263138130307,
0.020421253517270088,
-0.0388030968606472,
0.05519198253750801,
-0.018236514180898666,
-0.015637993812561035,
0.04397443309426308,
0.011633496731519699,
0.07591170072555542,
-0.00428784592077136,
-0.06052577495574951,
1.0681741043151128e-33,
0.06203586235642433,
-0.034080177545547485,
0.038620851933956146,
0.06655004620552063,
-0.014352092519402504,
-0.08344854414463043,
0.00468852836638689,
0.02506229653954506,
0.03550388664007187,
-0.09220071882009506,
-0.015501309186220169,
-0.03336882218718529,
-0.0037357972469180822,
0.031408440321683884,
0.021376226097345352,
0.012220364063978195,
0.01476308237761259,
0.050638455897569656,
-0.08467690646648407,
-0.02043847180902958,
0.08946972340345383,
-0.025399424135684967,
-0.010045691393315792,
-0.04962197691202164,
0.04615021497011185,
-0.004707647953182459,
0.07729010283946991,
-0.03360748663544655,
0.08329004049301147,
0.001270571956411004,
-0.03325607627630234,
-0.024960603564977646,
0.01791403815150261,
-0.02637738361954689,
0.01436058059334755,
-0.03358645364642143,
-0.012251554057002068,
0.01730782352387905,
-0.011183049529790878,
0.022002751007676125,
-0.005367175210267305,
0.006160676945000887,
0.035230860114097595,
-0.00654714647680521,
-0.07364135980606079,
-0.09502693265676498,
0.05268964171409607,
0.0006823633448220789,
-0.02788168005645275,
-0.055990055203437805,
0.03759704902768135,
-0.03362886607646942,
0.040617626160383224,
-0.010393591597676277,
-0.01103152520954609,
-0.01731139048933983,
-0.003661863738670945,
0.06722590327262878,
0.015926113352179527,
-0.03956872224807739,
-0.024931475520133972,
0.04082441329956055,
0.05933079868555069,
0.11693529784679413,
0.04280887544155121,
0.019434960559010506,
-0.022370336577296257,
0.09076524525880814,
0.07011546194553375,
-0.05662468448281288,
0.0006777268717996776,
-0.004492541309446096,
0.07759390026330948,
0.00016093904559966177,
0.054285530000925064,
0.012993081472814083,
0.01760539598762989,
-0.06271228194236755,
-0.009149211458861828,
-0.008662324398756027,
-0.0440027080476284,
-0.05175279080867767,
-0.02949940785765648,
-0.06126845255494118,
-0.001981158275157213,
-0.004156975541263819,
0.04411962255835533,
-0.053242310881614685,
-0.011242799460887909,
0.06864630430936813,
-0.06455276161432266,
-0.010205449536442757,
-0.11166632920503616,
0.009850266389548779,
0.0066568548791110516,
-8.949958684481578e-34,
-0.10660851001739502,
0.04175128787755966,
-0.023823607712984085,
0.017353463917970657,
0.06497224420309067,
-0.022228386253118515,
-0.006670631002634764,
0.013516126200556755,
0.030443880707025528,
0.016636013984680176,
-0.06803321838378906,
-0.05235195532441139,
0.06166285648941994,
-0.009373706765472889,
-0.007248093374073505,
0.05966299772262573,
-0.08465766161680222,
0.05771830305457115,
0.04448937997221947,
0.07272361218929291,
0.01783447153866291,
-0.00003197949263267219,
-0.07247085869312286,
0.025666330009698868,
0.027770116925239563,
0.04127397760748863,
-0.05765147879719734,
0.055383097380399704,
0.03231910243630409,
-0.025064514949917793,
-0.07733908295631409,
-0.09668566286563873,
-0.12687261402606964,
-0.07714606821537018,
-0.03470417112112045,
0.03167619928717613,
-0.027794500812888145,
0.015880698338150978,
-0.0618719644844532,
0.13326981663703918,
0.05634349212050438,
-0.023673968389630318,
-0.1429840475320816,
-0.02239128015935421,
-0.022336754947900772,
-0.08940791338682175,
-0.041650425642728806,
-0.10377476364374161,
-0.04792097955942154,
0.0027218537870794535,
-0.038293201476335526,
0.026434725150465965,
-0.08660557866096497,
0.032938748598098755,
-0.09166381508111954,
-0.0495537631213665,
0.03067592903971672,
0.007276102434843779,
0.05276808887720108,
0.07180257141590118,
-0.07211938500404358,
-0.020667789503932,
0.03295597806572914,
0.03647075966000557,
-0.025825805962085724,
-0.01641979068517685,
0.024373622611165047,
0.07468613982200623,
0.013904533348977566,
0.056539278477430344,
0.04983150213956833,
0.051764894276857376,
0.00869592372328043,
0.03361243009567261,
0.00882338359951973,
0.044621068984270096,
0.06593270599842072,
-0.08732574433088303,
0.029721705242991447,
-0.05500156059861183,
0.0692470371723175,
-0.06031224876642227,
0.05413942039012909,
0.029232870787382126,
0.07053688913583755,
0.09535552561283112,
0.044100191444158554,
0.05078725516796112,
-0.005619790405035019,
-0.01526736095547676,
-0.013689378276467323,
-0.036722563207149506,
-0.0033880642149597406,
0.05535247176885605,
-0.07015720754861832,
-4.492381577847482e-8,
-0.04174022376537323,
0.09386932104825974,
0.09115146100521088,
0.021989470347762108,
0.11224975436925888,
-0.005296336952596903,
-0.0879247859120369,
0.09214191883802414,
-0.03361329063773155,
0.0029883140232414007,
0.04596792906522751,
-0.01652180589735508,
0.015839384868741035,
-0.09822431951761246,
0.04484819993376732,
0.09834232926368713,
0.028099721297621727,
0.04638505354523659,
-0.047050025314092636,
0.003153942758217454,
0.11216071993112564,
-0.035981860011816025,
-0.058556415140628815,
0.014059830456972122,
0.06469385325908661,
-0.05585389584302902,
-0.019253259524703026,
0.06366155296564102,
0.014146038331091404,
-0.0010836058063432574,
-0.07756013423204422,
0.0033191011752933264,
-0.026520172134041786,
-0.02216954529285431,
0.0581192709505558,
0.10667020082473755,
0.016547372564673424,
-0.08581184595823288,
-0.07990050315856934,
-0.012199994176626205,
0.014071660116314888,
0.0641956552863121,
-0.00024106490309350193,
0.012641682289540768,
0.12277834862470627,
-0.030562840402126312,
-0.06430773437023163,
-0.11780412495136261,
0.013060353696346283,
-0.07558520883321762,
0.012901430949568748,
-0.10447143763303757,
-0.019591860473155975,
0.09951629489660263,
0.032057251781225204,
0.017374666407704353,
-0.010194501839578152,
-0.003211319213733077,
0.05330716818571091,
-0.002440636744722724,
0.025657081976532936,
0.03294452279806137,
-0.0628276839852333,
0.03515849635004997
] | 0 | 0 | 0 | 0 |
|
A Decade's Battle on Dataset Bias: Are We There Yet? | https://openreview.net/forum?id=SctfBCLmWo | [
"Zhuang Liu",
"Kaiming He"
] | Oral | We revisit the ``dataset classification'' experiment suggested by Torralba & Efros (2011) a decade ago, in the new era with large-scale, diverse, and hopefully less biased datasets as well as more capable neural network architectures. Surprisingly, we observe that modern neural networks can achieve excellent accuracy in classifying which dataset an image is from: e.g., we report 84.7% accuracy on held-out validation data for the three-way classification problem consisting of the YFCC, CC, and DataComp datasets. Our further experiments show that such a dataset classifier could learn semantic features that are generalizable and transferable, which cannot be explained by memorization. We hope our discovery will inspire the community to rethink issues involving dataset bias. | Vision datasets, Dataset bias, Deep learning | Modern large-scale vision datasets that are supposedly very general and diverse, are in fact still very biased | 407 | 2403.08632 | [
-0.06616389751434326,
-0.08431859314441681,
-0.011656607501208782,
0.038993436843156815,
0.11432990431785583,
0.02725379727780819,
-0.020347777754068375,
-0.0768556147813797,
-0.0740743950009346,
-0.04271848127245903,
-0.027138609439134598,
-0.016394339501857758,
0.018777428194880486,
0.032245662063360214,
-0.03063245117664337,
0.0021132694091647863,
0.07722786068916321,
0.07070419937372208,
-0.06261750310659409,
0.017145469784736633,
-0.045680828392505646,
0.03792343661189079,
0.06930462270975113,
-0.000839733169414103,
0.02204909548163414,
0.017537567764520645,
0.012009063735604286,
-0.008990513160824776,
-0.02706875652074814,
-0.03320052847266197,
0.06663617491722107,
0.041901860386133194,
-0.12006006389856339,
0.010497293435037136,
0.00199393043294549,
0.04119337350130081,
-0.031097762286663055,
0.023316849023103714,
0.0039708842523396015,
0.009518252685666084,
0.030876711010932922,
0.017015365883708,
0.028134578838944435,
0.05652453377842903,
0.0760110467672348,
0.05522551015019417,
0.03846706077456474,
-0.03838140890002251,
-0.0375179722905159,
0.005349636543542147,
-0.056747082620859146,
-0.031113769859075546,
-0.040865328162908554,
0.06996046751737595,
-0.06210566312074661,
0.021893547847867012,
0.058054279536008835,
0.07054755836725235,
-0.03467633202672005,
0.0496051199734211,
-0.013136157765984535,
-0.09358127415180206,
-0.002765775891020894,
0.029738474637269974,
-0.04073457792401314,
-0.0165272057056427,
-0.011869676411151886,
0.029306823387742043,
0.022311760112643242,
-0.07197604328393936,
0.00930581334978342,
0.10087992250919342,
-0.011292618699371815,
0.06846322864294052,
0.03574978560209274,
0.027985038235783577,
0.04265984147787094,
-0.016481613740324974,
0.09488090127706528,
-0.14187465608119965,
-0.016062630340456963,
-0.028481649234890938,
0.0397816076874733,
0.03142017498612404,
0.13413724303245544,
-0.038727935403585434,
-0.026311660185456276,
0.017999693751335144,
-0.09521254897117615,
-0.0005070376209914684,
0.02412262000143528,
-0.029165714979171753,
0.06825540214776993,
-0.02910595014691353,
-0.012443386018276215,
0.022862166166305542,
0.004056165926158428,
-0.05732233077287674,
-0.06369827687740326,
0.053843721747398376,
-0.07185523211956024,
0.09219959378242493,
-0.01309429295361042,
-0.0006533053237944841,
0.10013756155967712,
-0.023790905252099037,
0.05705331638455391,
-0.016221802681684494,
0.06644753366708755,
-0.07499471306800842,
-0.015986232087016106,
0.050978273153305054,
-0.044453658163547516,
-0.06854638457298279,
0.051310718059539795,
-0.0763922929763794,
-0.033419545739889145,
-0.0031742630526423454,
-0.041755158454179764,
0.04328851401805878,
-0.07689870893955231,
-0.007927036844193935,
0.0007971568848006427,
-0.0029400610364973545,
0.028220785781741142,
-0.03866957500576973,
-0.06987526267766953,
2.027710857437153e-33,
0.0695945993065834,
-0.007693955674767494,
0.02460244670510292,
0.03684655576944351,
0.06768510490655899,
-0.03778102621436119,
0.011135010980069637,
0.01889835298061371,
-0.043456725776195526,
0.04216688126325607,
-0.03039771132171154,
0.05762288346886635,
-0.02389182150363922,
0.10156615823507309,
0.01727990247309208,
0.035839762538671494,
-0.12999659776687622,
0.040057770907878876,
0.004086516331881285,
-0.034976180642843246,
0.08303308486938477,
0.022937724366784096,
0.08772636204957962,
-0.08411432057619095,
-0.040085602551698685,
-0.047268327325582504,
-0.012345192022621632,
-0.006542331073433161,
0.015966452658176422,
0.041032083332538605,
-0.0546877384185791,
0.04316532984375954,
0.007892024703323841,
0.08428080379962921,
0.040965739637613297,
-0.06482578814029694,
-0.0070917340926826,
-0.04534421116113663,
0.06388192623853683,
0.02857086807489395,
-0.026127489283680916,
0.002226316137239337,
-0.001922019524499774,
-0.003157022874802351,
-0.05569428950548172,
-0.0597367063164711,
0.04838106408715248,
-0.00776028074324131,
0.002116648480296135,
-0.023022456094622612,
0.05313736945390701,
-0.049843721091747284,
-0.040923282504081726,
0.0455077588558197,
0.0243417639285326,
0.024976462125778198,
-0.0355369858443737,
0.022073796018958092,
0.0369851291179657,
0.019026268273591995,
-0.01969946175813675,
0.10742099583148956,
0.02773786336183548,
0.033759281039237976,
-0.0641934797167778,
0.100445955991745,
-0.009682361036539078,
0.011624295264482498,
-0.02076042629778385,
-0.013928698375821114,
0.01480510551482439,
0.03982948511838913,
-0.12458919733762741,
-0.026215974241495132,
0.07030055671930313,
-0.0024283267557621002,
0.05245121940970421,
-0.053799550980329514,
-0.02996237576007843,
0.04001463204622269,
0.042551439255476,
-0.05408516526222229,
0.01632692851126194,
-0.04923434183001518,
-0.045181646943092346,
0.006859493907541037,
0.034146327525377274,
-0.058782853186130524,
0.0027139410376548767,
0.042968932539224625,
-0.0645260363817215,
0.0051411231979727745,
0.03276593238115311,
-0.032179154455661774,
-0.04245771840214729,
-2.2612477894829553e-33,
-0.08439228683710098,
0.02337639220058918,
-0.03836285322904587,
0.08305514603853226,
-0.061270784586668015,
-0.04106009006500244,
0.0009716062340885401,
0.02777727320790291,
-0.029335394501686096,
0.008081831969320774,
0.03218725323677063,
-0.027299078181385994,
0.013613119721412659,
-0.04781004413962364,
-0.05491311848163605,
-0.05289563909173012,
-0.07568473368883133,
0.0070678009651601315,
-0.04150153324007988,
0.025480447337031364,
0.04405709356069565,
0.030545497313141823,
-0.13739296793937683,
0.06479696929454803,
-0.06689431518316269,
0.07106294482946396,
-0.033864665776491165,
0.004977664444595575,
0.03297427296638489,
-0.00512564554810524,
-0.0560857392847538,
-0.042986124753952026,
-0.055458731949329376,
-0.06392234563827515,
-0.031968697905540466,
0.021718734875321388,
0.04219165816903114,
-0.02211977168917656,
-0.056945789605379105,
0.08856526762247086,
0.009087945334613323,
0.026028307154774666,
-0.13973328471183777,
-0.024659376591444016,
-0.009409531950950623,
-0.11153199523687363,
-0.022868862375617027,
0.045037418603897095,
0.03801315650343895,
0.04687851667404175,
0.0359540618956089,
-0.030971815809607506,
-0.040737152099609375,
0.07528739422559738,
0.0011876133503392339,
-0.028522828593850136,
0.0031105861999094486,
0.016098180785775185,
0.02870030328631401,
0.10231334716081619,
-0.06429474800825119,
0.007052971050143242,
-0.00016012665582820773,
-0.026718679815530777,
0.04207000881433487,
-0.015791170299053192,
-0.036168940365314484,
-0.016275541856884956,
0.011502757668495178,
0.04960467293858528,
0.09395299106836319,
-0.009989921934902668,
-0.07673350721597672,
-0.040339380502700806,
-0.048937343060970306,
0.017337901517748833,
0.03287501633167267,
-0.008627560921013355,
-0.007004051934927702,
0.005271495785564184,
0.09133783727884293,
-0.04914294555783272,
0.10397966206073761,
0.038339365273714066,
0.10615655034780502,
0.10072898864746094,
0.020936861634254456,
-0.026425780728459358,
0.03340192511677742,
-0.04690449312329292,
-0.0021796629298478365,
-0.046485207974910736,
-0.03582281246781349,
0.054138366132974625,
0.0018973842961713672,
-4.677934128949346e-8,
-0.09525015950202942,
0.07281991839408875,
0.03197962045669556,
0.046480365097522736,
-0.005032326094806194,
-0.07418206334114075,
0.012859471142292023,
0.0896119549870491,
-0.04639463499188423,
-0.024728907272219658,
0.15586145222187042,
0.09419196844100952,
-0.02954433113336563,
0.00945109874010086,
0.042231783270835876,
0.016458991914987564,
0.04698599502444267,
0.020581787452101707,
0.033750101923942566,
0.00971650518476963,
0.08613325655460358,
-0.03668566048145294,
0.07374287396669388,
-0.05054549127817154,
0.046666793525218964,
-0.10851439088582993,
-0.048583850264549255,
0.10632637143135071,
-0.010625049471855164,
0.026811519637703896,
-0.04981101304292679,
-0.05568462237715721,
-0.015723945572972298,
-0.09310854971408844,
0.10585329681634903,
0.052524156868457794,
-0.02351440116763115,
-0.04297512769699097,
-0.04705388844013214,
-0.07352413237094879,
-0.0479329377412796,
0.06497899442911148,
-0.036410994827747345,
0.03157176822423935,
-0.005030586384236813,
0.008725952357053757,
0.01520678773522377,
0.004804990720003843,
-0.04212591052055359,
0.009170686826109886,
0.01809958927333355,
0.018827199935913086,
-0.009907579980790615,
0.06565013527870178,
0.08191244304180145,
0.004918219055980444,
-0.040750984102487564,
-0.06693118810653687,
-0.05836765095591545,
0.04962840676307678,
-0.007200329564511776,
0.0051380968652665615,
-0.06905922293663025,
-0.040664270520210266
] | https://github.com/liuzhuang13/bias | 109 | 0 | 0 | 0 |
Knowing Your Target: Target-Aware Transformer Makes Better Spatio-Temporal Video Grounding | https://openreview.net/forum?id=WOzffPgVjF | [
"Xin Gu",
"Yaojie Shen",
"Chenxi Luo",
"Tiejian Luo",
"Yan Huang",
"Yuewei Lin",
"Heng Fan",
"Libo Zhang"
] | Oral | Transformer has attracted increasing interest in spatio-temporal video grounding, or STVG, owing to its end-to-end pipeline and promising result. Existing Transformer-based STVG approaches often leverage a set of object queries, which are initialized simply using zeros and then gradually learn target position information via iterative interactions with multimodal features, for spatial and temporal localization. Despite simplicity, these zero object queries, due to lacking target-specific cues, are hard to learn discriminative target information from interactions with multimodal features in complicated scenarios (e.g., with distractors or occlusion), resulting in degradation. Addressing this, we introduce a novel $\textbf{T}$arget-$\textbf{A}$ware Transformer for $\textbf{STVG}$ ($\textbf{TA-STVG}$), which seeks to adaptively generate object queries via exploring target-specific cues from the given video-text pair, for improving STVG. The key lies in two simple yet effective modules, comprising text-guided temporal sampling (TTS) and attribute-aware spatial activation (ASA), working in a cascade. The former focuses on selecting target-relevant temporal cues from a video utilizing holistic text information, while the latter aims at further exploiting the fine-grained visual attribute information of the object from previous target-aware temporal cues, which is applied for object query initialization. Compared to existing methods leveraging zero-initialized queries, object queries in our TA-STVG, directly generated from a given video-text pair, naturally carry target-specific cues, making them adaptive and better interact with multimodal features for learning more discriminative information to improve STVG. In our experiments on three benchmarks, including HCSTVG-v1/-v2 and VidSTG, TA-STVG achieves state-of-the-art performance and significantly outperforms the baseline, validating its efficacy. Moreover, TTS and ASA are designed for general purpose. When applied to existing methods such as TubeDETR and STCAT, we show substantial performance gains, verifying its generality. Code is released at https://github.com/HengLan/TA-STVG. | Spatio-Temporal Video Grounding | null | 284 | 2502.11168 | [
0.01838148944079876,
-0.06125182285904884,
0.024818111211061478,
0.02759162150323391,
0.10058116912841797,
0.008513526991009712,
0.040193915367126465,
-0.00916204508394003,
0.047875821590423584,
-0.08994857966899872,
-0.060786496847867966,
-0.06585429608821869,
-0.08294916152954102,
0.04246291518211365,
0.0378447100520134,
0.007037410978227854,
0.05974755436182022,
0.11010227352380753,
-0.034503426402807236,
-0.038843777030706406,
0.004086992237716913,
-0.020754000172019005,
0.003047608071938157,
-0.04526633024215698,
0.015150937251746655,
0.02712136320769787,
0.006388195790350437,
-0.06737088412046432,
0.0034969374537467957,
-0.06187107414007187,
0.0012975134886801243,
0.03314532712101936,
-0.01544004026800394,
0.046654362231492996,
-0.08505704998970032,
0.03748150169849396,
-0.007211468648165464,
-0.01412198320031166,
-0.04593677818775177,
-0.035348065197467804,
0.02531140297651291,
0.05756673216819763,
0.05039454996585846,
-0.044116489589214325,
0.010138588957488537,
0.04428156465291977,
0.09514190256595612,
-0.044114772230386734,
0.025565901771187782,
-0.026464102789759636,
-0.027872620150446892,
-0.08942535519599915,
-0.04044032841920853,
0.07164840400218964,
-0.0005123664741404355,
0.07490867376327515,
0.023925747722387314,
-0.022543741390109062,
0.06071052700281143,
-0.047365497797727585,
-0.016217151656746864,
-0.06576216220855713,
0.05590646341443062,
-0.056743163615465164,
-0.05445894971489906,
-0.08041586726903915,
0.12880295515060425,
-0.05662534013390541,
0.009266753681004047,
0.032864637672901154,
-0.08904733508825302,
0.053959883749485016,
-0.011193964630365372,
0.05481952801346779,
-0.03635617718100548,
0.09135420620441437,
0.02297569252550602,
-0.10157566517591476,
0.067494235932827,
-0.06795596331357956,
0.09721510857343674,
-0.03783820942044258,
-0.02415326237678528,
0.04332052916288376,
0.06043381243944168,
0.04858231917023659,
-0.03022451512515545,
0.06560415774583817,
0.0387229360640049,
0.00527211744338274,
-0.06858009845018387,
-0.054892025887966156,
0.019037572667002678,
-0.05124776065349579,
-0.028941260650753975,
-0.05103886500000954,
0.003269993932917714,
-0.0643102154135704,
0.008708085864782333,
0.08059632033109665,
0.06046532094478607,
-0.024757657200098038,
0.02430928684771061,
0.04568096250295639,
0.004714611917734146,
-0.007295619230717421,
-0.02526526153087616,
0.02634911611676216,
0.025464901700615883,
-0.039611972868442535,
0.004084800370037556,
0.0016988415736705065,
0.019423972815275192,
-0.08019660413265228,
-0.018902786076068878,
0.03616250306367874,
0.009068135172128677,
0.0010817639995366335,
0.008482703939080238,
-0.02533148229122162,
0.042251501232385635,
0.004899920895695686,
-0.09723643958568573,
0.026736322790384293,
0.07228300720453262,
-0.027193982154130936,
0.004247068427503109,
2.0731978970391675e-33,
0.01923433132469654,
0.048940449953079224,
0.04785330593585968,
0.01667274534702301,
0.046950336545705795,
-0.004671991802752018,
0.047813065350055695,
0.08561016619205475,
-0.00015854461526032537,
0.06460931897163391,
-0.024564526975154877,
-0.042811665683984756,
-0.049973007291555405,
0.05937759205698967,
0.014480683021247387,
-0.07707706093788147,
-0.06022506579756737,
-0.006916186772286892,
-0.019074834883213043,
0.03668303042650223,
0.01299363188445568,
0.010017245076596737,
-0.06020192429423332,
-0.0011293842690065503,
0.037667326629161835,
-0.036186348646879196,
0.033938582986593246,
-0.05354411527514458,
-0.012990684248507023,
-0.028698837384581566,
-0.10091517120599747,
0.04457473382353783,
-0.007610479835420847,
-0.006460118573158979,
0.062371350824832916,
-0.1160716861486435,
-0.08681541681289673,
-0.07017195969820023,
-0.01560308039188385,
0.0030437426175922155,
0.008085884153842926,
0.040098708122968674,
-0.03445281833410263,
-0.05112672597169876,
-0.062272392213344574,
-0.03742482140660286,
0.0015810851473361254,
0.0715995654463768,
-0.0782293900847435,
0.026516899466514587,
0.06823024153709412,
0.04089938849210739,
-0.05166679620742798,
-0.023930273950099945,
0.0007225007866509259,
0.027728142216801643,
0.05018511414527893,
0.014809347689151764,
0.037435971200466156,
0.0019919045735150576,
-0.028879418969154358,
-0.008533577434718609,
-0.0009981964249163866,
0.04035291075706482,
0.015501797199249268,
0.008642345666885376,
0.036698129028081894,
0.03647340461611748,
0.07388871163129807,
0.010546348989009857,
-0.01426293421536684,
0.06305655092000961,
0.019000757485628128,
-0.09450265020132065,
0.02772616595029831,
0.04608376324176788,
0.02543049119412899,
-0.012469077482819557,
0.06530890613794327,
0.02114139311015606,
-0.06966274976730347,
-0.044519659131765366,
0.05606672540307045,
-0.0716213583946228,
-0.009679763577878475,
0.018961159512400627,
0.0400729738175869,
-0.13369780778884888,
-0.067115917801857,
-0.04185838624835014,
-0.029899489134550095,
0.019703982397913933,
-0.09893639385700226,
-0.017186051234602928,
0.048795077949762344,
-9.723596516150843e-34,
0.04648607596755028,
0.018172919750213623,
-0.01519122626632452,
0.03954816237092018,
-0.04032899811863899,
-0.024526352062821388,
0.031472206115722656,
-0.03843218460679054,
0.03300829604268074,
-0.09901880472898483,
-0.054063744843006134,
-0.005709742195904255,
-0.0007055980386212468,
-0.020730797201395035,
0.039714209735393524,
0.05619579926133156,
-0.04618552699685097,
-0.00999880488961935,
-0.07374442368745804,
0.040241289883852005,
-0.013809927739202976,
0.05974595248699188,
-0.08344737440347672,
-0.020651843398809433,
-0.07936320453882217,
0.036626093089580536,
0.00012146557855885476,
0.03961694985628128,
0.015971500426530838,
-0.024062611162662506,
0.05201687663793564,
-0.032810308039188385,
0.030546443536877632,
-0.06068101525306702,
-0.08060350269079208,
0.059937264770269394,
0.056175149977207184,
-0.02405928447842598,
0.007991932332515717,
0.08206691592931747,
0.09148837625980377,
0.03063780814409256,
-0.027700135484337807,
0.09013289213180542,
-0.05970781669020653,
-0.0030089709907770157,
-0.060886628925800323,
0.06104336678981781,
0.015592473559081554,
0.04517102986574173,
-0.05269010365009308,
0.033848535269498825,
-0.1369207501411438,
-0.019268853589892387,
0.0415768176317215,
-0.08208651840686798,
0.0011278938036412,
-0.04149568825960159,
0.03365175053477287,
0.014561623334884644,
-0.04633869603276253,
-0.00929972156882286,
0.0012382857967168093,
-0.03764887526631355,
0.011730828322470188,
0.028882142156362534,
-0.009351355023682117,
-0.0175454281270504,
0.02991374023258686,
0.018412820994853973,
0.11794789135456085,
-0.012146200984716415,
0.002341386629268527,
-0.03493894264101982,
-0.03346007689833641,
0.00703749805688858,
0.02256670407950878,
0.01062755472958088,
0.027283525094389915,
-0.04744438827037811,
-0.030243460088968277,
-0.07083120942115784,
0.0030425109434872866,
0.09381508827209473,
0.07789020240306854,
0.14338301122188568,
0.0030868591275066137,
0.04263345152139664,
0.05496089905500412,
0.06701094657182693,
-0.0918763279914856,
0.058064255863428116,
-0.0017742107156664133,
0.058734215795993805,
-0.021111762151122093,
-4.543523246525183e-8,
-0.1149645447731018,
0.08805538713932037,
-0.1189233660697937,
-0.006385474465787411,
0.012331009842455387,
-0.027050968259572983,
0.011491295881569386,
0.0905572921037674,
0.004117082804441452,
-0.042824070900678635,
0.013435380533337593,
-0.06442219018936157,
0.005925406236201525,
0.0012276482302695513,
0.03531605377793312,
0.08122516423463821,
0.0004385857027955353,
-0.01235900353640318,
-0.04525527358055115,
0.044288113713264465,
-0.03162309527397156,
0.03926588594913483,
-0.010894544422626495,
-0.012072762474417686,
0.020335499197244644,
0.0254944059997797,
0.009631743654608727,
0.11320509761571884,
0.13431310653686523,
0.037771958857774734,
0.019136717543005943,
0.03324945271015167,
-0.009169784374535084,
-0.04090125486254692,
0.07176346331834793,
0.03869537264108658,
0.018751809373497963,
-0.04508485645055771,
-0.014942481182515621,
0.0004964639083482325,
-0.0047317915596067905,
0.03469816595315933,
-0.06698225438594818,
0.03415817394852638,
-0.11402565985918045,
0.06887558847665787,
-0.013093903660774231,
-0.14714913070201874,
-0.01909753866493702,
0.004529625177383423,
0.06414377689361572,
0.00010377552825957537,
-0.09101471304893494,
0.12110207229852676,
0.07108857482671738,
0.033427707850933075,
0.07708337903022766,
-0.011804738081991673,
0.027431229129433632,
0.10433124750852585,
-0.04719596728682518,
-0.004891077056527138,
-0.05668482556939125,
0.04238996282219887
] | https://github.com/HengLan/TA-STVG | 15 | 0 | 0 | 0 |
Open-World Reinforcement Learning over Long Short-Term Imagination | https://openreview.net/forum?id=vzItLaEoDa | [
"Jiajian Li",
"Qi Wang",
"Yunbo Wang",
"Xin Jin",
"Yang Li",
"Wenjun Zeng",
"Xiaokang Yang"
] | Oral | Training visual reinforcement learning agents in a high-dimensional open world presents significant challenges. While various model-based methods have improved sample efficiency by learning interactive world models, these agents tend to be “short-sighted”, as they are typically trained on short snippets of imagined experiences. We argue that the primary challenge in open-world decision-making is improving the exploration efficiency across a vast state space, especially for tasks that demand consideration of long-horizon payoffs. In this paper, we present LS-Imagine, which extends the imagination horizon within a limited number of state transition steps, enabling the agent to explore behaviors that potentially lead to promising long-term feedback. The foundation of our approach is to build a $\textit{long short-term world model}$. To achieve this, we simulate goal-conditioned jumpy state transitions and compute corresponding affordance maps by zooming in on specific areas within single images. This facilitates the integration of direct long-term values into behavior learning. Our method demonstrates significant improvements over state-of-the-art techniques in MineDojo. | World models, reinforcement learning, visual control | null | 242 | 2410.03618 | [
0.014406666159629822,
-0.03952448442578316,
0.08217339962720871,
0.03061799891293049,
0.1038714349269867,
-0.030093472450971603,
-0.03689434379339218,
0.02290530689060688,
-0.003570938017219305,
0.012515703216195107,
-0.020016377791762352,
-0.09200803190469742,
0.02970956452190876,
0.032533541321754456,
0.0480220653116703,
0.04059574380517006,
0.022886084392666817,
0.058272939175367355,
-0.05328372120857239,
-0.00013460716581903398,
0.0752619206905365,
-0.07091816514730453,
0.0595223642885685,
-0.026929089799523354,
-0.01690538600087166,
0.03149868920445442,
0.03221926838159561,
-0.030058305710554123,
0.03860905021429062,
-0.010874827392399311,
0.05999990552663803,
-0.023477541282773018,
-0.0034069668035954237,
-0.09853553771972656,
-0.021330364048480988,
0.09097711741924286,
-0.12001672387123108,
0.0014906194992363453,
-0.07161462306976318,
-0.06969932466745377,
0.014855526387691498,
0.002353021875023842,
0.037713490426540375,
0.03266586735844612,
0.013133437372744083,
0.015349174849689007,
0.05754626914858818,
-0.06291410326957703,
0.018803533166646957,
-0.02938411757349968,
-0.09617912769317627,
-0.08088385313749313,
0.009510375559329987,
-0.031023254618048668,
-0.0299823060631752,
0.02218136005103588,
0.008476601913571358,
-0.0227541271597147,
0.00007117330824257806,
-0.06313423067331314,
-0.05275475233793259,
-0.0754183679819107,
-0.004900241736322641,
-0.039763227105140686,
-0.003882326651364565,
-0.013397280126810074,
0.01493594329804182,
0.04219171404838562,
-0.0013609748566523194,
-0.040254417806863785,
0.008219712413847446,
0.011236904188990593,
0.03085433878004551,
-0.06348519027233124,
0.04350599646568298,
0.018453489989042282,
-0.024241754785180092,
-0.037626173347234726,
0.026272380724549294,
-0.03793422132730484,
0.018007392063736916,
-0.030886385589838028,
-0.027867665514349937,
0.02171006239950657,
0.03904774412512779,
0.034418802708387375,
-0.027907172217965126,
0.06503412872552872,
0.06406407058238983,
-0.0004981295205652714,
0.022962607443332672,
-0.07418687641620636,
-0.03645838052034378,
-0.044792041182518005,
0.010006628930568695,
-0.016577223315835,
-0.037966493517160416,
-0.1390140801668167,
-0.07706457376480103,
0.06286356598138809,
0.019102007150650024,
0.042723216116428375,
0.0402616485953331,
0.015871645882725716,
0.010836037807166576,
-0.0006470285006798804,
0.06535523384809494,
0.05207536369562149,
0.004485477693378925,
-0.07331466674804688,
0.03722686693072319,
-0.04443257302045822,
0.06979408860206604,
0.04618164524435997,
-0.038951605558395386,
-0.008544640615582466,
-0.02745727077126503,
-0.03953881934285164,
0.007738655898720026,
0.03006589226424694,
0.051063790917396545,
-0.012529772706329823,
0.036996010690927505,
0.010457892902195454,
0.020232178270816803,
-0.04342500492930412,
-0.01671293005347252,
-5.5377538867237965e-34,
0.05457477644085884,
-0.02274785190820694,
0.050066616386175156,
0.007650313898921013,
0.0774897113442421,
-0.004980419296771288,
0.04189888760447502,
0.0029873214662075043,
-0.04364490509033203,
0.05028030276298523,
0.019251948222517967,
0.03251359239220619,
-0.03190841153264046,
0.14651289582252502,
0.08815670013427734,
-0.050929922610521317,
-0.03497244790196419,
0.03958522528409958,
0.0022735607344657183,
0.0179291982203722,
0.021975066512823105,
-0.038909781724214554,
0.02788074128329754,
-0.04197151958942413,
0.03306705877184868,
0.0330565944314003,
0.023519054055213928,
0.02062949538230896,
-0.027379969134926796,
0.004516308661550283,
-0.022123757749795914,
0.06856349855661392,
-0.0719284936785698,
0.05519219860434532,
-0.01173142809420824,
0.04046861082315445,
-0.03283989429473877,
-0.0005626396741718054,
0.013295602053403854,
-0.011666595935821533,
-0.024530459195375443,
0.019986191764473915,
0.012255112640559673,
0.06951211392879486,
-0.06329698115587234,
-0.09102079272270203,
0.05706176534295082,
0.0429990254342556,
-0.12221012264490128,
0.056533072143793106,
0.03791734203696251,
-0.05199617147445679,
-0.041235022246837616,
-0.12723924219608307,
-0.05766027048230171,
-0.00431055249646306,
0.06849643588066101,
0.030799372121691704,
0.008938762359321117,
-0.005715101025998592,
-0.01533899549394846,
0.014589390717446804,
-0.017710410058498383,
0.0702136680483818,
-0.05629850924015045,
0.08177142590284348,
-0.02161167375743389,
0.057986605912446976,
0.0183382760733366,
0.04412389174103737,
-0.012933221645653248,
0.061881985515356064,
0.04048836976289749,
-0.11718808114528656,
0.09391938894987106,
-0.008130052126944065,
0.06475795805454254,
-0.07967399060726166,
0.018497072160243988,
0.02274101786315441,
0.03903969004750252,
-0.05461995303630829,
-0.05422292277216911,
0.012837190181016922,
0.036638498306274414,
-0.032637469470500946,
-0.0034177035558968782,
-0.11131089925765991,
-0.040977995842695236,
0.0024322927929461002,
-0.1388614922761917,
-0.054628465324640274,
-0.021133245900273323,
0.026096604764461517,
0.008615628816187382,
4.897393237208778e-34,
0.024279246106743813,
-0.03721073642373085,
0.05782477930188179,
0.03701614961028099,
-0.006066892761737108,
-0.039707258343696594,
-0.018202543258666992,
0.008774351328611374,
-0.011398711241781712,
-0.08250435441732407,
-0.08210486173629761,
0.06319200992584229,
0.03951480612158775,
0.021613238379359245,
0.01601078361272812,
-0.03941972926259041,
0.04638379439711571,
-0.004486000631004572,
0.007060097996145487,
0.0009972599800676107,
0.04310810938477516,
0.0522969551384449,
-0.10046235471963882,
-0.003314409637823701,
-0.05423104763031006,
0.06982140243053436,
0.023308590054512024,
0.0542706623673439,
-0.010142560116946697,
0.02760523557662964,
-0.06653432548046112,
-0.06926121562719345,
-0.013647567480802536,
-0.02399815246462822,
-0.04638701304793358,
0.08764976263046265,
0.062192995101213455,
-0.025484707206487656,
-0.05345253273844719,
0.11502443253993988,
0.02856668271124363,
-0.1793341338634491,
-0.023696571588516235,
0.016227539628744125,
-0.03258062154054642,
-0.0072456179186701775,
-0.08658091723918915,
0.042339909821748734,
-0.03133203461766243,
0.019004927948117256,
-0.05142639949917793,
0.07463180273771286,
-0.10454770177602768,
-0.034141987562179565,
-0.04333845525979996,
-0.07220069319009781,
-0.09439866989850998,
0.0510159507393837,
0.0661754235625267,
-0.019754683598876,
-0.03460656851530075,
-0.040336571633815765,
-0.011125223711133003,
0.05807638168334961,
-0.022985348477959633,
-0.020057234913110733,
-0.008826140314340591,
0.025931011885404587,
0.04497043788433075,
-0.030024871230125427,
0.01928034797310829,
0.00922465417534113,
0.008540663868188858,
-0.03738708794116974,
0.07456858456134796,
-0.013342211954295635,
0.009412095881998539,
-0.005138157866895199,
0.05481346696615219,
-0.03957102447748184,
0.0555950403213501,
-0.03516758605837822,
0.03064311109483242,
0.048963483422994614,
0.05322413891553879,
0.1090494841337204,
-0.030694659799337387,
0.10306186974048615,
-0.0017544117290526628,
-0.024309661239385605,
-0.04346024990081787,
0.047675590962171555,
0.012154428288340569,
0.10073554515838623,
-0.04270336031913757,
-4.155182153908754e-8,
-0.09184233844280243,
0.014219185337424278,
0.11025016009807587,
0.03964086249470711,
0.06325217336416245,
0.03923307731747627,
0.027666183188557625,
0.060152459889650345,
-0.027159158140420914,
0.004596375394612551,
0.12497834861278534,
0.03289620950818062,
0.057847920805215836,
-0.07568622380495071,
-0.008023218251764774,
0.038426171988248825,
0.028694594278931618,
-0.010833973996341228,
-0.003921329043805599,
0.0781894251704216,
0.018645523115992546,
0.014091876335442066,
0.07569576799869537,
-0.06919121742248535,
-0.07175059616565704,
-0.0661388486623764,
-0.07549547404050827,
0.03942568972706795,
0.03167927265167236,
-0.02682899683713913,
0.08837325125932693,
0.013946662656962872,
0.05749186873435974,
0.02464745193719864,
-0.0019093765877187252,
-0.0054268729873001575,
-0.011333001777529716,
-0.03248095512390137,
0.04088669642806053,
-0.03302071616053581,
-0.03597055748105049,
0.0979367345571518,
-0.0016822544857859612,
0.003718077903613448,
-0.05470021069049835,
0.030844634398818016,
-0.017348188906908035,
-0.10385369509458542,
0.050735581666231155,
-0.004223098512738943,
-0.0675884410738945,
-0.012273051775991917,
-0.04082106053829193,
0.09981245547533035,
0.09506923705339432,
-0.02793719619512558,
0.028600776568055153,
-0.01619577780365944,
0.02121134102344513,
0.14760351181030273,
0.021435461938381195,
-0.04032878205180168,
-0.1301414668560028,
0.002893372904509306
] | https://github.com/qiwang067/LS-Imagine | 101 | 0 | 0 | 0 |
OLMoE: Open Mixture-of-Experts Language Models | https://openreview.net/forum?id=xXTkbTBmqq | [
"Niklas Muennighoff",
"Luca Soldaini",
"Dirk Groeneveld",
"Kyle Lo",
"Jacob Morrison",
"Sewon Min",
"Weijia Shi",
"Evan Pete Walsh",
"Oyvind Tafjord",
"Nathan Lambert",
"Yuling Gu",
"Shane Arora",
"Akshita Bhagia",
"Dustin Schwenk",
"David Wadden",
"Alexander Wettig",
"Binyuan Hui",
"Tim Dettmers",
"Douwe Kiela",
"Ali Farhadi",
"et al. (4 additional authors not shown)"
] | Oral | We introduce OLMoE, a fully open, state-of-the-art language model leveraging sparse Mixture-of-Experts (MoE). OLMoE-1B-7B has 7 billion (B) parameters but uses only 1B per input token. We pretrain it on 5 trillion tokens and further adapt it to create OLMoE-1B-7B-Instruct. Our models outperform all available models with similar active parameters, even surpassing larger ones like Llama2-13B-Chat and DeepSeekMoE-16B. We present novel findings on MoE training, define and analyze new routing properties showing high specialization in our model, and open-source all our work: model weights, training data, code, and logs. | large language models, mixture-of-experts, open-source | A state-of-the-art Mixture-of-Experts LLM with 1B active and 7B total parameters trained for 5T tokens that is 100% open-source | 211 | 2409.02060 | [
-0.028544234111905098,
-0.11009832471609116,
0.047672927379608154,
0.10204517096281052,
0.025353703647851944,
-0.06629350781440735,
-0.016934234648942947,
0.050711460411548615,
-0.04559660702943802,
-0.06741649657487869,
-0.07885564863681793,
-0.027426814660429955,
-0.0020344522781670094,
-0.007161970250308514,
0.014162575826048851,
0.06888309866189957,
0.06754638254642487,
0.0558619350194931,
-0.11406043916940689,
-0.0828811451792717,
0.016077183187007904,
0.005752853117883205,
0.03270329535007477,
-0.013876968063414097,
0.031811412423849106,
0.012952969409525394,
0.032545093446969986,
-0.009947280399501324,
0.03270062059164047,
-0.07093504816293716,
0.023655962198972702,
0.060960803180933,
0.015295298770070076,
0.038759540766477585,
0.0034779703710228205,
0.046221762895584106,
-0.05844049155712128,
-0.047053027898073196,
0.0015112004475668073,
-0.05741040036082268,
0.004666413180530071,
0.03742179274559021,
0.020292315632104874,
0.0517466738820076,
0.025563599541783333,
-0.03545578569173813,
0.02198638580739498,
-0.03282846137881279,
-0.0003719487867783755,
0.006460397969931364,
-0.09136497229337692,
-0.0954902246594429,
0.006816740147769451,
0.06641297042369843,
0.0033736079931259155,
0.022981815040111542,
-0.02318575419485569,
-0.010202807374298573,
-0.009282952174544334,
0.0011323129292577505,
-0.07556385546922684,
-0.05892021581530571,
-0.0673762783408165,
0.003751866053789854,
0.01651947759091854,
-0.02469867281615734,
-0.0014391285367310047,
0.030050698667764664,
-0.002736821537837386,
0.01585214026272297,
0.004560175817459822,
0.08757532387971878,
0.00924272183328867,
0.11098320037126541,
0.008321958594024181,
0.005447122268378735,
0.056346431374549866,
0.03412758931517601,
0.046971261501312256,
-0.03594266250729561,
0.02796333096921444,
-0.009908835403621197,
-0.002076837932690978,
-0.037822920829057693,
0.13192452490329742,
-0.054148152470588684,
0.027905989438295364,
0.04408932477235794,
0.0800948217511177,
0.02440575696527958,
-0.08628790080547333,
-0.032471850514411926,
0.09098739922046661,
-0.03372512757778168,
-0.006060012150555849,
0.05344688892364502,
0.02346917986869812,
-0.02836184948682785,
-0.026874540373682976,
0.14940029382705688,
-0.055851101875305176,
0.0259820818901062,
0.04524187743663788,
-0.003133652964606881,
-0.033678021281957626,
0.03607066348195076,
0.06257981806993484,
0.045733384788036346,
0.03859901800751686,
0.011269468814134598,
0.02253858745098114,
0.01158571895211935,
0.05037162825465202,
-0.0389411635696888,
-0.02928978204727173,
0.015483314171433449,
-0.011532606557011604,
-0.05255994573235512,
-0.020511943846940994,
0.1155187264084816,
-0.0607353113591671,
-0.007607603911310434,
0.0356263630092144,
-0.0263820830732584,
-0.04153909534215927,
0.044192418456077576,
-0.011504740454256535,
1.5382955293467205e-33,
0.09596885740756989,
0.09206441789865494,
-0.012656284496188164,
0.03160437196493149,
0.011781369335949421,
0.021970830857753754,
0.0215314794331789,
-0.013025924563407898,
-0.07870912551879883,
-0.08279022574424744,
-0.07294777035713196,
0.07025361806154251,
-0.024306556209921837,
0.05363099277019501,
0.0448152981698513,
-0.022018365561962128,
-0.03261948749423027,
0.07557171583175659,
0.018023690208792686,
-0.05025561898946762,
0.07076183706521988,
0.010814190842211246,
-0.052791692316532135,
0.025590259581804276,
0.001366433105431497,
-0.011653779074549675,
0.09269978106021881,
-0.09171773493289948,
-0.0268437247723341,
0.031910669058561325,
-0.04418332129716873,
-0.018332626670598984,
-0.003582169534638524,
0.03483953699469566,
0.048226431012153625,
-0.012715929187834263,
-0.01953153870999813,
-0.06336045265197754,
0.020585939288139343,
-0.04392196238040924,
-0.04803867265582085,
0.011745287105441093,
0.004593747667968273,
-0.006031080149114132,
-0.11736901849508286,
-0.07426095753908157,
-0.043783485889434814,
-0.003335731802508235,
0.017091581597924232,
-0.01928713358938694,
0.036531176418066025,
-0.016058040782809258,
-0.006006218492984772,
0.004304889123886824,
0.026486113667488098,
-0.07987188547849655,
0.033282943069934845,
0.05829589441418648,
0.0831112340092659,
0.10082310438156128,
0.01573638804256916,
0.03483537212014198,
0.033234186470508575,
0.0272320955991745,
0.06884308904409409,
0.028458772227168083,
-0.048541732132434845,
0.10971447080373764,
0.07461051642894745,
-0.011830711737275124,
-0.04320521280169487,
-0.004759586416184902,
0.011800868436694145,
0.008610058575868607,
-0.023776208981871605,
0.05746838450431824,
0.032139863818883896,
-0.11560218781232834,
-0.007039335556328297,
-0.014220246113836765,
0.0032270445954054594,
0.021307792514562607,
0.035766128450632095,
-0.023118028417229652,
-0.032592110335826874,
-0.015262627974152565,
0.02372882515192032,
-0.06263628602027893,
-0.02988915704190731,
0.01477786060422659,
-0.03261873126029968,
-0.03417779505252838,
-0.04852566123008728,
0.003529556328430772,
0.015361612662672997,
1.0085813850508451e-34,
-0.033916160464286804,
0.07746214419603348,
-0.04647253453731537,
0.06991759687662125,
-0.08341781795024872,
-0.0968000516295433,
0.03011430986225605,
0.03626382350921631,
-0.030766598880290985,
0.018976835533976555,
-0.0010348669020459056,
-0.06405960768461227,
0.1388915628194809,
-0.0160817950963974,
0.1108754351735115,
-0.07469704747200012,
-0.1059243232011795,
-0.015808062627911568,
0.06924999505281448,
0.07768397033214569,
-0.05640348047018051,
0.02861666865646839,
-0.1253952980041504,
-0.0025657438673079014,
-0.0006537204026244581,
0.05363045260310173,
-0.06499183177947998,
0.0828595981001854,
-0.05858011916279793,
0.04084651917219162,
-0.0249166339635849,
-0.014113475568592548,
-0.058403048664331436,
-0.04694370552897453,
-0.006555838044732809,
0.1204589456319809,
0.04188043624162674,
0.01317670103162527,
0.05808968096971512,
0.016334453597664833,
0.01785476878285408,
-0.053168538957834244,
-0.042176395654678345,
0.0003536401782184839,
-0.09842142462730408,
0.012893513776361942,
-0.0763387382030487,
-0.03150566294789314,
-0.008483081124722958,
-0.07825625687837601,
0.0032497080974280834,
0.06336861848831177,
-0.06473296135663986,
-0.010281877592206001,
-0.03763489052653313,
-0.05208438262343407,
-0.020413625985383987,
-0.046084750443696976,
-0.014611631631851196,
-0.0148847047239542,
0.012566487304866314,
-0.03641688451170921,
0.04418470710515976,
0.021042149513959885,
0.03570117801427841,
-0.10204534977674484,
-0.01199446152895689,
0.008416621014475822,
-0.06146339699625969,
-0.0528155080974102,
0.006895528640598059,
-0.08560417592525482,
0.02592318132519722,
0.06985461711883545,
0.03159978613257408,
0.028000617399811745,
0.07532919198274612,
-0.011027174070477486,
-0.023678677156567574,
0.003637545509263873,
-0.0021112740505486727,
0.03098815307021141,
0.03952416032552719,
-0.006119106896221638,
0.08172634243965149,
0.11378498375415802,
0.030418476089835167,
0.08020365238189697,
0.009523862972855568,
0.08891543000936508,
-0.018476363271474838,
0.002933724783360958,
0.023278670385479927,
0.042913831770420074,
-0.06014201417565346,
-3.74175783690589e-8,
-0.11458586901426315,
0.012030516751110554,
0.020695885643363,
0.06624523550271988,
0.10303294658660889,
-0.036585960537195206,
-0.025091158226132393,
0.0557265467941761,
-0.0038528519216924906,
0.06707433611154556,
0.06026410311460495,
0.055429305881261826,
-0.10003601014614105,
-0.055772311985492706,
-0.009198927320539951,
0.09452217072248459,
-0.02796192839741707,
0.01403390895575285,
-0.041279371827840805,
-0.01813478209078312,
0.026098018512129784,
0.011025065556168556,
0.06270407140254974,
0.019349047914147377,
0.03743692487478256,
-0.07105918228626251,
-0.0172278955578804,
0.10223337262868881,
0.0421355739235878,
0.000003357428568051546,
-0.10875573754310608,
-0.0006972774863243103,
0.016073022037744522,
-0.07120100408792496,
0.0006149336695671082,
0.14825548231601715,
-0.008521453477442265,
-0.05394836887717247,
-0.08878961205482483,
-0.016581349074840546,
0.04969208315014839,
-0.04004736617207527,
-0.017374860122799873,
0.005560125689953566,
0.017302529886364937,
0.015893369913101196,
-0.05721543729305267,
-0.04677459970116615,
-0.002729184925556183,
-0.04574238136410713,
-0.05052792653441429,
-0.014164854772388935,
-0.015324180945754051,
0.07721356302499771,
0.026113560423254967,
-0.0049015190452337265,
-0.08720642328262329,
-0.030766990035772324,
0.044961754232645035,
0.052539095282554626,
0.0055514611303806305,
-0.022694673389196396,
0.006770862732082605,
0.029593845829367638
] | https://github.com/allenai/OLMoE | 726 | 0 | 0 | 0 |
Do Vision-Language Models Represent Space and How? Evaluating Spatial Frame of Reference under Ambiguities | https://openreview.net/forum?id=84pDoCD4lH | [
"Zheyuan Zhang",
"Fengyuan Hu",
"Jayjun Lee",
"Freda Shi",
"Parisa Kordjamshidi",
"Joyce Chai",
"Ziqiao Ma"
] | Oral | Spatial expressions in situated communication can be ambiguous, as their meanings vary depending on the frames of reference (FoR) adopted by speakers and listeners. While spatial language understanding and reasoning by vision-language models (VLMs) have gained increasing attention, potential ambiguities in these models are still under-explored. To address this issue, we present the COnsistent Multilingual Frame Of Reference Test (COMFORT), an evaluation protocol to systematically assess the spatial reasoning capabilities of VLMs. We evaluate nine state-of-the-art VLMs using COMFORT. Despite showing some alignment with English conventions in resolving ambiguities, our experiments reveal significant shortcomings of VLMs: notably, the models (1) exhibit poor robustness and consistency, (2) lack the flexibility to accommodate multiple FoRs, and (3) fail to adhere to language-specific or culture-specific conventions in cross-lingual tests, as English tends to dominate other languages. With a growing effort to align vision-language models with human cognitive intuitions, we call for more attention to the ambiguous nature and cross-cultural diversity of spatial reasoning. | vision-language models, spatial reasoning, multimodal reasoning | We present an evaluation protocol to systematically assess the spatial reasoning capabilities of vision language models, and shed light on the ambiguity and cross-cultural diversity of frame of reference in spatial reasoning. | 194 | 2410.17385 | [
0.07608702033758163,
-0.09904163330793381,
0.030067307874560356,
0.01640576310455799,
0.06267469376325607,
-0.03113461099565029,
0.07598066329956055,
-0.024347085505723953,
0.03353927657008171,
-0.043981678783893585,
-0.050133027136325836,
-0.08249907195568085,
0.044062063097953796,
0.10969847440719604,
0.02295926958322525,
-0.06951376050710678,
0.06304565817117691,
0.028223492205142975,
-0.05216577276587486,
0.03644835576415062,
0.03820401802659035,
0.010857632383704185,
0.05836432799696922,
0.03207464888691902,
0.05107532814145088,
-0.002639584708958864,
-0.023496730253100395,
-0.0581771619617939,
0.005846449173986912,
0.014836333692073822,
0.023142512887716293,
0.10695444792509079,
-0.013925833627581596,
0.09053683280944824,
0.03840279206633568,
0.03143055737018585,
-0.001816698000766337,
0.021461760625243187,
-0.024642586708068848,
-0.014371349476277828,
-0.12665973603725433,
-0.050563324242830276,
0.09166835993528366,
0.005570356268435717,
0.03011518530547619,
0.026135394349694252,
-0.01967475190758705,
0.024981684982776642,
-0.0745113417506218,
-0.004558535758405924,
-0.16179372370243073,
0.014216332696378231,
-0.014122727327048779,
-0.05367506295442581,
-0.057027433067560196,
-0.020627804100513458,
0.04777199402451515,
0.06566974520683289,
0.022669699043035507,
-0.01525385957211256,
-0.0058676679618656635,
-0.07894401252269745,
0.00396597059443593,
0.10863053798675537,
-0.07053422182798386,
0.017905225977301598,
-0.10395951569080353,
0.034523963928222656,
-0.05115262418985367,
-0.029431218281388283,
-0.053481657058000565,
0.03626430034637451,
0.029258156195282936,
0.012951057404279709,
0.0310335885733366,
0.08555634319782257,
-0.008984848856925964,
-0.06534584611654282,
0.016263777390122414,
-0.08539538085460663,
0.041261933743953705,
0.014686021953821182,
-0.02725013718008995,
-0.03341067209839821,
0.12283917516469955,
0.031422100961208344,
-0.015050557442009449,
0.02926449477672577,
-0.0053914375603199005,
-0.025644395500421524,
0.010263165459036827,
-0.046079933643341064,
0.00687218876555562,
0.028489883989095688,
0.12372173368930817,
0.028688497841358185,
0.014229381456971169,
-0.033188723027706146,
-0.05888477340340614,
0.0566398985683918,
0.051584284752607346,
0.05858020856976509,
0.03133733570575714,
-0.029762521386146545,
-0.004525606986135244,
-0.04286923632025719,
-0.0373677983880043,
-0.010762589983642101,
0.0006148077081888914,
-0.10657205432653427,
-0.028897572308778763,
0.027564136311411858,
0.0035968408919870853,
0.059643253684043884,
-0.036912765353918076,
-0.0315115712583065,
-0.004955440293997526,
0.015607870183885098,
0.1554366946220398,
-0.058700818568468094,
-0.01095626037567854,
0.036664679646492004,
0.009918510913848877,
-0.048369478434324265,
0.04026025906205177,
-0.0053925542160868645,
0.03399315103888512,
6.94621285077983e-34,
0.03203567489981651,
0.007750781252980232,
0.024737421423196793,
0.07685235142707825,
0.038506459444761276,
-0.030986705794930458,
-0.05322890728712082,
0.03613034635782242,
0.003977224696427584,
0.040152426809072495,
0.03893496096134186,
0.04208582639694214,
0.007358592003583908,
0.03420978784561157,
0.03054450824856758,
0.07744156569242477,
0.005106339696794748,
0.05915867164731026,
-0.07640563696622849,
0.055982980877161026,
-0.0034593085292726755,
-0.008598321117460728,
0.004689868073910475,
-0.005476855207234621,
-0.04646192491054535,
0.006750508211553097,
0.019459914416074753,
-0.05206489935517311,
-0.09070266038179398,
0.00521004619076848,
-0.12428177148103714,
0.0364082045853138,
-0.0228361077606678,
0.06904322654008865,
0.02098720706999302,
-0.08689739555120468,
-0.008674330078065395,
-0.06776729226112366,
-0.010743622668087482,
-0.013060785830020905,
-0.07213485985994339,
0.019881410524249077,
-0.07323450595140457,
0.06653334945440292,
0.030817018821835518,
0.03758896514773369,
-0.03494499996304512,
0.060056913644075394,
-0.10346654802560806,
0.0746074840426445,
0.023500727489590645,
0.046927500516176224,
0.004913519136607647,
-0.03428485989570618,
0.043692685663700104,
0.026806257665157318,
0.04568684846162796,
0.02603248506784439,
0.036419015377759933,
-0.037740740925073624,
-0.02238810807466507,
0.05995713919401169,
0.05356794595718384,
-0.006202538497745991,
0.050059136003255844,
-0.041642993688583374,
-0.0637606531381607,
0.0191671010106802,
0.03622272610664368,
0.015339847654104233,
-0.0024013121146708727,
0.04541053622961044,
-0.012297575362026691,
0.039182454347610474,
0.02003611996769905,
-0.03237785026431084,
-0.025947267189621925,
-0.026911476626992226,
0.054555200040340424,
-0.028072763234376907,
0.04275454208254814,
-0.016415627673268318,
-0.0288096833974123,
-0.0031196167692542076,
-0.02632051892578602,
-0.0856219157576561,
0.036885011941194534,
-0.08185736835002899,
-0.024763217195868492,
-0.07110118120908737,
0.017562737688422203,
-0.06931266188621521,
-0.009163021109998226,
-0.08982948213815689,
0.010967658832669258,
-2.049367871483115e-33,
-0.02233068458735943,
-0.06917009502649307,
-0.03636239469051361,
-0.015343835577368736,
-0.08092601597309113,
-0.10049452632665634,
0.07577797770500183,
-0.044994451105594635,
0.024757929146289825,
-0.06624509394168854,
-0.06152914837002754,
-0.03440683335065842,
0.03380783647298813,
0.06512776762247086,
0.012307796627283096,
-0.045902661979198456,
0.060418639332056046,
-0.0052381944842636585,
-0.002255386905744672,
0.07854323089122772,
0.07121265679597855,
0.03665698319673538,
-0.11243364959955215,
-0.008178184740245342,
-0.12465211004018784,
0.06502342969179153,
-0.005102981813251972,
-0.05835280567407608,
-0.11166507750749588,
0.009662977419793606,
0.005250055342912674,
0.03579455986618996,
-0.015803806483745575,
-0.03292812034487724,
-0.012697051279246807,
0.039226822555065155,
0.011307579465210438,
-0.09528063982725143,
-0.0635598748922348,
0.03628324344754219,
-0.0025282809510827065,
-0.022680118680000305,
-0.029149943962693214,
-0.006547880824655294,
-0.0046812244690954685,
0.05006342753767967,
-0.021182097494602203,
-0.002236910630017519,
-0.028999144211411476,
-0.01439017802476883,
-0.007568672299385071,
0.008153419010341167,
-0.05007539689540863,
-0.01244125235825777,
-0.005146948620676994,
-0.01824265532195568,
-0.04816698282957077,
-0.056945327669382095,
-0.0008896106737665832,
-0.005394544452428818,
0.016348227858543396,
-0.01148785650730133,
0.006584298796951771,
-0.0015535566490143538,
-0.05261637642979622,
0.009562958963215351,
-0.029809117317199707,
-0.034222256392240524,
0.06028930842876434,
-0.006390328519046307,
0.03535528481006622,
-0.09320145100355148,
-0.09827731549739838,
0.04871717095375061,
0.045067932456731796,
0.04744390770792961,
0.10068085044622421,
-0.021925760433077812,
0.0032933256588876247,
-0.03934067487716675,
-0.003507173852995038,
-0.09303317964076996,
0.054107122123241425,
0.05607721582055092,
0.0005891109467484057,
0.06467307358980179,
-0.04383616894483566,
0.032674871385097504,
0.049618616700172424,
0.010725741274654865,
-0.034647803753614426,
0.03471776470541954,
-0.07563763856887817,
0.06808966398239136,
-0.019107988104224205,
-4.5329564102303266e-8,
-0.07964624464511871,
0.05390700325369835,
-0.013131373561918736,
0.0018415898084640503,
-0.0674467384815216,
-0.04493102431297302,
-0.03926937282085419,
0.07207228988409042,
-0.044662151485681534,
0.0106301698833704,
-0.04353228211402893,
0.01081042643636465,
-0.028399024158716202,
0.019812753424048424,
0.03485400602221489,
0.06937989592552185,
0.014355898834764957,
0.009235136210918427,
-0.023316966369748116,
0.09592646360397339,
0.0977378711104393,
0.015187971293926239,
0.004406983498483896,
0.04079766571521759,
0.002562497043982148,
-0.014679147861897945,
-0.08316455036401749,
0.09872937947511673,
0.08835824579000473,
-0.09026414155960083,
0.13110144436359406,
0.037178799510002136,
-0.027154257521033287,
-0.02163872867822647,
-0.01103484258055687,
-0.00418526167050004,
-0.04616339132189751,
0.00029472942696884274,
0.08195295929908752,
0.005629043560475111,
-0.00217721750959754,
-0.0034356561955064535,
-0.0893508791923523,
0.06090608611702919,
0.06023012474179268,
0.03661787882447243,
-0.021257255226373672,
-0.07497415691614151,
-0.08339433372020721,
-0.02403576485812664,
-0.0952828973531723,
0.013828454539179802,
-0.04510056599974632,
0.07464230805635452,
0.03856876492500305,
-0.020259922370314598,
-0.017071353271603584,
0.0016187377041205764,
0.00206988793797791,
0.067470483481884,
0.01270973403006792,
0.158848375082016,
-0.08110390603542328,
-0.032019052654504776
] | https://github.com/sled-group/COMFORT | 8 | 0 | 0 | 0 |
SAM 2: Segment Anything in Images and Videos | https://openreview.net/forum?id=Ha6RTeWMd0 | [
"Nikhila Ravi",
"Valentin Gabeur",
"Yuan-Ting Hu",
"Ronghang Hu",
"Chaitanya Ryali",
"Tengyu Ma",
"Haitham Khedr",
"Roman Rädle",
"Chloe Rolland",
"Laura Gustafson",
"Eric Mintun",
"Junting Pan",
"Kalyan Vasudev Alwala",
"Nicolas Carion",
"Chao-Yuan Wu",
"Ross Girshick",
"Piotr Dollar",
"Christoph Feichtenhofer"
] | Oral | We present Segment Anything Model 2 (SAM 2), a foundation model towards solving promptable visual segmentation in images and videos. We build a data engine, which improves model and data via user interaction, to collect the largest video segmentation dataset to date. Our model is a simple transformer architecture with streaming memory for real-time video processing. SAM 2 trained on our data provides strong performance across a wide range of tasks. In video segmentation, we observe better accuracy, using 3x fewer interactions than prior approaches. In image segmentation, our model is more accurate and 6x faster than the Segment Anything Model (SAM). We believe that our data, model, and insights will serve as a significant milestone for video segmentation and related perception tasks. We are releasing our main model, the dataset, an interactive demo and code. | computer vision, video segmentation, image segmentation | null | 92 | 2408.00714 | [
-0.03322210535407066,
-0.12739214301109314,
0.030255960300564766,
-0.02390757016837597,
0.0696880966424942,
-0.020178457722067833,
-0.01264477614313364,
-0.004089456517249346,
0.025331497192382812,
-0.08456669747829437,
-0.045290134847164154,
-0.03457315266132355,
-0.03390704467892647,
0.08577428013086319,
-0.010564295575022697,
-0.01194017194211483,
0.05061478540301323,
0.05466444790363312,
-0.11721635609865189,
-0.03045518882572651,
0.004492200445383787,
-0.059013597667217255,
0.04962578043341637,
-0.011756549589335918,
0.028390031307935715,
0.07131790369749069,
0.04894913733005524,
-0.07603905349969864,
0.06986434012651443,
-0.03939249366521835,
0.06262166053056717,
-0.032056860625743866,
0.046588145196437836,
0.05564875155687332,
-0.055857520550489426,
-0.01807352714240551,
0.0375579372048378,
-0.029973527416586876,
-0.06594274193048477,
-0.06809361279010773,
-0.013729306869208813,
0.02447383664548397,
0.05468721315264702,
-0.043836429715156555,
0.04070417582988739,
0.01685629040002823,
0.018668703734874725,
-0.07373261451721191,
0.04286244884133339,
-0.000565056805498898,
-0.11592350155115128,
-0.08587342500686646,
-0.020396629348397255,
0.05229329690337181,
-0.07439438253641129,
0.07530178129673004,
0.05933448672294617,
-0.05469046160578728,
0.03710539638996124,
0.019023997709155083,
0.015572368167340755,
-0.04476802051067352,
0.01310831867158413,
0.030449246987700462,
-0.046129219233989716,
-0.02128305658698082,
0.05862019583582878,
-0.06048835441470146,
0.03693237528204918,
-0.04118599370121956,
-0.03375012427568436,
0.09431011974811554,
0.005963626783341169,
0.04404502734541893,
-0.038746319711208344,
-0.07334336638450623,
0.043851301074028015,
-0.04242601618170738,
0.0590767003595829,
-0.06473201513290405,
0.045556116849184036,
-0.062065619975328445,
-0.01659303344786167,
0.04754174128174782,
0.047454141080379486,
-0.04448749125003815,
-0.004400149453431368,
0.049545664340257645,
-0.06514568626880646,
0.04134013503789902,
-0.09683451056480408,
0.00629084836691618,
0.057662222534418106,
-0.03226673975586891,
0.0276964008808136,
0.019261837005615234,
-0.004131135065108538,
-0.1257125586271286,
0.0017577323596924543,
0.03193589672446251,
-0.04161540046334267,
0.03983068838715553,
0.04948113113641739,
-0.02092280611395836,
0.022746412083506584,
0.015893571078777313,
0.036704082041978836,
0.06341460347175598,
0.016664812341332436,
0.026943864300847054,
0.08946622908115387,
0.021129269152879715,
-0.002191650914028287,
-0.025348976254463196,
0.026433704420924187,
-0.0243108831346035,
-0.0025300749111920595,
-0.02539702132344246,
-0.03586886450648308,
0.07623273134231567,
0.03678331524133682,
0.01703394390642643,
-0.07153825461864471,
0.08078162372112274,
0.047530848532915115,
-0.04724094271659851,
-0.05322521552443504,
6.825214992816868e-34,
0.03189389035105705,
-0.02046298421919346,
0.09714657813310623,
0.016216089949011803,
0.06267012655735016,
0.012755737639963627,
0.0196828730404377,
-0.020827708765864372,
-0.029081817716360092,
-0.059407368302345276,
0.0030068252235651016,
-0.054221369326114655,
-0.014866478741168976,
0.09688306599855423,
0.049584317952394485,
-0.08648207038640976,
-0.01321410108357668,
-0.0010965467663481832,
-0.03536955639719963,
0.057433538138866425,
-0.002893090480938554,
-0.031020915135741234,
-0.010267820209264755,
0.00792765524238348,
0.007763288449496031,
0.009392485953867435,
0.04534279182553291,
-0.08175027370452881,
0.04680687189102173,
0.0020294126588851213,
-0.0643380731344223,
-0.00915815681219101,
0.016429845243692398,
0.053379546850919724,
0.032972875982522964,
-0.00521153723821044,
-0.048461273312568665,
-0.03192592039704323,
-0.005631200969219208,
-0.019290097057819366,
0.07797593623399734,
0.036899715662002563,
0.01800720952451229,
-0.03135545551776886,
-0.10061396658420563,
-0.055935006588697433,
0.09632765501737595,
0.14517933130264282,
-0.08657976984977722,
0.06548909097909927,
0.08020778000354767,
0.04526133090257645,
-0.01925223506987095,
-0.0213985126465559,
-0.07305242121219635,
0.01128340233117342,
0.08002892136573792,
0.045891404151916504,
0.006525351200252771,
0.00031793778180144727,
-0.033957138657569885,
0.034893669188022614,
0.0026680640876293182,
0.055698901414871216,
-0.0670221745967865,
-0.07255564630031586,
0.014341408386826515,
0.052685078233480453,
-0.00881518330425024,
0.038624390959739685,
-0.08883973956108093,
0.050069067627191544,
0.04045167565345764,
-0.003800599602982402,
0.04805010184645653,
-0.00038519385270774364,
-0.012856902554631233,
-0.026812436059117317,
-0.016748111695051193,
0.045184116810560226,
-0.027738584205508232,
-0.019607126712799072,
0.029660776257514954,
-0.012169703841209412,
0.049395542591810226,
0.00432969257235527,
0.08226160705089569,
-0.07095938175916672,
-0.07959282398223877,
-0.01796097680926323,
-0.03900865837931633,
-0.013626037165522575,
0.012087429873645306,
0.050752319395542145,
0.0579727441072464,
3.797205141179745e-34,
0.07778003811836243,
0.06586843729019165,
-0.021904069930315018,
0.05954381078481674,
-0.002843475202098489,
-0.08130770176649094,
0.03194551542401314,
0.04831935465335846,
0.0024686448741704226,
-0.10758332163095474,
-0.013011150062084198,
0.006205431651324034,
-0.027461709454655647,
-0.05942902714014053,
0.006629053968936205,
-0.07469616830348969,
0.02967650443315506,
-0.05522690340876579,
-0.03853731229901314,
0.004247339442372322,
0.0029817898757755756,
0.0920695811510086,
-0.04905657097697258,
-0.06394416093826294,
-0.1335468590259552,
0.05751882120966911,
-0.031369224190711975,
0.0809485912322998,
0.026457026600837708,
0.037101250141859055,
0.0540488064289093,
-0.017120234668254852,
-0.0020045337732881308,
-0.0790160745382309,
-0.07987762987613678,
0.0411459244787693,
0.09084248542785645,
-0.039000723510980606,
-0.027097919955849648,
0.038573551923036575,
0.05253778398036957,
-0.058760929852724075,
-0.02729986608028412,
0.08351588249206543,
-0.017908651381731033,
-0.0012582680210471153,
0.004184899386018515,
0.06552893668413162,
-0.14895474910736084,
-0.010889009572565556,
-0.14870436489582062,
0.09669594466686249,
-0.11529767513275146,
-0.01826048083603382,
-0.005886987783014774,
-0.054861608892679214,
-0.08994537591934204,
-0.014673572964966297,
0.04725821688771248,
-0.007528987713158131,
-0.029968108981847763,
0.03304341062903404,
-0.04635557904839516,
0.0103219635784626,
0.021208375692367554,
-0.015967193990945816,
-0.057086702436208725,
-0.01995454914867878,
-0.011574639938771725,
0.013348245061933994,
0.03066098876297474,
-0.04007454589009285,
0.014659330248832703,
0.023912573233246803,
0.01568625308573246,
-0.0007975315093062818,
-0.027380136772990227,
-0.00832068920135498,
0.12356029450893402,
-0.011999820359051228,
-0.03765629604458809,
-0.033335134387016296,
0.007173599675297737,
0.05107537657022476,
0.08579935133457184,
0.11934029310941696,
0.050323985517024994,
0.02335645630955696,
0.008741459809243679,
0.03363218531012535,
-0.09716617316007614,
0.009174016304314137,
0.010784079320728779,
0.036699555814266205,
-0.04379487782716751,
-3.541520143812704e-8,
-0.0711423009634018,
0.018832962960004807,
0.010873755440115929,
0.01833287440240383,
-0.018501613289117813,
-0.04402056336402893,
0.010462014004588127,
0.16251742839813232,
-0.01056663878262043,
-0.04317675158381462,
0.06853500753641129,
-0.0426894947886467,
-0.06755417585372925,
0.018836157396435738,
0.013901292346417904,
0.04653613641858101,
0.060864053666591644,
0.02467304840683937,
-0.029560059309005737,
0.02401348203420639,
-0.010545757599174976,
0.010972159914672375,
0.0253834817558527,
-0.05986421927809715,
0.023428238928318024,
-0.007932486943900585,
0.032922834157943726,
0.10461926460266113,
0.03450711816549301,
-0.05777822062373161,
-0.013727751560509205,
0.015313946641981602,
0.01798138953745365,
0.007021876983344555,
0.045720424503088,
-0.023315919563174248,
0.006923905108124018,
0.02715214155614376,
-0.017909573391079903,
-0.06368622183799744,
0.061750464141368866,
0.016621464863419533,
0.02396603859961033,
0.016312194988131523,
-0.03224446251988411,
0.11001171916723251,
0.03253461420536041,
-0.05107254162430763,
0.03459227457642555,
-0.0006318605737760663,
-0.036219291388988495,
-0.015329677611589432,
-0.043082088232040405,
0.09953749924898148,
0.06334871053695679,
-0.011045234277844429,
0.03760776296257973,
-0.033481426537036896,
0.03975531831383705,
0.07307785004377365,
-0.009183567948639393,
0.02403460443019867,
-0.10011076182126999,
0.016143066808581352
] | https://github.com/facebookresearch/sam2 | 15,178 | 0 | 0 | 0 |
A Computational Framework for Modeling Emergence of Color Vision in the Human Brain | https://openreview.net/forum?id=g3xuCtrG6H | [
"Atsunobu Kotani",
"Ren Ng"
] | Oral | It is a mystery how the brain decodes color vision purely from the optic nerve signals it receives, with a core inferential challenge being how it disentangles internal perception with the correct color dimensionality from the unknown encoding properties of the eye.
In this paper, we introduce a computational framework for modeling this emergence of human color vision by simulating both the eye and the cortex. Existing research often overlooks how the cortex develops color vision or represents color space internally, assuming that the color dimensionality is known a priori; however, we argue that the visual cortex has the capability and the challenge of inferring the color dimensionality purely from fluctuations in the optic nerve signals. To validate our theory, we introduce a simulation engine for biological eyes based on established vision science and generate optic nerve signals resulting from looking at natural images. Further, we propose a bio-plausible model of cortical learning based on self-supervised prediction of optic nerve signal fluctuations under natural eye motions. We show that this model naturally learns to generate color vision by disentangling retinal invariants from the sensory signals. When the retina contains $N$ types of color photoreceptors, our simulation shows that $N$-dimensional color vision naturally emerges, verified through formal colorimetry. Using this framework, we also present the first simulation work that successfully boosts the color dimensionality, as observed in gene therapy on squirrel monkeys, and demonstrates the possibility of enhancing human color vision from 3D to 4D. | color vision, computational neuroscience, retina simulation, cortical learning, self-supervised learning, color blindness | This paper introduces a novel computational framework for modeling the emergence of human color vision by simulating the eye and the cortex. | 20 | 2408.16916 | [
-0.04422564432024956,
-0.12623633444309235,
0.0601772777736187,
0.036014605313539505,
0.008668463677167892,
0.0043794335797429085,
0.09547644108533859,
-0.06428823620080948,
0.0512370690703392,
-0.013999308459460735,
-0.05979665741324425,
-0.1504860520362854,
0.006517889443784952,
0.03517385572195053,
-0.08419067412614822,
0.023908350616693497,
0.008195081725716591,
0.04929168522357941,
-0.036363277584314346,
-0.006389006040990353,
0.05711923912167549,
-0.0754360780119896,
-0.024622999131679535,
0.004478881135582924,
-0.03201868385076523,
0.021877523511648178,
0.1064053624868393,
0.04130729287862778,
-0.0037626386620104313,
0.004677854478359222,
0.057485129684209824,
-0.05960680544376373,
0.0011445478303357959,
0.0010045352391898632,
-0.029657304286956787,
0.012474608607590199,
0.013747557066380978,
-0.040483396500349045,
-0.0417884923517704,
0.03365353122353554,
0.020852256566286087,
0.08074285089969635,
-0.02540886215865612,
0.06603793054819107,
0.12819184362888336,
0.09836571663618088,
0.1393345594406128,
-0.02599628083407879,
-0.05555972084403038,
-0.06407847255468369,
-0.02990569733083248,
0.002509879879653454,
-0.06870754808187485,
0.007454593665897846,
0.023709271103143692,
0.08134859055280685,
0.005915258079767227,
-0.005542369093745947,
0.0016788208158686757,
-0.009585182182490826,
0.05166938155889511,
-0.05350976064801216,
0.01176610216498375,
-0.04180944338440895,
0.005244208965450525,
0.03549956902861595,
-0.005163436755537987,
0.005320800933986902,
0.005289215594530106,
-0.06444843113422394,
-0.024544304236769676,
0.10173580795526505,
-0.017104564234614372,
0.04093438759446144,
0.054428208619356155,
0.06815644353628159,
0.05441925674676895,
0.030285540968179703,
0.03815333917737007,
-0.09062332659959793,
0.044252533465623856,
0.06216059625148773,
0.010786120779812336,
0.031064419075846672,
0.13533031940460205,
0.07589603960514069,
-0.03475717827677727,
0.1016041710972786,
-0.020419808104634285,
-0.0063430131413042545,
-0.05178274214267731,
-0.028965892270207405,
-0.07971645891666412,
-0.07677876204252243,
-0.002410782501101494,
-0.02293979562819004,
0.0395953468978405,
-0.052470613270998,
0.0537262037396431,
-0.02300330251455307,
-0.03343053162097931,
-0.026424167677760124,
0.05549095571041107,
-0.0072813378646969795,
0.04940667375922203,
0.030322466045618057,
0.08064258098602295,
0.0632946714758873,
0.05152016505599022,
-0.024319138377904892,
0.07574769854545593,
0.021321982145309448,
-0.03861653059720993,
0.002758257556706667,
0.036018092185258865,
0.0434899777173996,
-0.017939584329724312,
0.01650073193013668,
0.11448772996664047,
0.029047677293419838,
0.031804054975509644,
-0.0026766255032271147,
-0.07426483929157257,
0.008785668760538101,
0.014432999305427074,
-0.052570246160030365,
-0.04620453715324402,
1.0182782950901674e-34,
-0.020587651059031487,
0.025412920862436295,
0.03134686499834061,
0.03280637413263321,
0.030676906928420067,
-0.0029306018259376287,
-0.004256754647940397,
-0.011290962807834148,
0.013329942710697651,
0.007651896215975285,
-0.0027361188549548388,
0.03476041182875633,
-0.05924161896109581,
0.09669390320777893,
0.04448109120130539,
0.014334997162222862,
-0.042875681072473526,
-0.009132632985711098,
-0.01649271883070469,
-0.057015612721443176,
0.008060693740844727,
0.05285850539803505,
-0.017395243048667908,
-0.033525604754686356,
-0.04088965803384781,
0.02999800443649292,
0.0033638596069067717,
-0.021353406831622124,
0.041814304888248444,
-0.03613551706075668,
-0.052287958562374115,
0.09666755795478821,
-0.004823650233447552,
0.01834740862250328,
-0.003074527019634843,
-0.06298696994781494,
0.04926039278507233,
0.036462679505348206,
-0.018815116956830025,
0.019474999979138374,
-0.02075517736375332,
0.09800175577402115,
0.014601781032979488,
0.0032552562188357115,
-0.046653471887111664,
-0.023356927558779716,
0.05897865071892738,
0.00714294845238328,
-0.05836809054017067,
-0.07227307558059692,
-0.02078285999596119,
-0.04693973809480667,
0.06628259271383286,
-0.10716577619314194,
-0.06639129668474197,
0.004327878821641207,
0.011126533150672913,
-0.02465221658349037,
0.05874764546751976,
0.01447995100170374,
0.0041610137559473515,
-0.018742000684142113,
0.047323085367679596,
0.10161183774471283,
0.026722809299826622,
-0.010240946896374226,
0.0056837634183466434,
-0.03381361439824104,
0.04342056065797806,
-0.004440442658960819,
-0.017673220485448837,
0.04841584339737892,
-0.0717453807592392,
-0.14293013513088226,
-0.019210832193493843,
-0.009136867709457874,
0.07412387430667877,
-0.024548234418034554,
-0.053014177829027176,
0.0020871644373983145,
-0.012943388894200325,
-0.03990838676691055,
-0.04239483177661896,
-0.013246442191302776,
-0.053673092275857925,
0.03408738970756531,
0.05170386657118797,
-0.03319382667541504,
-0.10070911794900894,
-0.08933795988559723,
-0.04131791368126869,
-0.022868622094392776,
0.08808298408985138,
-0.09847287088632584,
-0.02122967690229416,
-1.0198419092066274e-33,
-0.03641868382692337,
0.020684264600276947,
-0.05575994774699211,
0.019879546016454697,
-0.06104017049074173,
-0.028896180912852287,
0.0551350973546505,
0.013757125474512577,
0.001163797453045845,
-0.047882869839668274,
0.008075003512203693,
-0.02381676621735096,
0.026398304849863052,
0.11366602778434753,
-0.023671653121709824,
-0.029092343524098396,
-0.043949194252491,
0.06559332460165024,
-0.013956195674836636,
-0.026287242770195007,
-0.08334062248468399,
0.0837184339761734,
-0.06322694569826126,
-0.05056658387184143,
-0.033476464450359344,
0.09964621812105179,
-0.0334722138941288,
0.08909834176301956,
-0.09079867601394653,
0.04741760715842247,
-0.05078374221920967,
0.032042261213064194,
-0.03641389310359955,
-0.0379968024790287,
-0.044501788914203644,
0.03220093250274658,
-0.01614714413881302,
-0.07980010658502579,
-0.05348756164312363,
0.07616648077964783,
0.0450313426554203,
-0.05247153341770172,
-0.02009355090558529,
0.0047459714114665985,
-0.0036731050349771976,
-0.07870489358901978,
0.01491136010736227,
0.024351777508854866,
-0.0522698238492012,
0.05285628139972687,
-0.03232470527291298,
0.07647375017404556,
-0.12521202862262726,
-0.052320778369903564,
-0.04953862354159355,
0.009867430664598942,
-0.050314344465732574,
0.03513043373823166,
0.14374825358390808,
0.05229400843381882,
-0.06629206240177155,
-0.06976842880249023,
-0.00155721977353096,
0.018622994422912598,
-0.014120562933385372,
0.01947854459285736,
-0.0814451351761818,
0.020824506878852844,
0.09819557517766953,
-0.01658531464636326,
0.044833723455667496,
-0.004389065317809582,
0.003844820661470294,
0.01562109962105751,
0.036549005657434464,
-0.02378835715353489,
0.011741427704691887,
0.04465924948453903,
-0.024089982733130455,
0.07274062931537628,
0.044857289642095566,
-0.07060138136148453,
0.08189169317483902,
0.04903389886021614,
0.01450525876134634,
0.0024592746049165726,
0.030997145920991898,
-0.029768798500299454,
0.018723014742136,
-0.04669187217950821,
-0.020148390904068947,
-0.027668237686157227,
-0.0193520225584507,
0.04708883538842201,
-0.06696102023124695,
-3.580495899768721e-8,
-0.06875794380903244,
0.0021214918233454227,
0.09947790205478668,
0.03310533985495567,
0.07354661822319031,
-0.02344413660466671,
0.031615886837244034,
-0.004176367074251175,
-0.045554470270872116,
0.00113122898619622,
0.043130867183208466,
-0.07180730253458023,
-0.012063929811120033,
0.010321392677724361,
-0.012636679224669933,
0.06444679200649261,
-0.007739747874438763,
-0.008444189094007015,
-0.006085394416004419,
0.01866358332335949,
-0.057176265865564346,
0.025291966274380684,
0.006232651881873608,
0.0008173632668331265,
0.03691178932785988,
-0.0332133024930954,
-0.0907653272151947,
0.02727172151207924,
0.02890777960419655,
0.08150142431259155,
0.07885416597127914,
0.02966020442545414,
0.08270377665758133,
-0.03956473618745804,
-0.05497698113322258,
0.012469764798879623,
-0.045919131487607956,
0.08675821870565414,
-0.027196193113923073,
-0.10793197900056839,
-0.11641784757375717,
0.060081638395786285,
-0.039785608649253845,
0.020064247772097588,
-0.004735097289085388,
-0.04198195040225983,
0.09250064939260483,
-0.06018361821770668,
0.037637848407030106,
0.05798575282096863,
-0.051655229181051254,
0.013008755631744862,
0.024314140900969505,
0.027817843481898308,
0.014830049127340317,
-0.09562002867460251,
-0.04126569628715515,
0.008714437484741211,
-0.05847325176000595,
0.013559669256210327,
-0.003556398442015052,
0.05841177701950073,
-0.02619791403412819,
-0.05549014359712601
] | https://github.com/atsu-kotani/Matisse | 4 | 0 | 0 | 0 |
PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding | https://openreview.net/forum?id=Q6a9W6kzv5 | [
"Wei Chow",
"Jiageng Mao",
"Boyi Li",
"Daniel Seita",
"Vitor Campagnolo Guizilini",
"Yue Wang"
] | Oral | Understanding the physical world is a fundamental challenge in embodied AI, critical for enabling agents to perform complex tasks and operate safely in real-world environments. While Vision-Language Models (VLMs) have shown great promise in reasoning and task planning for embodied agents, their ability to comprehend physical phenomena remains extremely limited.
To close this gap, we introduce PhysBench, a comprehensive benchmark designed to evaluate VLMs' physical world understanding capability across a diverse set of tasks.
PhysBench contains 10,002 entries of interleaved video-image-text data, categorized into four major domains: physical object properties, physical object relationships, physical scene understanding, and physics-based dynamics, further divided into 19 subclasses and 8 distinct capability dimensions.
Our extensive experiments, conducted on 75 representative VLMs, reveal that while these models excel in common-sense reasoning, they struggle with understanding the physical world---likely due to the absence of physical knowledge in their training data and the lack of embedded physical priors.
To tackle the shortfall, we introduce PhysAgent, a novel framework that combines the generalization strengths of VLMs with the specialized expertise of vision models, significantly enhancing VLMs' physical understanding across a variety of tasks, including an 18.4\% improvement on GPT-4o.
Furthermore, our results demonstrate that enhancing VLMs' physical world understanding capabilities can help embodied agents such as MOKA.
We believe that PhysBench and PhysAgent offer valuable insights and contribute to bridging the gap between VLMs and physical world understanding. [Project Page is here](https://physbench.github.io/) | vision-language, multi-modal understanding | We propose PhysBench to evaluate VLMs' physical understanding, highlighting their limitations and introducing PhysAgent to enhance VLMs' physical understanding. | 3 | 2501.16411 | [
0.021364854648709297,
-0.07534443587064743,
0.017973177134990692,
0.014227704145014286,
0.10312778502702713,
-0.005220605991780758,
0.004633531905710697,
-0.017046639695763588,
-0.014926359988749027,
-0.0121281323954463,
-0.013335490599274635,
-0.13654136657714844,
-0.0072039151564240456,
0.11509888619184494,
0.031673792749643326,
-0.017583578824996948,
0.07815388590097427,
0.04232555627822876,
-0.16547192633152008,
0.004699962213635445,
0.023583538830280304,
0.050492897629737854,
0.10568364709615707,
0.00060275912983343,
0.02721666544675827,
0.023330537602305412,
-0.0023666266351938248,
-0.039416417479515076,
-0.02000739984214306,
-0.010920247994363308,
0.08509322255849838,
0.024463826790452003,
0.0007646704325452447,
0.04471956565976143,
0.005064439028501511,
0.08838839083909988,
-0.029382189735770226,
-0.031987302005290985,
-0.03158363327383995,
-0.027305996045470238,
-0.05203857645392418,
-0.011028453707695007,
0.03802390396595001,
-0.01658869907259941,
0.13016492128372192,
0.06945697218179703,
0.009170261211693287,
-0.045415136963129044,
0.03631000965833664,
-0.032657690346241,
-0.10113633424043655,
-0.04490622505545616,
0.016575124114751816,
0.006252292077988386,
-0.016878463327884674,
0.10255322605371475,
0.05771829932928085,
-0.040026549249887466,
0.02360871434211731,
-0.04272877424955368,
0.04417877271771431,
-0.08757100999355316,
0.025372756645083427,
0.026168379932641983,
-0.03287925198674202,
-0.03272987902164459,
0.025924423709511757,
-0.01808706670999527,
0.005230634473264217,
-0.027217010036110878,
-0.047341443598270416,
0.06777501106262207,
0.025716425850987434,
-0.001784652704373002,
0.020796671509742737,
-0.024746226146817207,
-0.004180533345788717,
-0.03011946566402912,
0.048537127673625946,
-0.09752676635980606,
0.014177186414599419,
-0.01201317086815834,
0.00913936272263527,
0.07003959268331528,
0.10943922400474548,
0.0726475939154625,
0.028329450637102127,
0.10100118815898895,
-0.03518487140536308,
0.00776549382135272,
-0.06631983071565628,
-0.11764383316040039,
0.036394909024238586,
0.045656412839889526,
0.02519998326897621,
0.031739167869091034,
0.0060437628999352455,
-0.15628229081630707,
-0.037649765610694885,
0.04585998132824898,
-0.01336263120174408,
0.016396142542362213,
0.041782207787036896,
0.05025586485862732,
0.04127594828605652,
0.002985103288665414,
0.057386502623558044,
0.06754807382822037,
-0.014795308001339436,
-0.058089498430490494,
0.050688959658145905,
0.0006369408220052719,
0.031716275960206985,
0.04288515821099281,
-0.007001261226832867,
-0.0035105023998767138,
0.02907458320260048,
0.020693670958280563,
0.01863371580839157,
-0.021969590336084366,
0.016153067350387573,
-0.02415529638528824,
0.05384582281112671,
0.05168933793902397,
0.03486790880560875,
-0.039140596985816956,
0.020817117765545845,
9.818900638998721e-34,
0.002029943512752652,
0.005070932675153017,
0.1198999211192131,
0.032389894127845764,
0.07380083203315735,
-0.02793581783771515,
-0.024980327114462852,
0.015047021210193634,
-0.049981944262981415,
0.018659232184290886,
-0.04959006607532501,
0.020794766023755074,
-0.0570027232170105,
0.10424546152353287,
0.03443942591547966,
0.017749443650245667,
-0.07150942087173462,
0.056259915232658386,
-0.051741041243076324,
0.05043961852788925,
0.06239936500787735,
-0.019505197182297707,
-0.019296634942293167,
-0.007274087518453598,
0.013286689296364784,
0.06428617238998413,
0.04089920222759247,
-0.025279218330979347,
-0.038892120122909546,
-0.02680225484073162,
-0.13096144795417786,
0.03022523783147335,
-0.0000240141453105025,
0.025418121367692947,
-0.012973648495972157,
-0.048850733786821365,
-0.026476731523871422,
-0.10427575558423996,
-0.014248741790652275,
-0.01065720897167921,
0.00006011111690895632,
0.04747324436903,
-0.030368411913514137,
-0.059939153492450714,
-0.039820604026317596,
-0.019028378650546074,
0.028938068076968193,
0.10826843976974487,
-0.09226834028959274,
-0.0061745066195726395,
-0.0170678049325943,
0.02172969840466976,
0.020242905244231224,
-0.08830355852842331,
-0.026936620473861694,
-0.04041662812232971,
0.07457021623849869,
0.0013596845092251897,
0.032747648656368256,
0.004624433349817991,
-0.011989587917923927,
0.08516301959753036,
0.009948006831109524,
0.10014356672763824,
-0.004187865182757378,
0.01416817307472229,
-0.006606655661016703,
0.09370052069425583,
0.014338048174977303,
0.007150481455028057,
-0.05289091914892197,
0.03396783396601677,
-0.00251781870611012,
-0.06045512482523918,
0.12056206911802292,
-0.008614195510745049,
0.04996276646852493,
-0.1462685763835907,
-0.023740950971841812,
0.05815443769097328,
-0.03257417678833008,
-0.02507113479077816,
-0.011742951348423958,
-0.03311078995466232,
-0.04911579564213753,
-0.03744133934378624,
0.03388750180602074,
-0.13628621399402618,
-0.024332020431756973,
-0.08195219933986664,
-0.11426513642072678,
-0.0923924520611763,
-0.029024938121438026,
-0.027946097776293755,
-0.022928684949874878,
-9.327405328755806e-34,
-0.006773282773792744,
-0.01972142606973648,
-0.06187668815255165,
0.06817184388637543,
0.01481742225587368,
-0.1148732379078865,
-0.039380721747875214,
0.0034151007421314716,
-0.033789243549108505,
-0.09440770000219345,
0.012865923345088959,
-0.05021149665117264,
-0.05610311031341553,
0.00014795130118727684,
0.019776422530412674,
0.042691171169281006,
-0.034268639981746674,
-0.06513015180826187,
0.03757665678858757,
0.0023777924943715334,
0.022690406069159508,
0.06280534714460373,
-0.05568954721093178,
0.0024163364432752132,
-0.03335181251168251,
0.07498464733362198,
-0.015053601935505867,
0.010803976096212864,
-0.025883423164486885,
0.005100003909319639,
-0.027731481939554214,
-0.01720261014997959,
-0.009670600295066833,
-0.003854830516502261,
-0.04638442397117615,
0.028724275529384613,
0.050126101821660995,
-0.03894779458642006,
-0.04125377908349037,
0.04222454875707626,
0.08224917203187943,
-0.07142245769500732,
-0.10312183201313019,
-0.02578929252922535,
-0.056308068335056305,
0.02881568670272827,
-0.04981737583875656,
-0.0033017550595104694,
-0.02804979868233204,
0.003918992355465889,
-0.01748814433813095,
0.03647425398230553,
-0.03794633969664574,
-0.016706369817256927,
-0.015618088655173779,
-0.018361113965511322,
-0.028018197044730186,
0.016518468037247658,
0.0445253923535347,
-0.04767439141869545,
-0.0052082836627960205,
0.0014703319175168872,
-0.02142915688455105,
0.07862227410078049,
-0.014703630469739437,
-0.0022970857098698616,
-0.07224351912736893,
0.04303491860628128,
0.029509305953979492,
-0.054788738489151,
0.013635664246976376,
-0.05487898737192154,
0.028634127229452133,
0.05580073595046997,
0.04430140182375908,
-0.014118892140686512,
-0.013205401599407196,
-0.001451281481422484,
0.04188571125268936,
0.008318085223436356,
0.04048728570342064,
-0.04412885010242462,
0.09643331915140152,
0.10542158037424088,
0.06297966092824936,
0.08138260990381241,
-0.04515305161476135,
0.010678427293896675,
0.027344658970832825,
0.05515701323747635,
-0.041837118566036224,
0.06895753741264343,
-0.02418188378214836,
0.0692608654499054,
-0.11703895777463913,
-4.126372132873257e-8,
-0.07102487981319427,
0.037170860916376114,
0.03512777015566826,
-0.039816562086343765,
-0.029522676020860672,
-0.012864555232226849,
-0.03176148980855942,
0.04020559415221214,
-0.04017788544297218,
-0.008858872577548027,
0.03697209060192108,
-0.07594231516122818,
-0.006377247162163258,
0.041313134133815765,
0.005487678572535515,
0.12112250179052353,
-0.03203776851296425,
-0.00607579480856657,
-0.020005516707897186,
0.03216481953859329,
0.07163424789905548,
0.009732772596180439,
0.014645933173596859,
-0.015337509103119373,
-0.013896295800805092,
-0.05314845219254494,
-0.10802270472049713,
0.08377490937709808,
0.042844805866479874,
-0.026328567415475845,
0.03803236410021782,
-0.013563430868089199,
0.06059176102280617,
0.0030141097959131002,
0.01924785226583481,
0.03673217445611954,
-0.04184088855981827,
-0.04669367894530296,
-0.02429891750216484,
0.009168307296931744,
-0.009208142757415771,
0.10256746411323547,
-0.030011070892214775,
0.01815143972635269,
0.0936167761683464,
0.005601505748927593,
-0.03847618028521538,
-0.10031891614198685,
0.05860235542058945,
0.05053839460015297,
-0.03763093054294586,
0.04682435840368271,
-0.029689054936170578,
0.14343689382076263,
0.020093809813261032,
0.05314461514353752,
-0.007808806374669075,
-0.04476583003997803,
0.04051344096660614,
0.03969312086701393,
0.007005675230175257,
0.026545628905296326,
-0.057870976626873016,
0.014435202814638615
] | 0 | 0 | 0 | 0 |
|
Beyond Random Masking: When Dropout meets Graph Convolutional Networks | https://openreview.net/forum?id=PwxYoMvmvy | [
"Yuankai Luo",
"Xiao-Ming Wu",
"Hao Zhu"
] | Poster | Graph Convolutional Networks (GCNs) have emerged as powerful tools for learning on graph-structured data, yet the behavior of dropout in these models remains poorly understood. This paper presents a comprehensive theoretical analysis of dropout in GCNs, revealing that its primary role differs fundamentally from standard neural networks - preventing oversmoothing rather than co-adaptation. We demonstrate that dropout in GCNs creates dimension-specific stochastic sub-graphs, leading to a form of structural regularization not present in standard neural networks. Our analysis shows that dropout effects are inherently degree-dependent, resulting in adaptive regularization that considers the topological importance of nodes. We provide new insights into dropout's role in mitigating oversmoothing and derive novel generalization bounds that account for graph-specific dropout effects. Furthermore, we analyze the synergistic interaction between dropout and batch normalization in GCNs, uncovering a mechanism that enhances overall regularization. Our theoretical findings are validated through extensive experiments on both node-level and graph-level tasks across 14 datasets. Notably, GCN with dropout and batch normalization outperforms state-of-the-art methods on several benchmarks, demonstrating the practical impact of our theoretical insights. | Graph neural networks, Dropout | null | 14,284 | null | [
-0.03459398075938225,
-0.03768260031938553,
0.09596733003854752,
0.06077318266034126,
0.08853375166654587,
0.027527999132871628,
-0.040357958525419235,
-0.05527006462216377,
-0.06014551222324371,
-0.09288386255502701,
-0.0400095209479332,
0.010113815777003765,
0.0026089607272297144,
0.025238456204533577,
-0.01025064941495657,
-0.04816802218556404,
0.04677624627947807,
0.04399922490119934,
-0.14396074414253235,
-0.04388710483908653,
0.029540754854679108,
-0.01463411282747984,
0.013037036173045635,
0.010607868432998657,
0.1114254891872406,
-0.025012604892253876,
-0.011390733532607555,
-0.07254199683666229,
0.05422845482826233,
-0.03658948466181755,
0.053042177110910416,
0.006014939863234758,
-0.016929127275943756,
0.034686170518398285,
-0.015475593507289886,
0.009428373537957668,
-0.01416221633553505,
0.05671268701553345,
0.06161043792963028,
0.006672274321317673,
0.03254083916544914,
0.043632373213768005,
0.030681675300002098,
0.08518286794424057,
0.026379821822047234,
0.009307458065450191,
-0.022587673738598824,
-0.06816506385803223,
0.006359349004924297,
-0.01129101775586605,
0.01952195353806019,
0.0031874242704361677,
-0.04041125252842903,
0.06996635347604752,
0.005134951788932085,
0.07825464755296707,
0.010203152894973755,
0.04410495609045029,
0.00042011411278508604,
0.05663461610674858,
0.03156561031937599,
-0.12618184089660645,
-0.05468197539448738,
-0.007789907045662403,
0.0003178452025167644,
-0.0023475780617445707,
0.045414261519908905,
0.019999781623482704,
-0.03504738211631775,
0.09382273256778717,
0.04207083210349083,
0.07499013841152191,
-0.07532443106174469,
0.08010878413915634,
0.07154755294322968,
0.05665815249085426,
0.009984727017581463,
-0.02748009003698826,
0.0022660591639578342,
-0.051959045231342316,
0.06316433101892471,
0.03616221994161606,
0.05244892090559006,
-0.0010567454155534506,
0.06136766076087952,
-0.024985942989587784,
-0.02629133127629757,
-0.038895100355148315,
-0.03438043221831322,
-0.005035633686929941,
-0.07891813665628433,
-0.008483467623591423,
-0.03597721830010414,
-0.03560963273048401,
0.019862793385982513,
0.01885300688445568,
0.04447227343916893,
-0.10026293992996216,
-0.020780853927135468,
0.08641624450683594,
-0.0266514141112566,
-0.019842715933918953,
0.030192440375685692,
0.03185225650668144,
0.030237117782235146,
0.020865119993686676,
0.054481834173202515,
0.0345301479101181,
0.046244267374277115,
-0.03911226615309715,
0.006069491151720285,
0.12212115526199341,
0.026394100859761238,
0.03607439994812012,
-0.016721507534384727,
-0.039202868938446045,
0.05698951706290245,
-0.06183933466672897,
0.01626577600836754,
0.04165258631110191,
-0.07455192506313324,
0.057443343102931976,
-0.009916062466800213,
-0.05270181596279144,
-0.05610739067196846,
0.05424842610955238,
-0.062145113945007324,
4.675290008105638e-33,
0.0370728000998497,
0.0011724496725946665,
0.010745881125330925,
-0.04498394951224327,
0.08778239041566849,
-0.009699120186269283,
-0.07550425082445145,
-0.0155539745464921,
-0.009946770034730434,
0.022120153531432152,
-0.04650058224797249,
-0.02530239149928093,
-0.06123760715126991,
0.07133758068084717,
0.0020151124335825443,
0.0316992923617363,
0.046141210943460464,
0.041288331151008606,
0.04055032134056091,
-0.1195160448551178,
-0.0010357500286772847,
-0.03627197816967964,
-0.0027146164793521166,
0.04733489826321602,
0.0006651597796007991,
0.006087927147746086,
-0.045956745743751526,
-0.0019385189516469836,
-0.003924515098333359,
-0.014197273179888725,
-0.05510474368929863,
0.10401324927806854,
0.056897055357694626,
0.039974603801965714,
0.01086525060236454,
-0.03859247639775276,
-0.0011297909077256918,
-0.05179435759782791,
-0.003250295529142022,
-0.07527437061071396,
-0.049967601895332336,
0.058189213275909424,
-0.028896387666463852,
-0.026031438261270523,
-0.05551427975296974,
-0.09711116552352905,
0.043006427586078644,
-0.11540370434522629,
-0.06560872495174408,
-0.08240356296300888,
-0.01908957213163376,
0.015146725811064243,
-0.04782763123512268,
0.004876403603702784,
0.0021298525389283895,
-0.03395010158419609,
0.10127503424882889,
0.0012516845017671585,
0.04172363132238388,
0.022572724148631096,
0.027882156893610954,
0.0370447114109993,
-0.010728317312896252,
0.04169822856783867,
0.047353800386190414,
0.052133094519376755,
-0.05411051586270332,
0.017632054165005684,
0.023052595555782318,
0.011418773792684078,
-0.056372568011283875,
0.07335485517978668,
-0.10980778187513351,
-0.05794570967555046,
-0.00064182432834059,
0.012941628694534302,
0.0073538185097277164,
-0.1477256417274475,
-0.023806730285286903,
0.04231086000800133,
-0.039571695029735565,
-0.010196996852755547,
0.011993944644927979,
-0.055723194032907486,
-0.10021347552537918,
-0.04316289350390434,
0.09567379951477051,
-0.08270780742168427,
0.04779874160885811,
-0.041387252509593964,
-0.04867223650217056,
-0.049059052020311356,
-0.00826248899102211,
-0.03260434418916702,
0.047078169882297516,
-4.532205529645687e-33,
-0.06811310350894928,
0.13473673164844513,
-0.02259122021496296,
0.06414016336202621,
0.007875490002334118,
0.01092615444213152,
0.023482132703065872,
-0.02648516371846199,
-0.01574045978486538,
-0.020258678123354912,
0.003919608425348997,
-0.04569694772362709,
0.0029614402446895838,
-0.04144807532429695,
0.032037798315286636,
-0.03101285733282566,
-0.022775446996092796,
0.02618953213095665,
-0.02101520448923111,
-0.029495401307940483,
0.06979908794164658,
0.05031326785683632,
-0.056040577590465546,
0.028522737324237823,
0.0857822448015213,
0.0027392134070396423,
-0.12163297832012177,
0.06247500702738762,
0.0360315777361393,
-0.037523943930864334,
-0.09327566623687744,
0.009106160141527653,
-0.03777064383029938,
-0.041432350873947144,
0.04312613233923912,
0.07240329682826996,
-0.03315996378660202,
0.08605770021677017,
-0.03601466119289398,
-0.013605363667011261,
-0.0015824375441297889,
-0.02610415779054165,
-0.0991247370839119,
0.06319791823625565,
0.049942050129175186,
-0.016772441565990448,
-0.02517026849091053,
0.00861973688006401,
-0.10315853357315063,
0.0711124837398529,
-0.06302108615636826,
0.03492288291454315,
-0.04808174446225166,
0.03516551852226257,
-0.05777779966592789,
-0.04268375784158707,
0.08142980188131332,
0.032719749957323074,
-0.00045099586714059114,
0.054186347872018814,
-0.04379762336611748,
-0.07528945058584213,
-0.03502423316240311,
-0.015600400976836681,
-0.025783728808164597,
-0.022939784452319145,
0.019889459013938904,
0.04333730787038803,
0.04009556397795677,
0.07172956317663193,
0.01936107873916626,
-0.017035169526934624,
-0.026358187198638916,
-0.015325195156037807,
-0.02545745298266411,
-0.012967023067176342,
-0.021718306466937065,
0.005283552221953869,
-0.06723597645759583,
0.03206857293844223,
0.004413374699652195,
0.027095716446638107,
-0.019452879205346107,
0.03179037198424339,
0.12580686807632446,
0.06463989615440369,
-0.06119680032134056,
0.05906016007065773,
0.06891153007745743,
0.06179191544651985,
-0.01698138192296028,
-0.02504662424325943,
-0.07107583433389664,
0.019934186711907387,
-0.0688672661781311,
-4.2348034412498237e-8,
-0.14300113916397095,
0.0348881296813488,
-0.026983866468071938,
0.027148550376296043,
0.0829315334558487,
-0.0524122379720211,
0.05870741605758667,
0.10381222516298294,
-0.01979396678507328,
0.02858417108654976,
0.047205887734889984,
-0.016721433028578758,
-0.021181894466280937,
-0.06674087047576904,
0.013844151049852371,
0.031549375504255295,
-0.0017870228039100766,
0.03637596592307091,
0.03803097456693649,
0.0423339381814003,
-0.03933612257242203,
-0.057930588722229004,
-0.033003512769937515,
-0.02405574358999729,
-0.005037299823015928,
-0.1051248237490654,
0.0037932805716991425,
0.12186538428068161,
-0.00024977364228107035,
0.012364029884338379,
-0.022165372967720032,
0.054523732513189316,
0.1080324575304985,
-0.04146735370159149,
0.01801440119743347,
0.12164261937141418,
0.06489454209804535,
0.013104546815156937,
-0.050494737923145294,
0.005182111635804176,
-0.0038434171583503485,
0.04170920327305794,
0.01410693023353815,
-0.016744911670684814,
-0.07016146183013916,
0.053910594433546066,
0.05776752904057503,
0.029873227700591087,
0.09079574048519135,
-0.06972011923789978,
0.035426266491413116,
-0.039830051362514496,
-0.0005263570346869528,
0.010863417759537697,
0.066848523914814,
0.013837904669344425,
-0.0016632615588605404,
-0.044130172580480576,
0.011664445511996746,
0.044998910278081894,
-0.04671334847807884,
-0.030149897560477257,
-0.1048264130949974,
-0.05388448387384415
] | 0 | 0 | 0 | 0 |
|
Self-supervised contrastive learning performs non-linear system identification | https://openreview.net/forum?id=ONfWFluZBI | [
"Rodrigo González Laiz",
"Tobias Schmidt",
"Steffen Schneider"
] | Poster | Self-supervised learning (SSL) approaches have brought tremendous success across many tasks and domains. It has been argued that these successes can be attributed to a link between SSL and identifiable representation learning: Temporal structure and auxiliary variables ensure that latent representations are related to the true underlying generative factors of the data. Here, we deepen this connection and show that SSL can perform system identification in latent space. We propose DynCL, a framework to uncover linear, switching linear and non-linear dynamics under a non-linear observation model, give theoretical guarantees and validate them empirically. | system identification, dynamics learning, identifiability, self-supervised learning | null | 14,280 | 2410.14673 | [
-0.09598539769649506,
-0.11692765355110168,
-0.004986651707440615,
0.052883587777614594,
0.020885517820715904,
-0.0005877905059605837,
-0.019151510670781136,
-0.08535099029541016,
0.08011952042579651,
-0.04150418937206268,
-0.05675389617681503,
-0.005256868898868561,
0.009540311992168427,
0.06601879000663757,
-0.06891115754842758,
-0.04121682047843933,
0.07469840347766876,
0.037294574081897736,
-0.021950893104076385,
-0.04512743279337883,
0.024082712829113007,
0.07005123794078827,
-0.04094377160072327,
0.06458784639835358,
0.05807926133275032,
0.02518821693956852,
0.041257094591856,
0.020982740446925163,
-0.0012274202890694141,
-0.07970274239778519,
0.05696793645620346,
0.047578442841768265,
-0.043391916900873184,
-0.026720931753516197,
-0.016744175925850868,
0.03226251155138016,
-0.012996191158890724,
0.02951078675687313,
-0.08006619662046432,
-0.03200593963265419,
0.029275307431817055,
-0.015203190967440605,
-0.016235744580626488,
0.04266364872455597,
0.00041029421845451,
-0.04706348106265068,
0.008112006820738316,
-0.1087120845913887,
-0.031398653984069824,
-0.014752696268260479,
-0.08516674488782883,
0.031101610511541367,
0.0067631700076162815,
0.043092288076877594,
-0.025375233963131905,
0.028621748089790344,
0.08202427625656128,
0.06322047114372253,
0.01780731976032257,
0.06787422299385071,
0.05433344841003418,
-0.010593519546091557,
-0.00348892854526639,
0.006720268167555332,
0.018942464143037796,
0.12803016602993011,
0.0012052112724632025,
0.04353439062833786,
-0.021016882732510567,
-0.03703509643673897,
-0.057401902973651886,
0.07301532477140427,
-0.042960066348314285,
0.030682405456900597,
0.0845453292131424,
-0.09067922085523605,
0.021654773503541946,
-0.010471486486494541,
0.11817938089370728,
-0.06581103056669235,
0.009133854880928993,
0.06089169159531593,
-0.04328779876232147,
0.08465775847434998,
0.08920496702194214,
0.043771352618932724,
-0.03336211293935776,
0.019505834206938744,
-0.027562225237488747,
-0.013314135372638702,
0.0074853417463600636,
-0.04493270814418793,
-0.07243727147579193,
-0.026617707684636116,
0.03170044347643852,
-0.03199734538793564,
0.09774862974882126,
-0.014971439726650715,
0.07197168469429016,
0.06050475314259529,
-0.028940780088305473,
0.003930860664695501,
0.02504892833530903,
0.08955749869346619,
0.0033807549625635147,
-0.03923914581537247,
0.1158197820186615,
-0.003933315631002188,
0.007205743808299303,
-0.11212518811225891,
0.03528270125389099,
-0.008411115035414696,
-0.02717592939734459,
0.050571490079164505,
0.029482176527380943,
-0.07257603853940964,
0.029779896140098572,
-0.03804968297481537,
0.0290616936981678,
-0.04364214837551117,
-0.010635944083333015,
-0.034954868257045746,
0.06654410809278488,
0.06130441650748253,
0.041053321212530136,
0.042873188853263855,
-0.06303926557302475,
5.161714621428189e-34,
0.00803567748516798,
0.013934175483882427,
0.012518013827502728,
-0.04935887083411217,
-0.04652982950210571,
-0.05393846705555916,
-0.00005157989653525874,
-0.0365060456097126,
0.03812362253665924,
0.018691403791308403,
0.03332330286502838,
0.0906708613038063,
-0.01948133297264576,
0.01409169565886259,
0.020473994314670563,
0.00856225285679102,
-0.0633273646235466,
0.030814161524176598,
0.0402463935315609,
-0.046444229781627655,
0.021319948136806488,
0.0627400204539299,
-0.012355922721326351,
-0.09144461154937744,
0.014318948611617088,
0.004885378759354353,
0.013107644394040108,
0.007165062241256237,
0.009785660542547703,
0.01976878196001053,
-0.023815348744392395,
0.03604481369256973,
-0.026736587285995483,
0.030260300263762474,
0.04384360834956169,
0.04700203239917755,
0.016006940975785255,
-0.015279090963304043,
0.04796500876545906,
-0.059414997696876526,
0.08432373404502869,
-0.0035314366687089205,
-0.024416768923401833,
-0.03554682061076164,
-0.12229951471090317,
-0.045036982744932175,
0.03159181401133537,
-0.03437681868672371,
-0.03367861732840538,
-0.048744019120931625,
-0.024629641324281693,
0.061264317482709885,
-0.010391946882009506,
-0.12182488292455673,
-0.012974021956324577,
0.006715656258165836,
-0.030680544674396515,
0.09650890529155731,
0.057773657143116,
0.0809369906783104,
-0.05337679013609886,
-0.022548841312527657,
-0.014625594019889832,
0.01442220713943243,
0.03179680183529854,
0.032741788774728775,
0.006614021956920624,
0.03799908608198166,
0.06771957129240036,
-0.053091805428266525,
-0.0027231641579419374,
-0.011161919683218002,
-0.05545177310705185,
-0.06996701657772064,
0.12598173320293427,
-0.05126112699508667,
0.028197094798088074,
-0.024026846513152122,
-0.014798540621995926,
-0.016027644276618958,
0.024946440011262894,
-0.060670118778944016,
0.009393995627760887,
0.02530522271990776,
-0.07681668549776077,
0.009029553271830082,
-0.001485001528635621,
-0.08120555430650711,
-0.09489677846431732,
0.011118856258690357,
-0.012640421278774738,
-0.058777615427970886,
0.05311978980898857,
0.12342387437820435,
0.009507724083960056,
-2.1629070341191166e-33,
-0.06961085647344589,
0.02795889601111412,
-0.018841808661818504,
-0.0008539388072676957,
0.04273896664381027,
-0.0405864343047142,
0.003263574093580246,
0.05604569986462593,
-0.010120706632733345,
0.0236421599984169,
0.06645820289850235,
-0.06351673603057861,
-0.03074815683066845,
0.049146901816129684,
0.05429190769791603,
0.04676489531993866,
0.0137915825471282,
0.007420484442263842,
0.03742435947060585,
-0.029522638767957687,
-0.021853985264897346,
0.09799856692552567,
-0.006674846634268761,
0.07753583788871765,
0.004922159481793642,
0.0292240921407938,
-0.03726901486515999,
0.044875435531139374,
-0.00854126363992691,
-0.057870373129844666,
0.00301870028488338,
-0.0029924409464001656,
-0.030956188216805458,
-0.02443733811378479,
-0.09663736075162888,
0.06841349601745605,
-0.03747669979929924,
-0.020937945693731308,
-0.03275419399142265,
0.02863042987883091,
0.057147134095430374,
-0.023532439023256302,
-0.056581590324640274,
-0.01047001127153635,
0.019354883581399918,
-0.06786532700061798,
-0.02516135945916176,
0.07696516066789627,
0.07147467881441116,
0.05005430057644844,
-0.01078016310930252,
0.060074273496866226,
-0.05964845418930054,
-0.0532512404024601,
-0.05492242053151131,
-0.03056159056723118,
0.1253388226032257,
0.01242563035339117,
0.07713695615530014,
0.04332563281059265,
-0.07535094022750854,
-0.06394124776124954,
0.037591613829135895,
-0.020618624985218048,
0.018219953402876854,
-0.15975771844387054,
-0.04515225440263748,
-0.007712087128311396,
0.05318813398480415,
-0.03982515260577202,
0.060611385852098465,
0.005757505539804697,
-0.06322018057107925,
-0.009018069133162498,
0.023139066994190216,
-0.13648942112922668,
0.0049801599234342575,
-0.01974710077047348,
-0.04836808890104294,
-0.03843095898628235,
0.05260901153087616,
0.011478623375296593,
0.06094159185886383,
0.007366636767983437,
-0.018970055505633354,
0.07656842470169067,
0.018566634505987167,
-0.05931337550282478,
-0.00321038905531168,
-0.012260433286428452,
-0.01558778714388609,
0.0014610890066251159,
-0.1285162717103958,
0.05075125768780708,
0.010449748486280441,
-3.752380095534136e-8,
-0.0752289667725563,
0.021842217072844505,
0.04381183907389641,
0.03775390610098839,
0.033124666661024094,
-0.016121575608849525,
0.01750251278281212,
0.04982665553689003,
0.005805606953799725,
0.01630282960832119,
0.10734770447015762,
-0.049597036093473434,
0.014766623266041279,
-0.1264379769563675,
-0.02670246921479702,
0.0330805703997612,
0.02441064454615116,
0.007060231640934944,
0.022110117599368095,
-0.020865842700004578,
0.09391019493341446,
-0.030255412682890892,
-0.0432785302400589,
-0.05455085262656212,
0.027763256803154945,
-0.09318819642066956,
-0.07087267935276031,
0.05632845312356949,
0.02991786040365696,
-0.015433226712048054,
-0.04217725992202759,
0.0631496012210846,
-0.02081885188817978,
-0.02705814316868782,
0.022940926253795624,
0.05999765917658806,
0.025396496057510376,
-0.0531308576464653,
-0.0238144900649786,
0.01144159585237503,
-0.055557217448949814,
0.1071576401591301,
-0.0747772753238678,
0.004441143479198217,
-0.058276768773794174,
0.046786606311798096,
-0.026604386046528816,
-0.0986533984541893,
0.05415157601237297,
-0.023995541036128998,
0.07036108523607254,
-0.1375093162059784,
0.02750135213136673,
0.05562344565987587,
0.07115982472896576,
0.00523363845422864,
0.0612541027367115,
-0.015776218846440315,
0.00046149129047989845,
0.07411086559295654,
-0.0018228640547022223,
0.05073926970362663,
-0.022047504782676697,
-0.018048102036118507
] | https://github.com/dynamical-inference/dyncl | 10 | 0 | 0 | 0 |
Sparse autoencoders reveal selective remapping of visual concepts during adaptation | https://openreview.net/forum?id=imT03YXlG2 | [
"Hyesu Lim",
"Jinho Choi",
"Jaegul Choo",
"Steffen Schneider"
] | Poster | Adapting foundation models for specific purposes has become a standard approach to build machine learning systems for downstream applications. Yet, it is an open question which mechanisms take place during adaptation. Here we develop a new Sparse Autoencoder (SAE) for the CLIP vision transformer, named PatchSAE, to extract interpretable concepts at granular levels (e.g., shape, color, or semantics of an object) and their patch-wise spatial attributions. We explore how these concepts influence the model output in downstream image classification tasks and investigate how recent state-of-the-art prompt-based adaptation techniques change the association of model inputs to these concepts. While activations of concepts slightly change between adapted and non-adapted models, we find that the majority of gains on common adaptation tasks can be explained with the existing concepts already present in the non-adapted foundation model. This work provides a concrete framework to train and use SAEs for Vision Transformers and provides insights into explaining adaptation mechanisms. | interpretability, vision-language models, sparse autoencoder, adaptation | null | 14,240 | 2412.05276 | [
0.0022846655920147896,
-0.09708346426486969,
0.013985211960971355,
0.018972227349877357,
0.16626137495040894,
-0.0013961922377347946,
0.006066768430173397,
-0.03756232559680939,
0.01637984812259674,
-0.043895870447158813,
-0.07324789464473724,
-0.08336096256971359,
0.014544513076543808,
0.02268441952764988,
-0.03504621237516403,
-0.019791072234511375,
0.04983479157090187,
0.1079307273030281,
-0.10574664920568466,
-0.05945883318781853,
0.023573368787765503,
-0.030778024345636368,
0.028523052111268044,
0.04600819572806358,
0.06055152416229248,
0.007142581511288881,
-0.0011381176300346851,
-0.009947560727596283,
0.026935752481222153,
-0.04979298263788223,
0.0035939901135861874,
0.011226868256926537,
-0.015281178057193756,
0.09218655526638031,
-0.050010692328214645,
0.04451437294483185,
-0.0011864107800647616,
0.11839651316404343,
-0.027154438197612762,
-0.024660324677824974,
-0.0003458200371824205,
-0.03833732753992081,
-0.013046957552433014,
-0.020190618932247162,
0.018168307840824127,
-0.014198051765561104,
0.01780538074672222,
-0.09400133788585663,
-0.045969802886247635,
-0.03075985051691532,
-0.01101983804255724,
-0.02381610870361328,
-0.08982142806053162,
0.0021664423402398825,
-0.05606886371970177,
0.05926715210080147,
0.08133899420499802,
-0.0011987753678113222,
0.029901618137955666,
-0.007587716914713383,
-0.01218322105705738,
-0.08864214271306992,
0.016603535041213036,
0.08656392991542816,
-0.005043592303991318,
-0.047613732516765594,
0.014116397127509117,
0.02999209426343441,
-0.038227200508117676,
-0.0826701819896698,
-0.016375217586755753,
0.05695387348532677,
-0.0016301736468449235,
0.006138964090496302,
0.09082074463367462,
-0.003034730441868305,
0.021051030606031418,
0.017672277987003326,
0.024388279765844345,
-0.06470317393541336,
0.013960583135485649,
0.04580189287662506,
-0.002501256065443158,
0.029579995200037956,
0.12925928831100464,
0.01590139977633953,
-0.004651010036468506,
0.06245054304599762,
-0.0324234701693058,
0.019480852410197258,
-0.043812867254018784,
-0.13173027336597443,
0.060314372181892395,
-0.011178684420883656,
-0.0049685831181705,
0.010964248329401016,
-0.011808979324996471,
-0.15793964266777039,
-0.031650152057409286,
0.0583219937980175,
-0.02407396025955677,
0.021566715091466904,
0.030064212158322334,
0.023868385702371597,
0.03748450055718422,
-0.02453605830669403,
0.07716186344623566,
0.05549050495028496,
0.004277690313756466,
-0.03063502535223961,
0.0674060508608818,
0.07248062640428543,
0.043435271829366684,
-0.026378361508250237,
-0.009050051681697369,
-0.03238408640027046,
-0.015282120555639267,
-0.10733344405889511,
-0.006319917272776365,
0.00015524320770055056,
-0.005669604521244764,
0.0010994687909260392,
-0.06754834204912186,
-0.009548279456794262,
0.056792452931404114,
-0.004654829856008291,
-0.05703386291861534,
5.308226676111244e-34,
-0.004296918399631977,
0.047278232872486115,
0.07323887199163437,
0.029933596029877663,
0.045964982360601425,
-0.0760025903582573,
0.030545733869075775,
-0.003696987172588706,
0.015654975548386574,
0.0041454993188381195,
-0.026778345927596092,
0.05832682549953461,
-0.030293269082903862,
0.1483820676803589,
-0.024313660338521004,
-0.030661815777420998,
0.0002503452997189015,
0.03862787038087845,
0.0007308595813810825,
0.01934792846441269,
-0.060437291860580444,
0.014852208085358143,
-0.03981618955731392,
-0.010642526671290398,
0.04302890971302986,
-0.04033048823475838,
0.026508113369345665,
-0.011862093582749367,
-0.051422663033008575,
0.015934642404317856,
-0.0629412829875946,
0.08020734041929245,
0.08250483870506287,
0.007987933233380318,
0.006472813431173563,
-0.06783847510814667,
0.05694425478577614,
-0.01051408238708973,
0.01705726981163025,
-0.03154052421450615,
0.019356902688741684,
0.02593575417995453,
-0.023522958159446716,
-0.053740244358778,
-0.02148672379553318,
-0.044410597532987595,
0.061623960733413696,
0.07694857567548752,
-0.13154466450214386,
0.013896279036998749,
0.08488219976425171,
0.050211917608976364,
-0.08303045481443405,
-0.03861803561449051,
-0.014303915202617645,
-0.02301439456641674,
-0.009539770893752575,
-0.03622512146830559,
0.04467915743589401,
-0.007735216524451971,
0.008007438853383064,
0.04505890980362892,
-0.024243297055363655,
0.08411174267530441,
0.0820077508687973,
0.029917770996689796,
0.08996117115020752,
-0.007938397116959095,
-0.050438057631254196,
0.06311395019292831,
-0.08415605872869492,
0.08984212577342987,
-0.08841758966445923,
-0.0216889176517725,
0.047476742416620255,
-0.010418985038995743,
0.026398666203022003,
-0.04593270644545555,
0.030980754643678665,
0.0366031676530838,
-0.028532255440950394,
-0.009334682486951351,
0.006290231831371784,
-0.016360003501176834,
-0.06323093920946121,
0.024501247331500053,
0.11278519779443741,
-0.13789598643779755,
-0.07439360022544861,
-0.08079218864440918,
0.022179551422595978,
-0.08065165579319,
-0.040646765381097794,
-0.03514867648482323,
0.04253251105546951,
-1.6485669095229323e-33,
0.02943408116698265,
0.03600897267460823,
-0.035257864743471146,
0.016335396096110344,
-0.04946274310350418,
0.0012791733024641871,
-0.007601826451718807,
0.07514233142137527,
-0.026222117245197296,
-0.09252133220434189,
-0.04676549509167671,
-0.02455574832856655,
-0.0027769124135375023,
0.0011515753576532006,
-0.011906922794878483,
-0.03312363103032112,
-0.031215591356158257,
0.021301718428730965,
-0.04312790185213089,
0.041032079607248306,
0.04654837027192116,
0.13999472558498383,
-0.06174002215266228,
-0.016475388780236244,
-0.083713598549366,
0.07028795778751373,
-0.07272464036941528,
0.09823433309793472,
-0.005760854575783014,
-0.014285441488027573,
-0.04401127249002457,
-0.004944389220327139,
-0.053045567125082016,
-0.07202223688364029,
-0.06646206229925156,
0.07245194166898727,
0.01304717268794775,
-0.03598732873797417,
-0.07710935175418854,
0.09401458501815796,
0.002200970659032464,
-0.08062881231307983,
-0.04000508785247803,
0.03325222060084343,
-0.06757085770368576,
-0.07454915344715118,
0.02218121849000454,
0.01160393189638853,
0.014689913019537926,
0.026263998821377754,
-0.012958870269358158,
0.03403633087873459,
-0.08934143930673599,
0.022618284448981285,
-0.027662895619869232,
-0.05086963623762131,
-0.0016961611108854413,
-0.01938200369477272,
0.052681297063827515,
0.007604075130075216,
-0.021158713847398758,
0.02237689681351185,
-0.0531444326043129,
-0.02694394439458847,
0.008735202252864838,
-0.04822218418121338,
-0.0839964970946312,
-0.010442492552101612,
0.023594828322529793,
0.02595304511487484,
0.043355219066143036,
-0.04031623527407646,
0.02786562778055668,
-0.032033126801252365,
0.04287009313702583,
-0.059527404606342316,
0.009468807838857174,
-0.027581052854657173,
-0.013602174818515778,
-0.009311316534876823,
-0.026232274249196053,
-0.03949708491563797,
0.03896820917725563,
0.08219309151172638,
0.024193070828914642,
0.1255638152360916,
-0.006747727748006582,
-0.019611235707998276,
0.12211019545793533,
-0.018766257911920547,
-0.06650698184967041,
0.05275178328156471,
-0.02574423886835575,
0.1003258228302002,
-0.049251165241003036,
-4.32750688617034e-8,
-0.05050153657793999,
0.0983862355351448,
-0.02591363899409771,
0.0026431838050484657,
0.006034931633621454,
-0.03781171143054962,
-0.03230667859315872,
0.07233541458845139,
-0.07057203352451324,
-0.014815477654337883,
0.017075462266802788,
-0.015771623700857162,
0.036542896181344986,
-0.0013484933879226446,
0.03774803504347801,
0.10452241450548172,
0.03554581478238106,
0.08171921223402023,
-0.01970021426677704,
0.026187866926193237,
-0.008207743987441063,
0.024495990946888924,
0.023975810036063194,
-0.05002719908952713,
0.03715723007917404,
-0.1090267151594162,
-0.022321179509162903,
0.0886901244521141,
0.061417192220687866,
-0.0012922270689159632,
0.0033325485419481993,
0.0653422474861145,
-0.00508917449042201,
-0.00463706161826849,
0.012766044586896896,
0.09840194135904312,
0.07415133714675903,
-0.010294771753251553,
-0.035378701984882355,
-0.023849274963140488,
-0.0031164605170488358,
0.030405733734369278,
-0.03770364075899124,
0.018777547404170036,
-0.09372156113386154,
0.03848501667380333,
0.011192601174116135,
-0.06386413425207138,
0.002987056039273739,
-0.04536955803632736,
-0.010486472398042679,
0.017099672928452492,
-0.027020031586289406,
0.10789738595485687,
0.041433561593294144,
-0.051161039620637894,
0.0612008199095726,
-0.03254326060414314,
0.04491613432765007,
0.05652163550257683,
-0.024692261591553688,
0.038082268089056015,
-0.021671153604984283,
-0.014613323844969273
] | https://github.com/dynamical-inference/patchsae | 11 | 0 | 0 | 0 |
PIED: Physics-Informed Experimental Design for Inverse Problems | https://openreview.net/forum?id=w7P92BEsb2 | [
"Apivich Hemachandra",
"Gregory Kang Ruey Lau",
"See-Kiong Ng",
"Bryan Kian Hsiang Low"
] | Poster | In many science and engineering settings, system dynamics are characterized by governing partial differential equations (PDEs), and a major challenge is to solve inverse problems (IPs) where unknown PDE parameters are inferred based on observational data gathered under limited budget.
Due to the high costs of setting up and running experiments, experimental design (ED) is often done with the help of PDE simulations to optimize for the most informative design parameters (e.g., sensor placements) to solve such IPs, prior to actual data collection. This process of optimizing design parameters is especially critical when the budget and other practical constraints make it infeasible to adjust the design parameters between trials during the experiments.
However, existing experimental design (ED) methods tend to require sequential and frequent design parameter adjustments between trials. Furthermore, they also have significant computational bottlenecks due to the need for complex numerical simulations for PDEs, and do not exploit the advantages provided by physics informed neural networks (PINNs) in solving IPs for PDE-governed systems, such as its meshless solutions, differentiability, and amortized training.
This work presents Physics-Informed Experimental Design (PIED), the first ED framework that makes use of PINNs in a fully differentiable architecture to perform continuous optimization of design parameters for IPs for one-shot deployments.
PIED overcomes existing methods' computational bottlenecks through parallelized computation and meta-learning of PINN parameter initialization, and proposes novel methods to effectively take into account PINN training dynamics in optimizing the ED parameters.
Through experiments based on noisy simulated data and even real world experimental data, we empirically show that given limited observation budget, PIED significantly outperforms existing ED methods in solving IPs, including for challenging settings where the inverse parameters are unknown functions rather than just finite-dimensional. | Physics-Informed Neural Network, PINNs, Experimental Design, AI For Science, Active Learning, Data Selection | An experimental design framework for PDE-based inverse problems that uses PINNs and its training dynamics, in a fully differentiable architecture to perform continuous optimization of design parameters. | 14,224 | 2503.07070 | [
-0.1049974113702774,
0.0024828386958688498,
0.08718296140432358,
0.10129265487194061,
0.024468140676617622,
-0.1117759719491005,
-0.0628473088145256,
0.021515216678380966,
0.0076170507818460464,
0.03274243324995041,
-0.030620897188782692,
-0.026231348514556885,
-0.05584986135363579,
-0.04285964369773865,
-0.07902264595031738,
-0.019012099131941795,
0.035516370087862015,
0.014601847156882286,
-0.032315824180841446,
-0.02627735398709774,
-0.00283761415630579,
0.00769691402092576,
0.015701763331890106,
-0.017578812316060066,
-0.05526917427778244,
-0.04717690497636795,
0.07237115502357483,
-0.01722235605120659,
-0.08618457615375519,
0.014306281693279743,
0.14147581160068512,
0.00028090583509765565,
-0.10964406281709671,
-0.03378317877650261,
-0.0672840029001236,
0.028477726504206657,
-0.05162360146641731,
-0.006922000087797642,
-0.03405429422855377,
-0.02519768290221691,
-0.04949497431516647,
0.05882991477847099,
0.058538638055324554,
0.031204167753458023,
0.13626782596111298,
-0.057789236307144165,
-0.009120148606598377,
-0.04634036868810654,
-0.001492342445999384,
0.07869549095630646,
-0.015310055576264858,
-0.021713269874453545,
0.037835493683815,
-0.05455935746431351,
0.06915754079818726,
0.02236926555633545,
0.06010255590081215,
-0.00561187881976366,
-0.01681472547352314,
0.01520059909671545,
0.06636170297861099,
-0.02419605851173401,
-0.08952809870243073,
-0.037180814892053604,
-0.023403989151120186,
0.011707646772265434,
0.05085119977593422,
0.06777298450469971,
0.05235650762915611,
0.010368594899773598,
0.005509970709681511,
0.07184327393770218,
-0.04975194111466408,
-0.05500148981809616,
0.03637463226914406,
0.03588905185461044,
0.031874172389507294,
0.036691486835479736,
0.039366960525512695,
-0.05962177738547325,
-0.0017012394964694977,
0.01517548318952322,
-0.031503256410360336,
-0.06541357934474945,
0.010958353988826275,
0.006128375884145498,
-0.04952534660696983,
0.1308153122663498,
0.005533188581466675,
-0.06327803432941437,
0.04411207512021065,
-0.0324036180973053,
-0.08559839427471161,
-0.04542532190680504,
0.0422818660736084,
0.009301678277552128,
0.06650738418102264,
-0.045545876026153564,
0.004046251997351646,
0.06272649019956589,
-0.0051465402357280254,
0.03231751173734665,
-0.020666692405939102,
0.09881426393985748,
-0.027738288044929504,
0.022361529991030693,
0.02111596055328846,
0.03735974431037903,
-0.009625610895454884,
0.011086871847510338,
0.055101215839385986,
0.0015915753319859505,
0.02816907875239849,
0.010009155608713627,
0.01602422073483467,
-0.045189350843429565,
-0.0010518342023715377,
-0.037705447524785995,
0.011964026838541031,
0.08968718349933624,
-0.033269766718149185,
-0.046611487865448,
-0.011906526982784271,
0.019507281482219696,
0.04058906063437462,
0.01259712502360344,
-0.0545152947306633,
4.645602898275621e-33,
-0.015278023667633533,
-0.03856920823454857,
0.007632386405020952,
-0.005640305578708649,
0.06389300525188446,
-0.03343139588832855,
0.08899489045143127,
0.0695909634232521,
-0.0008444187697023153,
0.011684499680995941,
-0.07127547264099121,
-0.023602325469255447,
-0.07021571695804596,
0.052690327167510986,
0.05376509577035904,
-0.00877203419804573,
0.013900870457291603,
0.061657749116420746,
-0.06683037430047989,
-0.08059506863355637,
0.047117799520492554,
-0.12417639046907425,
-0.0003059506416320801,
-0.06273046880960464,
0.07724810391664505,
-0.012973829172551632,
-0.030410895124077797,
0.07065556943416595,
-0.08604297041893005,
-0.003595065325498581,
-0.011962639167904854,
0.04851783812046051,
-0.003495936281979084,
0.07303973287343979,
-0.01736205816268921,
-0.0009118943708017468,
0.07938304543495178,
0.0027722585946321487,
-0.01731831394135952,
-0.004927963018417358,
0.07361792773008347,
0.04147091880440712,
0.07003500312566757,
0.04641934856772423,
-0.06872614473104477,
-0.04408728703856468,
-0.008043711073696613,
-0.0014583389274775982,
-0.0020123175345361233,
-0.014978860504925251,
-0.041548218578100204,
-0.024687806144356728,
0.028644653037190437,
-0.0809793695807457,
0.057114288210868835,
-0.08366864174604416,
0.05190906301140785,
-0.06173565983772278,
0.05821967124938965,
0.12136726081371307,
-0.064513199031353,
0.01882774941623211,
0.003729300806298852,
-0.013554719276726246,
-0.007365007884800434,
0.10416284948587418,
0.03553919121623039,
0.037522926926612854,
0.04683706909418106,
-0.06350592523813248,
-0.028602810576558113,
0.016848742961883545,
-0.005243396852165461,
-0.02621614560484886,
0.09067779034376144,
0.025559213012456894,
0.08353392034769058,
0.04781205952167511,
-0.003377768211066723,
-0.048788249492645264,
-0.017847826704382896,
0.09933383762836456,
-0.030765390023589134,
0.07185054570436478,
-0.05021823197603226,
-0.04993128776550293,
-0.020078333094716072,
0.007448901887983084,
-0.0906958132982254,
-0.12637391686439514,
-0.05286119505763054,
-0.02813030406832695,
0.00760941207408905,
-0.03250107541680336,
-0.004044760018587112,
-4.921404362363254e-33,
-0.1119958758354187,
0.006806883029639721,
-0.04673490300774574,
-0.005376680288463831,
0.019426634535193443,
-0.019986238330602646,
-0.02200237102806568,
-0.023267272859811783,
0.04405513033270836,
-0.09433465451002121,
-0.040938448160886765,
-0.07794301956892014,
-0.02346710115671158,
0.03988193720579147,
0.029587090015411377,
0.09698855131864548,
-0.12868934869766235,
-0.0979437604546547,
0.06516238301992416,
0.03928300738334656,
-0.04373864084482193,
0.08167116343975067,
-0.027542270720005035,
-0.05663120001554489,
-0.02945283055305481,
-0.019639616832137108,
-0.008848076686263084,
0.005297807510942221,
0.041524581611156464,
-0.019089698791503906,
-0.0980282798409462,
-0.009608235210180283,
-0.028906671330332756,
0.02194116823375225,
-0.011313037015497684,
0.12458047270774841,
0.09131314605474472,
0.013468614779412746,
0.003646255936473608,
-0.008036348968744278,
0.039113614708185196,
0.006737768184393644,
-0.09806996583938599,
0.024529442191123962,
0.03685285896062851,
-0.033131763339042664,
-0.017759602516889572,
-0.003977279178798199,
0.017119375988841057,
0.010545935481786728,
-0.02514221891760826,
0.006897063925862312,
0.042440034449100494,
-0.028902744874358177,
-0.03680757060647011,
-0.016548359766602516,
0.02393987402319908,
0.06638050824403763,
0.06793243438005447,
0.015178165398538113,
-0.021908875554800034,
-0.09278137236833572,
0.10968615114688873,
0.0137163195759058,
-0.05626612901687622,
0.061585698276758194,
-0.026336463168263435,
0.08650903403759003,
0.041508276015520096,
-0.06530693918466568,
-0.028385508805513382,
-0.017349103465676308,
0.06718750298023224,
-0.006342652253806591,
0.040903374552726746,
-0.08486972004175186,
0.04052698239684105,
-0.01880040392279625,
-0.024226300418376923,
0.04698009416460991,
-0.00754591403529048,
0.06371238082647324,
0.01078120805323124,
0.06786135584115982,
0.03548527508974075,
0.01087817084044218,
-0.02011501044034958,
0.05616768077015877,
0.026346415281295776,
-0.053627315908670425,
0.029941502958536148,
0.014494973234832287,
0.06493758410215378,
0.034623149782419205,
-0.020493298768997192,
-5.3314810344318175e-8,
-0.021436361595988274,
0.0064725978299975395,
0.13031181693077087,
-0.01936066895723343,
0.024392062798142433,
0.04940766468644142,
-0.0038446742109954357,
-0.03517092019319534,
-0.00335143250413239,
-0.07772377133369446,
0.02971150167286396,
0.0037095153238624334,
-0.005566268693655729,
-0.038539040833711624,
-0.05018109828233719,
0.059776779264211655,
0.022178148850798607,
0.0258368831127882,
-0.049939870834350586,
0.0046535274013876915,
0.01583205908536911,
0.10599202662706375,
0.01047446858137846,
-0.09615529328584671,
0.09866086393594742,
-0.08151256293058395,
-0.05594324320554733,
-0.0343024767935276,
0.051538094878196716,
0.02747105062007904,
-0.053239788860082626,
-0.05704633146524429,
0.05367046594619751,
0.07131071388721466,
0.06052316725254059,
0.00898047350347042,
-0.01352304220199585,
-0.02008865773677826,
-0.06504097580909729,
-0.03465038537979126,
-0.07948649674654007,
0.007491040509194136,
-0.05574340745806694,
0.01182868704199791,
0.0511649064719677,
-0.022994469851255417,
0.007886556908488274,
-0.025935260578989983,
0.07507821917533875,
0.06358738243579865,
-0.035255316644907,
0.007110367529094219,
-0.05316353216767311,
-0.016432998701930046,
0.07490093261003494,
0.054259199649095535,
-0.05142723396420479,
-0.019541321322321892,
0.04345070943236351,
0.02008483000099659,
0.03588271141052246,
0.051139719784259796,
-0.09185335785150528,
0.00897307600826025
] | https://github.com/apivich-h/pied | 2 | 0 | 0 | 0 |
AgentRefine: Enhancing Agent Generalization through Refinement Tuning | https://openreview.net/forum?id=FDimWzmcWn | [
"Dayuan Fu",
"Keqing He",
"Yejie Wang",
"Wentao Hong",
"Zhuoma GongQue",
"Weihao Zeng",
"Wei Wang",
"Jingang Wang",
"Xunliang Cai",
"Weiran Xu"
] | Poster | Large Language Model (LLM) based agents have proved their ability to perform complex tasks like humans. However, there is still a large gap between open-sourced LLMs and commercial models like the GPT series. In this paper, we focus on improving the agent generalization capabilities of LLMs via instruction tuning. We first observe that the existing agent training corpus exhibits satisfactory results on held-in evaluation sets but fails to generalize to held-out sets. These agent-tuning works face severe formatting errors and are frequently stuck in the same mistake for a long while. We analyze that the poor generalization ability comes from overfitting to several manual agent environments and a lack of adaptation to new situations. They struggle with the wrong action steps and can not learn from the experience but just memorize existing observation-action relations. Inspired by the insight, we propose a novel AgentRefine framework for agent-tuning. The core idea is to enable the model to learn to correct its mistakes via observation in the trajectory. Specifically, we propose an agent synthesis framework to encompass a diverse array of environments and tasks and prompt a strong LLM to refine its error action according to the environment feedback. AgentRefine significantly outperforms state-of-the-art agent-tuning work in terms of generalization ability on diverse agent tasks. It also has better robustness facing perturbation and can generate diversified thought in inference. Our findings establish the correlation between agent generalization and self-refinement and provide a new paradigm for future research. | agent, self-refine, diversity, generalization, data synthesis | The self-refine data can expand the search space of LLM agent and improve the reason quality, leading a generalized performance in agent tasks. | 14,212 | 2501.01702 | [
-0.009128888137638569,
-0.0765603631734848,
0.00013434316497296095,
0.027041614055633545,
0.023990599438548088,
-0.015748437494039536,
0.06016635149717331,
-0.03396623209118843,
-0.03197750821709633,
-0.0199661236256361,
-0.08421716839075089,
-0.01847175508737564,
0.04302573204040527,
0.028352146968245506,
0.036121685057878494,
-0.035604000091552734,
0.08921551704406738,
0.08479516953229904,
-0.06442902237176895,
-0.0363212488591671,
0.021603437140583992,
0.06812729686498642,
0.036096084862947464,
-0.02307494357228279,
-0.08358249068260193,
-0.01047450304031372,
-0.006326928269118071,
-0.07384636253118515,
0.020867548882961273,
-0.028622286394238472,
0.047100987285375595,
0.055759355425834656,
-0.04994634911417961,
0.021116074174642563,
-0.005421580281108618,
0.11530543863773346,
-0.12967908382415771,
0.029374100267887115,
-0.006315273232758045,
-0.027573326602578163,
-0.020335720852017403,
0.02055060677230358,
0.002071868861094117,
0.006394211668521166,
0.08572878688573837,
-0.018055938184261322,
-0.03656498342752457,
-0.043570175766944885,
-0.010892770253121853,
-0.010603160597383976,
-0.11589278280735016,
-0.04481310769915581,
-0.01942642778158188,
0.007299385964870453,
-0.0041067935526371,
0.02369743026793003,
0.03540350869297981,
0.03963325917720795,
0.0398387685418129,
-0.11513539403676987,
0.008989616297185421,
-0.09634557366371155,
-0.048893362283706665,
-0.04533637687563896,
-0.07095897197723389,
-0.017595579847693443,
0.002191402018070221,
0.04704465717077255,
0.0021038763225078583,
0.004614024888724089,
-0.06656937301158905,
0.025899238884449005,
0.036288317292928696,
0.0362551249563694,
-0.01774517260491848,
0.04800081625580788,
-0.043040789663791656,
-0.021593553945422173,
0.06629188358783722,
-0.01652255840599537,
-0.03421109542250633,
-0.042317863553762436,
0.02542247623205185,
0.055451586842536926,
0.10215313732624054,
-0.022671829909086227,
0.04987892135977745,
0.03491734340786934,
0.09140157699584961,
0.04780349135398865,
-0.010843761265277863,
-0.03984309360384941,
0.0656256452202797,
0.015570742078125477,
0.003501706523820758,
0.09475710988044739,
-0.019342780113220215,
-0.11950475722551346,
-0.006747644394636154,
0.10426288843154907,
0.0014584214659407735,
0.08428416401147842,
-0.04191720113158226,
0.01984674111008644,
-0.06470402330160141,
0.001656368374824524,
0.04107067734003067,
0.06444109976291656,
0.015140721574425697,
-0.07947525382041931,
-0.016915548592805862,
0.02159489318728447,
0.05068182945251465,
-0.00819189939647913,
0.05584783852100372,
-0.027267422527074814,
0.04166702181100845,
-0.0060661304742097855,
-0.05215727537870407,
0.05418168008327484,
0.0539793036878109,
0.019810372963547707,
0.02264888770878315,
0.017779238522052765,
0.09545055031776428,
0.010068022646009922,
0.014440358616411686,
-5.605886182145929e-34,
0.08154289424419403,
0.054111577570438385,
0.02244618535041809,
0.07057218998670578,
0.05120037868618965,
0.030362743884325027,
0.025308378040790558,
0.020003193989396095,
-0.01029657106846571,
0.008935382589697838,
0.05289978161454201,
0.02010836824774742,
-0.060295701026916504,
0.07643438130617142,
0.02504781447350979,
0.0242929570376873,
-0.04409679025411606,
0.08931884169578552,
-0.017108041793107986,
-0.041974253952503204,
0.011194792576134205,
0.01399702113121748,
0.04633777588605881,
-0.08527610450983047,
0.015167192555963993,
0.029436353594064713,
0.042663171887397766,
-0.043568551540374756,
0.01712946593761444,
-0.02577710896730423,
-0.03805491328239441,
0.00807139091193676,
0.025399304926395416,
0.12643635272979736,
-0.04845796525478363,
-0.055535607039928436,
-0.06750975549221039,
-0.06539998203516006,
0.04968377947807312,
0.016717493534088135,
-0.02360345982015133,
0.051348645240068436,
0.06498804688453674,
-0.028465088456869125,
-0.03694881126284599,
-0.0770660787820816,
-0.0012549360981211066,
0.07116640359163284,
-0.016200566664338112,
-0.008105535060167313,
0.03502531722187996,
0.044974107295274734,
0.01203353051096201,
-0.0022903578355908394,
-0.03269774094223976,
0.07963810861110687,
0.010038618929684162,
0.05154403671622276,
-0.03917425870895386,
0.0017476666253060102,
0.041621748358011246,
0.05179482698440552,
0.001649409532546997,
0.09045155346393585,
0.03581647574901581,
-0.007157345302402973,
-0.02307846024632454,
0.1459941864013672,
0.04823094606399536,
0.013179019093513489,
-0.04431132599711418,
0.009257754310965538,
-0.0025483418721705675,
0.025818541646003723,
0.0075323018245399,
-0.06562511622905731,
0.08971672505140305,
-0.11299590766429901,
-0.01714560016989708,
-0.028924578800797462,
0.005401467438787222,
0.027760110795497894,
-0.015492990612983704,
-0.021647784858942032,
-0.04177852347493172,
-0.12341934442520142,
-0.012952527962625027,
-0.05822357162833214,
0.004834542982280254,
0.015411639586091042,
-0.11916860938072205,
-0.05207686126232147,
-0.08981607854366302,
0.017911219969391823,
-0.018118685111403465,
8.089342306457378e-35,
-0.010706248693168163,
-0.07440941035747528,
-0.05641600117087364,
0.06123453006148338,
-0.040224600583314896,
-0.09055226296186447,
-0.04421033710241318,
-0.012479137629270554,
-0.005483780987560749,
-0.05650738254189491,
-0.07571812719106674,
-0.035724204033613205,
0.025138074532151222,
0.030698910355567932,
-0.022751139476895332,
-0.07410025596618652,
-0.019551580771803856,
-0.016444770619273186,
0.038845520466566086,
0.009407993406057358,
-0.006335459183901548,
0.0896574854850769,
-0.07222732156515121,
0.03571958839893341,
0.005168714560568333,
0.02324472740292549,
0.024505382403731346,
0.09767686575651169,
0.03978854417800903,
0.03705667331814766,
0.010883825831115246,
-0.002512105507776141,
-0.014716346748173237,
0.04629029706120491,
-0.028965217992663383,
0.07076422870159149,
0.06473715603351593,
0.011339842341840267,
-0.0349765419960022,
0.11637932062149048,
0.07832330465316772,
-0.08707112818956375,
-0.03707757592201233,
-0.00797513872385025,
0.010905412957072258,
0.004248886834830046,
-0.07269392907619476,
0.0377609021961689,
-0.0068236831575632095,
-0.04089558497071266,
-0.06725741922855377,
-0.04344450309872627,
-0.08621462434530258,
-0.03459770232439041,
-0.028579438105225563,
-0.07040304690599442,
0.04335302487015724,
-0.05506296455860138,
0.007670578081160784,
-0.05423899367451668,
-0.048532672226428986,
-0.007876324467360973,
0.045340199023485184,
-0.04197470843791962,
-0.009201910346746445,
-0.005829062312841415,
-0.053201403468847275,
-0.03132421150803566,
0.047975506633520126,
-0.009995505213737488,
0.05268126726150513,
-0.05885821580886841,
-0.08165394514799118,
-0.007877633906900883,
0.04044957458972931,
-0.02318803034722805,
-0.014708343893289566,
-0.046667519956827164,
0.014325417578220367,
-0.08275128155946732,
0.042566072195768356,
0.03782716765999794,
0.02728806994855404,
0.06260804086923599,
-0.00029692245880141854,
0.08924584835767746,
0.01816551201045513,
0.18135260045528412,
0.014265707693994045,
0.009403087198734283,
-0.08633973449468613,
0.03682735562324524,
0.020535562187433243,
0.021898509934544563,
-0.13673661649227142,
-4.098546924069524e-8,
-0.083480603992939,
0.0517510250210762,
0.04664215072989464,
0.07043386250734329,
-0.00916412565857172,
-0.08599220216274261,
-0.020631399005651474,
0.027922766283154488,
0.003771626390516758,
-0.040133655071258545,
0.059247083961963654,
-0.038929443806409836,
-0.0018385121366009116,
0.0035624534357339144,
0.034980807453393936,
0.04881485924124718,
-0.043097056448459625,
0.002182421972975135,
-0.0796303004026413,
-0.0005825533298775554,
0.030011525377631187,
0.026899123564362526,
-0.010535644367337227,
-0.03405635803937912,
0.056271910667419434,
-0.09489062428474426,
-0.02529717981815338,
0.07172717899084091,
0.031256720423698425,
0.06079256162047386,
0.049567755311727524,
-0.003165287198498845,
0.03809697553515434,
-0.023578647524118423,
0.008375612087547779,
0.03395642340183258,
0.03370044380426407,
-0.04161902144551277,
0.07632645219564438,
-0.08956115692853928,
-0.01834055967628956,
0.0582708902657032,
-0.0057276333682239056,
0.030463235452771187,
0.008780073374509811,
-0.035164330154657364,
-0.06451618671417236,
-0.13301248848438263,
0.061916906386613846,
-0.09207168966531754,
-0.013832448981702328,
-0.017687305808067322,
-0.012630811892449856,
0.005278223194181919,
0.06909442692995071,
0.05567814037203789,
-0.0018556492868810892,
-0.04951905086636543,
0.0733688548207283,
0.06316685676574707,
-0.08990387618541718,
0.04183993488550186,
-0.04304942861199379,
-0.0317472442984581
] | 0 | 0 | 0 | 0 |
|
TabM: Advancing tabular deep learning with parameter-efficient ensembling | https://openreview.net/forum?id=Sd4wYYOhmY | [
"Yury Gorishniy",
"Akim Kotelnikov",
"Artem Babenko"
] | Poster | Deep learning architectures for supervised learning on tabular data range from simple multilayer perceptrons (MLP) to sophisticated Transformers and retrieval-augmented methods.
This study highlights a major, yet so far overlooked opportunity for substantially improving tabular MLPs; namely, parameter-efficient ensembling -- a paradigm for imitating an ensemble of models with just one model.
We start by describing TabM -- a simple model based on MLP and BatchEnsemble (an existing technique), improved with our custom modifications.
Then, we perform a large scale evaluation of tabular DL architectures on public benchmarks in terms of both task performance and efficiency, which renders the landscape of tabular DL in a new light.
In particular, we find that TabM outperforms prior tabular DL models, while the complexity of attention- and retrieval-based methods does not pay off.
Lastly, we conduct a detailed empirical analysis, that sheds some light on the high performance of TabM.
For example, we show that parameter-efficient ensembling is not an arbitrary trick, but rather a highly effective way to reduce overfitting and improve optimization dynamics of tabular MLPs.
Overall, our work brings an impactful technique to tabular DL, analyses its behaviour, and advances the performance-efficiency tradeoff with TabM -- a simple and powerful baseline for researchers and practitioners. | tabular, tabular data, deep learning, architecture | Parameter-efficient ensembling has a massive positive impact on tabular MLPs, and TabM is a new SOTA architecture illustrating that. | 14,197 | 2410.24210 | [
-0.03404494374990463,
-0.1380775421857834,
0.0401514507830143,
0.0015381715493276715,
-0.010931179858744144,
0.014496713876724243,
-0.009267032146453857,
-0.05496812239289284,
-0.07624692469835281,
-0.028270544484257698,
-0.048621632158756256,
-0.044447172433137894,
-0.030566299334168434,
-0.008722812868654728,
-0.08987604081630707,
0.02993880584836006,
0.08278201520442963,
0.08774102479219437,
-0.12023061513900757,
0.03564411401748657,
-0.015602714382112026,
-0.0014568488113582134,
0.02648218721151352,
-0.01962067373096943,
0.005822019185870886,
0.01678551733493805,
-0.029285430908203125,
-0.0365726500749588,
-0.009897707961499691,
-0.038074202835559845,
0.01665187068283558,
-0.017311466857790947,
-0.07593013346195221,
0.06199830397963524,
-0.018053583800792694,
0.008840122260153294,
-0.07299821823835373,
0.0038325143977999687,
0.031594742089509964,
-0.05131804943084717,
-0.019477877765893936,
-0.048644643276929855,
-0.024343883618712425,
-0.004310063552111387,
0.055454906076192856,
-0.007190124597400427,
0.0302035603672266,
-0.06717497110366821,
0.011682646349072456,
-0.0073872338980436325,
-0.05723317712545395,
0.02486792765557766,
-0.04091371223330498,
0.0037313210777938366,
-0.03841685131192207,
0.0449041873216629,
0.0018905890174210072,
-0.022512171417474747,
-0.02650373801589012,
0.005900985095649958,
-0.053373198956251144,
-0.05873904377222061,
-0.013962311670184135,
-0.01437205821275711,
-0.011327031999826431,
-0.029188476502895355,
0.07424971461296082,
0.006537873763591051,
0.039133839309215546,
-0.006702188868075609,
0.06638048589229584,
0.10916608572006226,
-0.045309197157621384,
0.04506496712565422,
0.005295224487781525,
0.028277020901441574,
0.06014268845319748,
-0.03970090672373772,
0.03024897910654545,
-0.06549952179193497,
0.007665024138987064,
0.09363172203302383,
0.02498357743024826,
0.03900059685111046,
0.18519838154315948,
-0.046683840453624725,
0.008627400733530521,
0.04972885176539421,
0.018975283950567245,
-0.030061423778533936,
-0.03243853524327278,
-0.057347919791936874,
0.0010876166634261608,
-0.05737406760454178,
0.009152160957455635,
0.07280176132917404,
0.05120551958680153,
-0.04638712853193283,
-0.04584746062755585,
0.10391342639923096,
-0.04421938210725784,
0.0827573910355568,
0.08428604900836945,
-0.018893804401159286,
-0.04393359273672104,
-0.07337760925292969,
0.08804583549499512,
0.039001110941171646,
0.03215343877673149,
-0.13012568652629852,
0.10664501041173935,
0.010404218919575214,
0.035340867936611176,
-0.026494385674595833,
0.060158148407936096,
-0.008513132110238075,
-0.014880970120429993,
-0.06078365072607994,
0.04036759212613106,
0.06577098369598389,
-0.07447566092014313,
0.02440841868519783,
-0.04769807308912277,
0.08727249503135681,
-0.0277873408049345,
0.036159027367830276,
-0.07725358009338379,
3.205465463490881e-33,
0.010358410887420177,
0.019306879490613937,
0.04963262379169464,
0.0162274781614542,
0.08413178473711014,
-0.02753996104001999,
0.04957209527492523,
0.07081097364425659,
-0.027786264196038246,
0.05817548558115959,
-0.09264323115348816,
0.019063379615545273,
-0.03848985210061073,
0.10158625990152359,
0.032611601054668427,
0.023326588794589043,
-0.08380437642335892,
0.048236649483442307,
-0.07797932624816895,
-0.030173664912581444,
0.0199495330452919,
-0.04591716453433037,
0.010240967385470867,
-0.03377542272210121,
-0.01501380279660225,
-0.002778079826384783,
0.02730134129524231,
0.03891995921730995,
-0.007181152235716581,
0.028800273314118385,
-0.017728203907608986,
0.03543476387858391,
-0.0225861594080925,
0.02789466083049774,
0.06012687459588051,
-0.07950510829687119,
-0.010738514363765717,
0.010108081623911858,
0.0691886767745018,
0.018716005608439445,
0.0006462938617914915,
0.08291278779506683,
-0.02190781943500042,
-0.03644366189837456,
-0.08444693684577942,
-0.06975770741701126,
0.009951339103281498,
0.0075490777380764484,
-0.030184654518961906,
-0.07954806834459305,
0.11014048755168915,
-0.026634657755494118,
-0.05605059489607811,
-0.04768707603216171,
0.001500063925050199,
-0.01338195987045765,
0.02557399310171604,
0.06284471601247787,
0.05981910601258278,
0.11942701041698456,
-0.022376790642738342,
0.048650823533535004,
-0.03637837618589401,
0.08335956931114197,
0.007964689284563065,
0.04121676832437515,
0.024641720578074455,
0.11619766801595688,
0.05559943616390228,
-0.13141438364982605,
-0.007325808983296156,
-0.0002210152888437733,
0.020001795142889023,
-0.10878817737102509,
0.08567313849925995,
0.006850670091807842,
0.05841727554798126,
-0.09531424939632416,
0.00940797571092844,
-0.014086033217608929,
0.060512661933898926,
0.10557519644498825,
-0.05339862406253815,
-0.05382230877876282,
-0.04636058956384659,
-0.03344348445534706,
0.040732093155384064,
-0.09692959487438202,
0.019396400079131126,
0.013290606439113617,
-0.060758333653211594,
0.012437306344509125,
-0.05058886110782623,
-0.09682416170835495,
-0.04155155271291733,
-2.373303625185013e-33,
0.0029542187694460154,
-0.02812427654862404,
-0.06716985255479813,
0.002634631935507059,
-0.015025400556623936,
0.008727231062948704,
-0.003432435216382146,
-0.03932378068566322,
-0.037646833807229996,
-0.026140665635466576,
-0.0009776095394045115,
-0.041861627250909805,
0.06536152958869934,
-0.030000530183315277,
0.043076809495687485,
0.015319736674427986,
-0.092008575797081,
0.009916149079799652,
-0.04504784569144249,
0.0022153567988425493,
-0.023620158433914185,
0.0664369985461235,
-0.007969897240400314,
0.010462287813425064,
-0.08316309005022049,
0.005256920121610165,
-0.08934812247753143,
0.07024063169956207,
0.029981128871440887,
-0.05041421577334404,
-0.0204524714499712,
-0.007443636655807495,
-0.025470014661550522,
-0.010627859272062778,
-0.024267589673399925,
0.1010904535651207,
0.042924776673316956,
-0.06630507856607437,
-0.008456595242023468,
0.02872289903461933,
0.03680987283587456,
0.023647604510188103,
-0.0062345173209905624,
0.028666505590081215,
-0.005329489707946777,
-0.01194063015282154,
-0.02849935181438923,
-0.047084178775548935,
-0.02634168229997158,
0.023982573300600052,
-0.03606124222278595,
-0.03706962987780571,
-0.09454098343849182,
0.006901307497173548,
-0.04319101199507713,
-0.033084969967603683,
-0.04287492483854294,
-0.026270097121596336,
0.040838878601789474,
-0.009239652194082737,
-0.07980912178754807,
-0.013148677535355091,
0.02599363960325718,
-0.0030418462119996548,
0.021722638979554176,
-0.008351224474608898,
0.009947715327143669,
-0.0007626647129654884,
-0.007557384669780731,
0.0029341087210923433,
-0.0021053997334092855,
-0.0578429140150547,
-0.006149786524474621,
0.002464169403538108,
-0.013091236352920532,
0.026728589087724686,
0.017872270196676254,
-0.007994702085852623,
-0.04030224680900574,
-0.0788968950510025,
0.02696126699447632,
-0.059391070157289505,
-0.03081863559782505,
-0.0025317862164229155,
0.06853363662958145,
0.059781353920698166,
0.09407433867454529,
0.05547449737787247,
-0.025655046105384827,
0.022451842203736305,
0.00827325414866209,
0.012437662109732628,
0.00933857262134552,
0.09026764333248138,
0.059328459203243256,
-3.9657123807046446e-8,
-0.04960579425096512,
0.10132493823766708,
-0.006489010993391275,
0.050196509808301926,
0.05515740439295769,
-0.07679739594459534,
0.012119627557694912,
0.18493390083312988,
0.005069841165095568,
-0.032302018254995346,
0.15266188979148865,
-0.027204332873225212,
-0.037369780242443085,
-0.031252048909664154,
0.04236266389489174,
0.11019834131002426,
0.03800453245639801,
-0.015339513309299946,
-0.0075484830886125565,
0.005820083431899548,
0.03248492255806923,
0.07239258289337158,
0.07820813357830048,
0.016367139294743538,
-0.003866330487653613,
-0.05590216815471649,
-0.07952505350112915,
0.07623770087957382,
0.04953581467270851,
-0.0077833631075918674,
0.0315336212515831,
0.029056653380393982,
-0.010737837292253971,
-0.07461149245500565,
0.03532685711979866,
0.0834059864282608,
0.027889370918273926,
-0.016917847096920013,
-0.012780669145286083,
-0.005931586492806673,
0.0110058868303895,
0.029074056074023247,
-0.06798681616783142,
0.00040301447734236717,
0.0022154017351567745,
-0.0018175122095271945,
-0.046457018703222275,
-0.0765223577618599,
0.01136966235935688,
0.01697695441544056,
0.007697344291955233,
-0.017908548936247826,
0.04172546789050102,
0.0885675773024559,
-0.02464311756193638,
0.036254290491342545,
0.003882755758240819,
-0.03281070291996002,
0.025102825835347176,
0.08382702618837357,
0.019391905516386032,
0.005711950361728668,
-0.08423137664794922,
0.06639079749584198
] | https://github.com/yandex-research/tabm | 257 | 0 | 0 | 0 |
Multi-Label Test-Time Adaptation with Bound Entropy Minimization | https://openreview.net/forum?id=75PhjtbBdr | [
"Xiangyu Wu",
"Feng Yu",
"Yang Yang",
"Qing-Guo Chen",
"Jianfeng Lu"
] | Poster | Mainstream test-time adaptation (TTA) techniques endeavor to mitigate distribution shifts via entropy minimization for multi-class classification, inherently increasing the probability of the most confident class. However, when encountering multi-label instances, the primary challenge stems from the varying number of labels per image, and prioritizing only the highest probability class inevitably undermines the adaptation of other positive labels. To address this issue, we investigate TTA within multi-label scenario (ML--TTA), developing Bound Entropy Minimization (BEM) objective to simultaneously increase the confidence of multiple top predicted labels. Specifically, to determine the number of labels for each augmented view, we retrieve a paired caption with yielded textual labels for that view. These labels are allocated to both the view and caption, called weak label set and strong label set with the same size k. Following this, the proposed BEM considers the highest top-k predicted labels from view and caption as a single entity, respectively, learning both view and caption prompts concurrently. By binding top-k predicted labels, BEM overcomes the limitation of vanilla entropy minimization, which exclusively optimizes the most confident class. Across the MSCOCO, VOC, and NUSWIDE multi-label datasets, our ML--TTA framework equipped with BEM exhibits superior performance compared to the latest SOTA methods, across various model architectures, prompt initialization, and varying label scenarios. The code is available at https://github.com/Jinx630/ML-TTA. | Vision-Language Models, Zero-Shot Multi-Label Generalization, Test-Time Adaptation | A Multi-Label Test-Time Adaptation method with Bound Entropy Minimization objective. | 14,187 | 2502.03777 | [
-0.005727848038077354,
-0.036358706653118134,
-0.0302236657589674,
-0.014454134739935398,
0.09835796803236008,
0.04663262143731117,
0.04099293053150177,
-0.004317175596952438,
0.041440825909376144,
-0.06256671249866486,
-0.04335267096757889,
-0.09535063803195953,
0.041594862937927246,
0.0030724541284143925,
-0.004806621931493282,
-0.05803954228758812,
0.10468044131994247,
0.045182004570961,
-0.08522544801235199,
0.012357218191027641,
0.010604597628116608,
-0.030183719471096992,
0.08139228075742722,
0.022814171388745308,
0.011630011722445488,
-0.03560745343565941,
-0.05387343093752861,
0.002065452514216304,
0.024124881252646446,
-0.08135805279016495,
-0.029852524399757385,
0.029854267835617065,
0.004939477890729904,
0.0887521505355835,
-0.08451689034700394,
0.0012031573569402099,
-0.11307140439748764,
0.030749693512916565,
0.025377493351697922,
0.03012913465499878,
-0.03126111254096031,
0.040807485580444336,
0.04580783471465111,
0.002021363703534007,
-0.03636481985449791,
0.08491062372922897,
0.025407563894987106,
-0.11174682527780533,
-0.02585465833544731,
0.027585607022047043,
-0.017538363113999367,
0.0014313041465356946,
-0.04792807623744011,
0.01176406629383564,
-0.11044915020465851,
0.10005039721727371,
0.09004170447587967,
-0.036625444889068604,
0.008875319734215736,
0.07680194079875946,
-0.039703283458948135,
-0.028573540970683098,
0.0155648123472929,
0.055266764014959335,
0.0465831495821476,
-0.006783845368772745,
0.06448494642972946,
0.023070896044373512,
-0.07720422744750977,
0.04557526111602783,
-0.022957855835556984,
0.05480283126235008,
0.018663737922906876,
0.02277555502951145,
0.042391665279865265,
0.06066625565290451,
0.02371939644217491,
-0.04679480195045471,
0.025244953110814095,
-0.063095822930336,
0.0268561914563179,
0.08635276556015015,
0.030786219984292984,
-0.036662887781858444,
0.1635764241218567,
-0.03950817510485649,
-0.0005826964043080807,
-0.008471831679344177,
-0.03938726335763931,
0.020796768367290497,
-0.06998974084854126,
-0.04632663354277611,
0.04022618755698204,
0.04421660676598549,
-0.04676629975438118,
0.030784141272306442,
-0.045032698661088943,
-0.004953077994287014,
0.04346628487110138,
0.054598353803157806,
-0.042535144835710526,
0.013688747771084309,
0.00024994462728500366,
-0.007900764234364033,
0.03660570830106735,
-0.09443604946136475,
0.1097295731306076,
-0.05712050944566727,
0.07833545655012131,
-0.021041959524154663,
0.07321802526712418,
0.0286425594240427,
-0.023634545505046844,
-0.006813927553594112,
-0.001736489706672728,
0.05082935839891434,
0.0006080599268898368,
-0.03309020400047302,
0.010556562803685665,
0.02394726313650608,
-0.06785566359758377,
0.014541895128786564,
0.06108764559030533,
-0.07083085924386978,
0.014319219626486301,
0.030772002413868904,
0.018810098990797997,
1.617450839372697e-33,
0.05319036915898323,
-0.017911365255713463,
0.08175282180309296,
-0.03518814221024513,
0.05017939582467079,
-0.006194007117301226,
-0.03435984253883362,
0.007710842415690422,
-0.04580889269709587,
-0.011650539934635162,
0.06050271913409233,
0.015740102156996727,
-0.03955133259296417,
0.0374370738863945,
0.011383150704205036,
0.047239333391189575,
-0.017531385645270348,
0.07722421735525131,
-0.03152100369334221,
0.0059951962903141975,
-0.03478846326470375,
0.00005462169065140188,
0.028641995042562485,
-0.10867071896791458,
0.026679955422878265,
0.0022243885323405266,
-0.0032140049152076244,
-0.01335839368402958,
0.01638980209827423,
0.005095094442367554,
-0.12095583975315094,
-0.01056559756398201,
0.04861842840909958,
0.024466803297400475,
0.045299045741558075,
-0.10047324746847153,
-0.044954393059015274,
0.03582533076405525,
-0.04111539572477341,
0.006113226525485516,
-0.06375592947006226,
0.05298880115151405,
0.059368815273046494,
-0.09083794802427292,
-0.014156284742057323,
-0.012323183007538319,
-0.04935889318585396,
0.03944060206413269,
-0.11494465172290802,
-0.08432487398386002,
-0.00684594688937068,
-0.03187158703804016,
-0.0454929955303669,
-0.06310255080461502,
-0.058091554790735245,
-0.007557259406894445,
0.01667994260787964,
0.04657270014286041,
0.0026595075614750385,
0.016709361225366592,
-0.06616737693548203,
0.039030056446790695,
0.02022477425634861,
0.058834537863731384,
0.037576738744974136,
0.11371484398841858,
0.03900422900915146,
-0.006944424007087946,
0.11377694457769394,
0.0051847356371581554,
-0.014374482445418835,
-0.045116472989320755,
-0.049123648554086685,
-0.04812321439385414,
0.05764327570796013,
0.014413163997232914,
0.060380157083272934,
-0.06831613928079605,
0.028395485132932663,
0.10092373937368393,
0.022867796942591667,
-0.0030429456382989883,
0.0446927547454834,
-0.02320287376642227,
-0.04083654657006264,
-0.050561461597681046,
0.03689411282539368,
-0.0785982608795166,
-0.03829780966043472,
0.013468200340867043,
-0.05892815813422203,
0.018520250916481018,
-0.03311967849731445,
-0.02924909070134163,
-0.02264099009335041,
-1.7160494688097474e-33,
0.012270387262105942,
0.014348450116813183,
0.0014331226702779531,
0.023933706805109978,
-0.013603256084024906,
-0.03661586344242096,
0.024169355630874634,
0.06474651396274567,
-0.051776986569166183,
-0.04382733628153801,
-0.003245761850848794,
-0.09807790070772171,
0.04158010706305504,
-0.032670747488737106,
-0.04456682503223419,
-0.006175664719194174,
-0.013844652101397514,
0.04411468654870987,
-0.03373633697628975,
0.018928248435258865,
0.12441423535346985,
0.13687944412231445,
-0.060716211795806885,
0.06566058844327927,
-0.02985573001205921,
0.0028600040823221207,
-0.0030248973052948713,
0.08826620876789093,
-0.01616417057812214,
-0.0807347521185875,
-0.02232414484024048,
-0.05145065113902092,
0.024887999519705772,
0.0013690645573660731,
-0.02670246735215187,
0.06453606486320496,
0.0061769322492182255,
-0.052264951169490814,
-0.09804575145244598,
0.1575763076543808,
0.015847589820623398,
-0.05544958636164665,
-0.01921674981713295,
-0.06934605538845062,
-0.06646867841482162,
-0.01480199210345745,
-0.04402134567499161,
-0.04479619860649109,
0.0485076829791069,
0.07827098667621613,
-0.07255494594573975,
-0.023493297398090363,
-0.08582840859889984,
0.08854454010725021,
-0.03564610332250595,
-0.12082045525312424,
-0.0968584194779396,
-0.0213276669383049,
0.05952635034918785,
0.05188474804162979,
-0.022245343774557114,
0.024998655542731285,
0.031077886000275612,
-0.004314746707677841,
0.028061402961611748,
-0.004336600657552481,
-0.01980317197740078,
-0.04128135368227959,
-0.016906019300222397,
0.02162322588264942,
0.03856111690402031,
0.03567281365394592,
0.005462923552840948,
-0.00472727045416832,
-0.041945844888687134,
0.018011748790740967,
0.035047177225351334,
0.0008198862196877599,
-0.0006587235257029533,
-0.07741852849721909,
-0.009517221711575985,
-0.023346828296780586,
-0.02264876663684845,
0.12597796320915222,
0.0415056087076664,
0.0756247267127037,
0.10460828989744186,
0.050950661301612854,
0.085995614528656,
-0.06076549366116524,
-0.00695395702496171,
0.024105776101350784,
-0.01856270246207714,
0.06442289799451828,
-0.058650292456150055,
-4.373763218268323e-8,
-0.019359104335308075,
0.009576058946549892,
-0.008377818390727043,
0.0191226527094841,
0.03327580913901329,
0.01679203473031521,
-0.039660073816776276,
0.06885524839162827,
-0.0035036818590015173,
-0.03129690885543823,
0.09559913724660873,
0.04792093485593796,
-0.019312556833028793,
0.0009331853943876922,
-0.07527641952037811,
0.01936483010649681,
-0.04403551295399666,
0.03861183673143387,
0.016888583078980446,
-0.06524986773729324,
-0.037234194576740265,
-0.03143111616373062,
0.0963100716471672,
-0.07621466368436813,
-0.03081524930894375,
-0.05335305631160736,
0.01792898401618004,
0.08166725933551788,
0.07867980748414993,
-0.05406304448843002,
-0.02853691577911377,
0.0017542954301461577,
-0.0920184850692749,
0.017505871132016182,
0.011966829188168049,
0.04503057524561882,
-0.04998830705881119,
0.00941371824592352,
0.02423105388879776,
0.05854824557900429,
-0.004184532444924116,
0.013525417074561119,
-0.050752341747283936,
0.0507965162396431,
0.03694034367799759,
-0.0005808410351164639,
0.000397978990804404,
-0.03693956136703491,
0.04842531308531761,
-0.04437750205397606,
-0.014582185074687004,
-0.047505587339401245,
0.012629345059394836,
0.0966988056898117,
0.07754768431186676,
0.003622740739956498,
0.01992594450712204,
-0.019281381741166115,
0.013099882751703262,
0.08913697302341461,
0.034808091819286346,
-0.06344977766275406,
-0.08341342210769653,
-0.021270770579576492
] | 0 | 0 | 0 | 0 |
|
ToolGen: Unified Tool Retrieval and Calling via Generation | https://openreview.net/forum?id=XLMAMmowdY | [
"Renxi Wang",
"Xudong Han",
"Lei Ji",
"Shu Wang",
"Timothy Baldwin",
"Haonan Li"
] | Poster | As large language models (LLMs) advance, their inability to autonomously execute tasks by directly interacting with external tools remains a critical limitation. Traditional methods rely on inputting tool descriptions as context, which is constrained by context length and requires separate, often inefficient, retrieval mechanisms. We introduce ToolGen, a paradigm shift that integrates tool knowledge directly into the LLM’s parameters by representing each tool as a unique token. This enables the LLM to generate tool calls and arguments as part of its next token prediction capabilities, seamlessly blending tool invocation with language generation. Our framework allows the LLM to access and utilize a vast amount of tools with no additional retrieval step, significantly enhancing both performance and scalability. Experimental results with over 47,000 tools show that ToolGen not only achieves superior results in both tool retrieval and autonomous task completion but also sets the stage for a new era of AI agents that can adapt to tools across diverse domains. By fundamentally transforming tool retrieval into a generative process, ToolGen paves the way for more versatile, efficient, and autonomous AI systems. ToolGen enables end-to-end tool learning and opens opportunities for integration with other advanced techniques such as chain-of-thought and reinforcement learning, thereby expanding the practical capabilities of LLMs | Agent, Tool Learning, Virtual Token | Unified tool retrieval and calling by transforming tools into virtual tokens | 14,183 | 2410.03439 | [
-0.08345653861761093,
-0.07017704099416733,
-0.004973486997187138,
0.0829802080988884,
0.020388131961226463,
-0.020123785361647606,
-0.01175718940794468,
0.025240423157811165,
0.007274015806615353,
-0.049375709146261215,
-0.06631842255592346,
-0.11679244041442871,
0.046282120048999786,
0.011719451285898685,
0.05008268728852272,
0.044500645250082016,
0.09659525752067566,
0.009722022339701653,
-0.06260670721530914,
-0.05576546490192413,
0.024435773491859436,
0.09628227353096008,
0.031170150265097618,
-0.01681790128350258,
0.023822901770472527,
-0.0012927480274811387,
-0.0004699735145550221,
-0.07387793809175491,
0.03647386282682419,
0.005839054938405752,
0.05271532014012337,
0.08392100781202316,
-0.08123768121004105,
0.01627737283706665,
-0.015125864185392857,
0.06272363662719727,
-0.06104842573404312,
0.020745012909173965,
0.03867119550704956,
-0.12455618381500244,
-0.018420767039060593,
-0.010587535798549652,
-0.0012002931907773018,
-0.008057307451963425,
0.06986398249864578,
0.007448709104210138,
-0.052240945398807526,
-0.07415850460529327,
0.01819799654185772,
0.033008985221385956,
-0.06547922641038895,
-0.06686017662286758,
-0.012317627668380737,
0.014295630156993866,
-0.016378596425056458,
0.009131720289587975,
0.04516148194670677,
0.004509450867772102,
0.00393802160397172,
-0.036450743675231934,
-0.044484708458185196,
-0.08699876815080643,
-0.03611227869987488,
-0.022363683208823204,
-0.06978185474872589,
-0.06829289346933365,
0.0829351618885994,
0.035626120865345,
-0.04319924861192703,
0.012484461069107056,
-0.022959072142839432,
0.10677487403154373,
-0.012542552314698696,
0.07793284952640533,
0.023721424862742424,
0.09009874612092972,
0.010292816907167435,
-0.02005944773554802,
0.018146220594644547,
-0.11786985397338867,
0.01538327056914568,
0.021858230233192444,
0.04079936072230339,
0.07197221368551254,
0.060987841337919235,
0.05913708359003067,
0.06410502642393112,
0.02578410692512989,
0.09131382405757904,
0.04871997982263565,
-0.0726608857512474,
-0.08833292126655579,
0.05159236490726471,
-0.029469899833202362,
0.03915989398956299,
0.025257840752601624,
0.01932084932923317,
-0.10933004319667816,
-0.048655640333890915,
0.07825902104377747,
-0.012298050336539745,
0.07790543138980865,
-0.0016718617407605052,
-0.1215253621339798,
-0.02157084457576275,
0.056391410529613495,
0.008528368547558784,
-0.025177085772156715,
0.015591186471283436,
-0.039319947361946106,
0.001665654475800693,
-0.0021361513063311577,
0.051507607102394104,
-0.0027361882384866476,
0.06600876152515411,
-0.007687422912567854,
0.03094973787665367,
-0.05889314413070679,
0.05381888151168823,
0.03403807431459427,
-0.0009032738162204623,
0.08945831656455994,
-0.05533791333436966,
0.04905501380562782,
0.0016204704297706485,
0.01933147944509983,
0.011151985265314579,
2.7947191649948315e-35,
0.04953852295875549,
-0.0169014111161232,
0.003027598839253187,
0.09131516516208649,
0.040499404072761536,
-0.013262838125228882,
0.052623502910137177,
0.07816935330629349,
-0.049583468586206436,
-0.022469907999038696,
-0.07551510632038116,
0.01997784711420536,
-0.03356461971998215,
0.0499843992292881,
0.030083270743489265,
0.016542645171284676,
-0.04062184318900108,
0.08084855228662491,
-0.06656397134065628,
-0.021136650815606117,
-0.015316559933125973,
-0.014802509918808937,
-0.011637658812105656,
0.04129314050078392,
0.07110760360956192,
0.013937548734247684,
0.12346196919679642,
-0.05603828281164169,
0.014732745476067066,
-0.02625291980803013,
-0.118483766913414,
-0.04202595353126526,
0.06942591816186905,
0.06833458691835403,
-0.018539387732744217,
0.03474052995443344,
-0.038431309163570404,
-0.018124187365174294,
0.0022562940139323473,
0.0328543595969677,
0.016817893832921982,
0.01195408683270216,
0.013478808104991913,
-0.05137721076607704,
-0.10669034719467163,
-0.12370678037405014,
-0.04964803159236908,
0.05977446585893631,
0.022196777164936066,
-0.010311904363334179,
0.031971149146556854,
0.056514348834753036,
0.04542430490255356,
0.01395341195166111,
0.014223233796656132,
-0.01417918037623167,
-0.04726012796163559,
0.04221427068114281,
0.03914393484592438,
0.0025337173137813807,
-0.012321656569838524,
0.04766698554158211,
-0.02722897008061409,
0.1335860788822174,
0.06298469752073288,
0.04985610395669937,
0.06517015397548676,
0.14358288049697876,
0.07970397919416428,
-0.012729602865874767,
-0.005208965856581926,
-0.012721816077828407,
0.012453030794858932,
-0.05108976364135742,
0.05987156555056572,
-0.0030693572480231524,
0.023702410981059074,
-0.15276971459388733,
0.023345325142145157,
0.0009848990011960268,
0.011217904277145863,
-0.001670439844019711,
-0.027722816914319992,
-0.03917807713150978,
0.024012340232729912,
-0.06125776842236519,
0.03254140913486481,
-0.09783582389354706,
-0.028934797272086143,
-0.025830579921603203,
-0.0633629858493805,
-0.08695051074028015,
-0.08019089698791504,
0.05078206956386566,
-0.04804103821516037,
-1.3860565212877568e-33,
0.02996114455163479,
-0.055095188319683075,
-0.03546300530433655,
0.045440904796123505,
-0.04096775874495506,
-0.08866030722856522,
-0.08342558890581131,
-0.04460877925157547,
-0.0872780829668045,
-0.04031463339924812,
-0.021984193474054337,
-0.012079392559826374,
0.0507761612534523,
-0.05896584317088127,
0.062177762389183044,
-0.00939285010099411,
-0.05301961675286293,
-0.04534823074936867,
-0.004237501882016659,
0.03701029345393181,
-0.017144694924354553,
0.07910751551389694,
-0.03958698734641075,
0.011859812773764133,
-0.09397678077220917,
0.004141189623624086,
0.004185070749372244,
0.006138934753835201,
-0.02295367419719696,
0.003866782644763589,
-0.013400275260210037,
-0.0023941064719110727,
-0.010767188854515553,
-0.06144808605313301,
-0.0359065979719162,
0.044256724417209625,
0.11055249720811844,
0.025125518441200256,
-0.02247164025902748,
0.02489018812775612,
0.03909248113632202,
-0.02626209892332554,
-0.016860418021678925,
0.025827011093497276,
-0.04666098952293396,
0.04901731386780739,
-0.08132714778184891,
0.011297606863081455,
0.012578933499753475,
-0.0013342632446438074,
0.003169306553900242,
0.0006475010304711759,
-0.06851992756128311,
-0.07420317083597183,
-0.057864781469106674,
-0.07795941084623337,
-0.028303656727075577,
-0.08791858702898026,
-0.0462019219994545,
-0.0029567114543169737,
-0.00719439098611474,
-0.025779670104384422,
0.046604547649621964,
-0.03560618683695793,
0.020056575536727905,
0.012981862761080265,
-0.049942370504140854,
-0.056418683379888535,
-0.04967811703681946,
0.017268896102905273,
0.053584981709718704,
-0.016285132616758347,
0.017879264429211617,
0.0007381202885881066,
0.023784194141626358,
0.008904038928449154,
0.025340434163808823,
-0.10441122949123383,
0.014911118894815445,
-0.08022252470254898,
0.02461860328912735,
0.0019894479773938656,
0.06408458948135376,
0.04369555413722992,
-0.03254616633057594,
0.018030814826488495,
-0.04236706718802452,
0.14563493430614471,
0.03772415593266487,
0.055201075971126556,
0.009923628531396389,
0.0001856943272287026,
-0.02007056586444378,
0.07436909526586533,
-0.007796688470989466,
-3.9063223766788724e-8,
-0.014570184051990509,
0.11790888011455536,
-0.03683524578809738,
0.08142757415771484,
-0.03857755288481712,
0.024952957406640053,
-0.05124661326408386,
0.09931018948554993,
0.014506658539175987,
-0.005332132335752249,
0.04592333361506462,
-0.07117006927728653,
-0.01668471284210682,
-0.022032558917999268,
-0.02421889267861843,
0.009612369351089,
-0.024895895272493362,
0.014162158593535423,
-0.0008870834135450423,
-0.06782820075750351,
0.0445236936211586,
0.019701583310961723,
0.03265826404094696,
-0.011710422113537788,
0.04960010573267937,
-0.06312163919210434,
-0.10563798993825912,
0.1121765524148941,
0.052959688007831573,
0.0025044118519872427,
0.003727591596543789,
-0.008659310638904572,
-0.0337643101811409,
-0.03153415396809578,
0.020674163475632668,
0.036360543221235275,
-0.0028186270501464605,
-0.05201399326324463,
0.010848596692085266,
-0.04594537243247032,
0.05000781640410423,
0.06321895122528076,
-0.04984059929847717,
-0.006593620870262384,
-0.023250505328178406,
-0.03078591637313366,
0.020975427702069283,
-0.08656279742717743,
0.03642047941684723,
-0.03670330345630646,
-0.07165253162384033,
-0.03461766242980957,
0.0468086339533329,
0.048134468495845795,
0.040398817509412766,
0.04798198491334915,
-0.028889916837215424,
-0.017629504203796387,
0.07563149183988571,
0.06457442045211792,
-0.014311852864921093,
-0.036249175667762756,
0.00456229830160737,
0.06660260260105133
] | https://github.com/Reason-Wang/ToolGen | 138 | 0 | 0 | 0 |
Activation Gradient based Poisoned Sample Detection Against Backdoor Attacks | https://openreview.net/forum?id=VNMJfBBUd5 | [
"Danni Yuan",
"Mingda Zhang",
"Shaokui Wei",
"Li Liu",
"Baoyuan Wu"
] | Poster | This work studies the task of poisoned sample detection for defending against data poisoning based backdoor attacks. Its core challenge is finding a generalizable and discriminative metric to distinguish between clean and various types of poisoned samples (e.g., various triggers, various poisoning ratios). Inspired by a common phenomenon in backdoor attacks that the backdoored model tend to map significantly different poisoned and clean samples within the target class to similar activation areas, we introduce a novel perspective of the circular distribution of the gradients w.r.t. sample activation, dubbed gradient circular distribution (GCD). And, we find two interesting observations based on GCD. One is that the GCD of samples in the target class is much more dispersed than that in the clean class. The other is that in the GCD of target class, poisoned and clean samples are clearly separated. Inspired by above two observations, we develop an innovative three-stage poisoned sample detection approach, called Activation Gradient based Poisoned sample Detection (AGPD). First, we calculate GCDs of all classes from the model trained on the untrustworthy dataset. Then, we identify the target class(es) based on the difference on GCD dispersion between target and clean classes. Last, we filter out poisoned samples within the identified target class(es) based on the clear separation between poisoned and clean samples. Extensive experiments under various settings of backdoor attacks demonstrate the superior detection performance of the proposed method to existing poisoned detection approaches according to sample activation-based metrics. | Backdoor Defense, Poisoned Sample Detection, AI security | null | 14,155 | 2312.06230 | [
-0.055565860122442245,
-0.09267767518758774,
-0.028956862166523933,
-0.07295555621385574,
0.1877855807542801,
-0.027963414788246155,
0.04208485409617424,
-0.02238943986594677,
0.07863176614046097,
-0.004443817306309938,
0.04277878999710083,
-0.024124249815940857,
0.10131670534610748,
-0.03745593875646591,
-0.0000708831794327125,
-0.04674272611737251,
0.05440840125083923,
0.07552976161241531,
-0.11077280342578888,
-0.004701580386608839,
-0.05549798533320427,
0.04260119050741196,
0.02139032818377018,
0.004122248850762844,
-0.08052055537700653,
0.02367459237575531,
0.06287948042154312,
-0.08498551696538925,
-0.04286128655076027,
-0.06357597559690475,
0.05342348292469978,
0.008613062091171741,
0.025650307536125183,
0.00577550707384944,
-0.04487871751189232,
-0.026740001514554024,
-0.0332937017083168,
0.07089614868164062,
0.007199060637503862,
0.023793108761310577,
0.03267928957939148,
0.05352505296468735,
0.02864544466137886,
0.05768413841724396,
-0.02911710925400257,
-0.017016790807247162,
-0.03026248887181282,
-0.004124684724956751,
0.0010353083489462733,
-0.013149083591997623,
-0.014174697920680046,
0.036435890942811966,
-0.09297283738851547,
0.08864033222198486,
-0.017879394814372063,
-0.13219985365867615,
0.021153859794139862,
-0.019932756200432777,
-0.03025476634502411,
0.04253631457686424,
-0.023670803755521774,
-0.045416783541440964,
-0.017378997057676315,
-0.013559142127633095,
0.05290345475077629,
-0.03758866339921951,
-0.017781484872102737,
-0.016499480232596397,
0.09020628035068512,
0.0027581520844250917,
0.005216156132519245,
0.04886002466082573,
-0.10539048165082932,
-0.04152108356356621,
0.010968469083309174,
0.14034943282604218,
-0.019868237897753716,
0.011340843513607979,
0.008426140062510967,
-0.09480193257331848,
0.04369422048330307,
0.02117997035384178,
0.12236219644546509,
0.014283297583460808,
0.10047465562820435,
0.03160107508301735,
-0.04639501869678497,
-0.023410065099596977,
0.022094471380114555,
0.0741143599152565,
0.027975477278232574,
0.06338462978601456,
0.022450190037488937,
0.00904835294932127,
-0.00694043654948473,
0.05497948080301285,
0.006328853312879801,
-0.0871688723564148,
0.05215160921216011,
0.10853669792413712,
-0.037601862102746964,
-0.04952995479106903,
-0.06990867108106613,
-0.02364582195878029,
0.05204808712005615,
-0.019679393619298935,
0.06315542757511139,
0.01565421186387539,
-0.014520040713250637,
-0.05732114985585213,
-0.026588812470436096,
-0.0002869554446078837,
0.0014346997486427426,
-0.010443250648677349,
0.022777589038014412,
-0.05721946805715561,
0.06120138615369797,
0.017770832404494286,
-0.02872471511363983,
0.003305138787254691,
0.035534705966711044,
-0.039898667484521866,
-0.03526596352458,
-0.01840049959719181,
-0.0011632750974968076,
0.015935244038701057,
-0.046983469277620316,
3.2390324393625806e-33,
0.02761102467775345,
0.0739106833934784,
-0.027454162016510963,
-0.024900583550333977,
-0.03460657596588135,
-0.04278058931231499,
-0.02931351587176323,
0.0418965108692646,
0.051549773663282394,
0.11104988306760788,
-0.06027043238282204,
-0.020386850461363792,
-0.049047380685806274,
0.08493199944496155,
0.0690091997385025,
0.10215293616056442,
-0.011742528527975082,
0.036410246044397354,
-0.009629953652620316,
-0.07376101613044739,
0.056880883872509,
-0.08232654631137848,
-0.01526403333991766,
-0.014597895555198193,
-0.024053849279880524,
0.04627756029367447,
-0.01083712000399828,
0.026712145656347275,
-0.05552588775753975,
0.028775183483958244,
-0.039741452783346176,
0.05062619969248772,
0.00552584370598197,
0.03921100124716759,
0.007624405436217785,
0.04466712102293968,
-0.0335678867995739,
0.006227417383342981,
-0.04700443893671036,
-0.015428183600306511,
-0.032344989478588104,
0.008418981917202473,
0.005098591558635235,
-0.020777126774191856,
0.016581682488322258,
-0.03290611878037453,
-0.08530640602111816,
-0.01296981517225504,
0.03116900473833084,
-0.07740452140569687,
0.020623983815312386,
-0.011173995211720467,
0.10095405578613281,
-0.011073512956500053,
-0.10858427733182907,
0.1130436584353447,
0.0016339797293767333,
0.01949121057987213,
0.012562660500407219,
0.10627129673957825,
-0.004986078944057226,
0.08639981597661972,
-0.06924418359994888,
0.03447887673974037,
-0.015626147389411926,
-0.009612786583602428,
-0.0652083158493042,
-0.06966395676136017,
0.06395358592271805,
0.0602981336414814,
0.038488663733005524,
0.033528540283441544,
-0.07836879789829254,
0.00619511678814888,
-0.06040487438440323,
-0.038173411041498184,
0.06487880647182465,
0.0340806320309639,
0.02328386716544628,
-0.03152493014931679,
0.017160657793283463,
-0.08693341910839081,
0.029854988679289818,
-0.007823583669960499,
-0.10744816064834595,
0.03123757615685463,
-0.035795293748378754,
-0.09946241974830627,
-0.07197140157222748,
0.044731855392456055,
-0.013438181951642036,
-0.032245833426713943,
-0.019911417737603188,
0.03899591043591499,
-0.03507847338914871,
-3.161364588867875e-33,
-0.12567093968391418,
0.05664139613509178,
-0.02832026779651642,
0.026140326634049416,
-0.011491524055600166,
0.0015178078319877386,
-0.034504689276218414,
0.029483843594789505,
0.009374034591019154,
0.05912895128130913,
-0.04650084301829338,
0.006472875829786062,
0.030863316729664803,
0.053059399127960205,
0.03627274930477142,
-0.002263600006699562,
0.011183297261595726,
0.06148463487625122,
-0.061168063431978226,
0.020989416167140007,
0.027740778401494026,
0.06920556724071503,
-0.08796777576208115,
-0.0038198037073016167,
0.0005378855857998133,
0.0216669961810112,
-0.011041496880352497,
0.09558485448360443,
0.024008749052882195,
0.03718606010079384,
-0.10523273795843124,
0.06821836531162262,
0.00482370937243104,
0.051574308425188065,
-0.03878167271614075,
0.005675361957401037,
0.050029221922159195,
-0.06714290380477905,
-0.07226637750864029,
0.033070582896471024,
0.027431001886725426,
0.048420995473861694,
-0.09517647325992584,
-0.013289181515574455,
0.021853255107998848,
-0.04778043180704117,
0.06811270862817764,
0.08017624914646149,
0.04259369522333145,
-0.04422100633382797,
0.007743224035948515,
0.045816417783498764,
-0.000009395225788466632,
0.12099733203649521,
-0.028867168352007866,
-0.005388714373111725,
0.031256482005119324,
-0.04498834162950516,
0.10213438421487808,
0.044854793697595596,
-0.07352927327156067,
0.012923793867230415,
-0.09634598344564438,
0.024459967389702797,
-0.010221957229077816,
-0.0583011656999588,
0.02854638174176216,
0.005913089960813522,
0.009792034514248371,
-0.0076872315257787704,
0.06419345736503601,
0.03162553533911705,
0.04041547328233719,
-0.03494446352124214,
-0.0637550875544548,
-0.07058801501989365,
-0.04036741703748703,
-0.06678498536348343,
-0.044334765523672104,
0.01993742771446705,
0.056479644030332565,
-0.06319675594568253,
0.04341273382306099,
0.062265604734420776,
0.0333181731402874,
-0.0036481525748968124,
0.0551365502178669,
0.002856668783351779,
0.018049169331789017,
-0.004914878401905298,
-0.06275039166212082,
-0.016123531386256218,
-0.030138693749904633,
0.014649360440671444,
0.005217433907091618,
-4.415894849785218e-8,
0.022331614047288895,
0.01191034447401762,
0.007885826751589775,
0.009741702117025852,
0.05709293484687805,
0.06341586261987686,
-0.060380708426237106,
0.046262212097644806,
-0.039916276931762695,
-0.11634428799152374,
0.05285162851214409,
0.030588189139962196,
-0.07849721610546112,
-0.07922673225402832,
-0.003307127160951495,
0.08437006175518036,
-0.04680900648236275,
0.039773691445589066,
-0.04853944480419159,
0.0172903910279274,
-0.02229566127061844,
0.06077153608202934,
0.043384943157434464,
-0.01132096815854311,
0.03498713672161102,
-0.044891104102134705,
0.02023782953619957,
0.039818063378334045,
0.0758662298321724,
0.01314114686101675,
-0.07471394538879395,
-0.027906468138098717,
0.0020757883321493864,
0.038367487490177155,
-0.0017319673206657171,
0.15486744046211243,
-0.012575126253068447,
-0.02816794440150261,
-0.033620547503232956,
-0.01032323855906725,
-0.08926587551832199,
0.019567430019378662,
-0.06859221309423447,
-0.010033223778009415,
-0.005231822375208139,
-0.03154248744249344,
-0.01901926100254059,
-0.038375124335289,
0.018599942326545715,
-0.039699312299489975,
-0.018981771543622017,
-0.05844574794173241,
-0.02260659448802471,
0.04202450439333916,
0.0149098951369524,
-0.009044909849762917,
-0.014091459102928638,
-0.08142144232988358,
-0.014700817875564098,
0.0018619814654812217,
0.08862515538930893,
0.02497822232544422,
0.008839738555252552,
-0.09103719890117645
] | 0 | 0 | 0 | 0 |
|
Causally Motivated Sycophancy Mitigation for Large Language Models | https://openreview.net/forum?id=yRKelogz5i | [
"Haoxi Li",
"Xueyang Tang",
"Jie ZHANG",
"Song Guo",
"Sikai Bai",
"Peiran Dong",
"Yue Yu"
] | Poster | Incorporating user preferences into large language models (LLMs) can enhance the personalization and reliability of model outputs and facilitate the application of LLMs to real-world scenarios. However, leveraging user preferences can be a double-edged sword. Recent studies have found that improper utilization can incur sycophancy, where LLMs prioritize alignment with user preferences over the correctness of their outputs. To address sycophancy in LLMs, we analyze and model the problem through the lens of structured causal models (SCMs). We attribute sycophancy to LLMs' reliance on spurious correlations between user preferences and model outputs in this paper. Based on the proposed SCMs, we develop a novel framework, termed **CAUSM**, to mitigate sycophancy in LLMs by exploiting a significant causal signature. Specifically, we eliminate the spurious correlations embedded in the intermediate layers of LLMs through causally motivated head reweighting, and then calibrate the intra-head knowledge along the causal representation direction. Extensive experiments are conducted across diverse language tasks to demonstrate the superiority of our method over state-of-the-art competitors in mitigating sycophancy in LLMs. | Large Language Model; Sycophancy; Causal Modeling | null | 14,154 | null | [
-0.00568648986518383,
-0.07895482331514359,
0.0490499846637249,
0.014822307974100113,
0.04109494760632515,
0.033414293080568314,
0.03503834456205368,
-0.006746972445398569,
0.03585335239768028,
-0.002822014968842268,
-0.07384181767702103,
-0.04203106462955475,
0.1206645593047142,
0.05757758021354675,
0.014899175614118576,
-0.04604613780975342,
0.13516227900981903,
-0.009496855549514294,
-0.0681639015674591,
-0.0355873666703701,
-0.013911852613091469,
-0.012626832351088524,
-0.04489211365580559,
0.042093947529792786,
-0.016661731526255608,
-0.04473448172211647,
-0.027255278080701828,
-0.01052043866366148,
-0.03770366683602333,
0.0042259651236236095,
0.029125822708010674,
0.03100425750017166,
-0.055024269968271255,
-0.09062208235263824,
-0.031233487650752068,
-0.03922150656580925,
-0.054687533527612686,
-0.0356106199324131,
-0.01481783390045166,
-0.1251540631055832,
-0.06684885174036026,
-0.04870612546801567,
0.017126793041825294,
0.06310184299945831,
-0.024776482954621315,
0.024164488539099693,
0.027090365067124367,
0.0263528972864151,
-0.03785470500588417,
0.01788444072008133,
-0.07823940366506577,
-0.021703634411096573,
0.030092773959040642,
-0.0046143317595124245,
0.019173095002770424,
0.04607333987951279,
0.03453356772661209,
0.082221120595932,
-0.07502566277980804,
-0.03148990124464035,
-0.14011000096797943,
-0.09702350199222565,
-0.11511828750371933,
0.041730042546987534,
0.0227322019636631,
0.024458395317196846,
-0.03134480491280556,
0.055330052971839905,
-0.07471592724323273,
0.15128488838672638,
0.02894820086658001,
0.04704289510846138,
-0.003545454703271389,
0.035121068358421326,
0.030452599748969078,
0.03686670586466789,
-0.0009990300750359893,
-0.06447580456733704,
0.021505599841475487,
-0.06655742228031158,
-0.04662604629993439,
0.06821843236684799,
-0.017351362854242325,
0.005670917686074972,
0.1406228393316269,
0.008013077080249786,
0.04463212564587593,
0.010046743787825108,
0.04084516316652298,
0.06980764120817184,
-0.0005769723211415112,
-0.06789068132638931,
0.14180335402488708,
0.021388577297329903,
0.04225103184580803,
0.09882903099060059,
0.037562936544418335,
0.009780053980648518,
0.03540152311325073,
0.09142666310071945,
0.029723847284913063,
0.06415791809558868,
0.006544152274727821,
-0.027301885187625885,
0.0292068924754858,
-0.034615371376276016,
-0.061985503882169724,
0.08486035466194153,
-0.006709796376526356,
-0.07506249845027924,
-0.03138987720012665,
0.026960985735058784,
0.026488330215215683,
-0.02724117413163185,
-0.02456243894994259,
-0.033935513347387314,
0.05315670743584633,
-0.020637700334191322,
0.01955566555261612,
0.01959446258842945,
-0.0026313643902540207,
-0.01571029983460903,
-0.0021501618903130293,
0.002754984190687537,
0.0288854893296957,
-0.013501957058906555,
-0.056579139083623886,
2.1786204548537335e-33,
0.02375425584614277,
0.02544129453599453,
0.009333394467830658,
0.02612161450088024,
0.05224715918302536,
0.016260147094726562,
-0.02494093030691147,
0.03957658261060715,
-0.05700402706861496,
0.0016350361984223127,
0.03221069648861885,
0.02175995707511902,
-0.021840747445821762,
0.08456084877252579,
0.0287744402885437,
0.041970472782850266,
-0.050355322659015656,
0.08390290290117264,
-0.07430032640695572,
-0.04075642302632332,
0.053216319531202316,
-0.04306585341691971,
-0.008352220058441162,
-0.013372828252613544,
0.004477787762880325,
-0.0012267186539247632,
0.01470121555030346,
0.013811519369482994,
-0.02037198282778263,
0.0329907052218914,
-0.005236009135842323,
-0.06634315103292465,
0.07370628416538239,
-0.012012822553515434,
-0.03275173529982567,
-0.06310690194368362,
-0.05373341590166092,
-0.0820779800415039,
0.02910449728369713,
0.016713736578822136,
0.016238590702414513,
0.01938714273273945,
-0.002390095265582204,
0.028596116229891777,
-0.10872288793325424,
-0.038997989147901535,
-0.016330325976014137,
0.06119455769658089,
-0.02800845168530941,
-0.06290290504693985,
0.07881651073694229,
0.062230031937360764,
0.046773266047239304,
0.00711739482358098,
-0.04897216707468033,
-0.016127726063132286,
-0.0522925890982151,
-0.03600697219371796,
-0.03306661918759346,
-0.04015437141060829,
0.03651070594787598,
0.05117180570960045,
0.02844519168138504,
0.025712810456752777,
0.06016005948185921,
-0.004896251484751701,
-0.04483460634946823,
0.008873128332197666,
0.04296740144491196,
0.019018759950995445,
-0.04640704393386841,
0.03098398819565773,
-0.07408808916807175,
-0.05031059682369232,
0.00480952113866806,
-0.0024092854000627995,
-0.004325921181589365,
-0.055972110480070114,
0.05216562747955322,
-0.043139196932315826,
0.057765379548072815,
-0.05122017115354538,
0.04038398340344429,
-0.03257567435503006,
0.0022489724215120077,
-0.03741937875747681,
-0.01759938709437847,
-0.08023912459611893,
-0.06787338852882385,
-0.03923166170716286,
0.0606614388525486,
-0.040985528379678726,
0.0005849121953360736,
0.019206218421459198,
-0.06235431879758835,
-3.427298879906309e-33,
-0.055635716766119,
-0.08771676570177078,
0.08139784634113312,
0.0038827911484986544,
-0.05784616246819496,
-0.023216674104332924,
0.014817625284194946,
-0.021509602665901184,
-0.03351406753063202,
-0.09345521777868271,
-0.06280674040317535,
-0.045784205198287964,
0.04229825362563133,
0.03726482391357422,
0.034849390387535095,
0.012022354640066624,
0.033607471734285355,
-0.002947122324258089,
0.015019183047115803,
0.0033569782972335815,
-0.0051887063309550285,
0.07520797103643417,
-0.06212375685572624,
0.003126564435660839,
-0.026510821655392647,
-0.04159407317638397,
-0.034779489040374756,
0.03516409173607826,
-0.0040319086983799934,
0.028296062722802162,
-0.05451400950551033,
-0.00341530772857368,
-0.0889829769730568,
-0.08411034941673279,
-0.06190576031804085,
0.07444408535957336,
-0.005242141894996166,
0.022640246897935867,
-0.07478735595941544,
0.11458054184913635,
0.01918218471109867,
-0.019926095381379128,
-0.0648883655667305,
-0.03203573077917099,
-0.02265065349638462,
-0.010801920667290688,
-0.11941178888082504,
0.06828105449676514,
0.02178005315363407,
-0.009690470062196255,
-0.03109598346054554,
0.015382125042378902,
0.07418904453516006,
0.06346473842859268,
-0.0851893499493599,
-0.06336682289838791,
-0.007174695376306772,
-0.10031722486019135,
0.08800745755434036,
-0.009668386541306973,
-0.08153706043958664,
-0.018244493752717972,
0.018358586356043816,
-0.07151729613542557,
0.022602779790759087,
-0.06909822672605515,
0.03233395144343376,
-0.021625185385346413,
0.018947718665003777,
-0.066825270652771,
0.024795297533273697,
-0.04545878246426582,
-0.09786426275968552,
-0.0636604055762291,
-0.00256227795034647,
0.04625909402966499,
0.014024052768945694,
-0.0385805107653141,
-0.03333412855863571,
-0.024650782346725464,
0.05260122939944267,
0.06483945995569229,
0.06580056250095367,
-0.014850401319563389,
0.06496168673038483,
0.06849843263626099,
0.012097371742129326,
0.12766022980213165,
0.01694146730005741,
0.028537772595882416,
-0.04131976142525673,
0.03850618377327919,
-0.017744820564985275,
0.1278822124004364,
-0.06319993734359741,
-4.6667160802371654e-8,
-0.04448288679122925,
-0.022831615060567856,
-0.017274385318160057,
0.13262052834033966,
0.019985921680927277,
-0.020663566887378693,
-0.02565932460129261,
0.06430664658546448,
-0.002471831627190113,
0.0009454800747334957,
0.002010979223996401,
0.03637990728020668,
-0.04544917866587639,
-0.013539685867726803,
0.010320478118956089,
0.0811675563454628,
-0.00893997773528099,
0.07319188117980957,
-0.0765145942568779,
-0.02528023160994053,
0.07842808961868286,
0.039289314299821854,
-0.00895371288061142,
0.005305521190166473,
0.049322158098220825,
-0.00598190538585186,
-0.019513394683599472,
0.1360844075679779,
0.07736780494451523,
0.01193768996745348,
-0.01404868345707655,
-0.016919933259487152,
-0.08103720843791962,
-0.07850566506385803,
0.03708228841423988,
0.08444173634052277,
-0.02297607809305191,
-0.030000384896993637,
0.0013390866806730628,
-0.04873337224125862,
0.10987254232168198,
0.050854671746492386,
0.0036515353713184595,
0.08244138211011887,
0.05741563439369202,
-0.012911754660308361,
-0.04113424941897392,
-0.05852250009775162,
0.02668979950249195,
-0.029780352488160133,
-0.025998800992965698,
-0.0251315850764513,
-0.011070038191974163,
0.055519524961709976,
0.04651527851819992,
-0.009139802306890488,
0.02513296529650688,
0.026195809245109558,
0.00841442309319973,
0.04758332669734955,
0.11705174297094345,
0.08959753811359406,
0.007852242328226566,
-0.026558412238955498
] | 0 | 0 | 0 | 0 |
|
Compositional simulation-based inference for time series | https://openreview.net/forum?id=uClUUJk05H | [
"Manuel Gloeckler",
"Shoji Toyota",
"Kenji Fukumizu",
"Jakob H. Macke"
] | Poster | Amortized simulation-based inference (SBI) methods train neural networks on simulated data to perform Bayesian inference. While this strategy avoids the need for tractable likelihoods, it often requires a large number of simulations and has been challenging to scale to time series data. Scientific simulators frequently emulate real-world dynamics through thousands of single-state transitions over time. We propose an SBI approach that can exploit such Markovian simulators by locally identifying parameters consistent with individual state transitions. We then compose these local results to obtain a posterior over parameters that align with the entire time series observation. We focus on applying this approach to neural posterior score estimation but also show how it can be applied, e.g., to neural likelihood (ratio) estimation. We demonstrate that our approach is more simulation-efficient than directly estimating the global posterior on several synthetic benchmark tasks and simulators used in ecology and epidemiology. Finally, we validate scalability and simulation efficiency of our approach by applying it to a high-dimensional Kolmogorov flow simulator with around one million data dimensions. | Simulation-based inference, Bayesian inference, time series, markovian simulators, Amortized Bayesian inference | Simulation-based inference for Markovian simulators leveraging the factorization | 14,141 | 2411.02728 | [
-0.11366958916187286,
-0.10669245570898056,
0.07515063136816025,
0.05205332860350609,
0.050310004502534866,
0.013943432830274105,
-0.08745196461677551,
-0.05549377575516701,
0.06458761543035507,
-0.06172223389148712,
-0.05229398235678673,
-0.08020621538162231,
-0.043937280774116516,
0.03882984444499016,
-0.09911394119262695,
-0.025944968685507774,
-0.014320907182991505,
0.07628672569990158,
0.0011376215843483806,
0.010920766741037369,
0.005559548735618591,
0.09016349166631699,
-0.00798892043530941,
0.004061273764818907,
0.0257511418312788,
0.02484055981040001,
-0.030249956995248795,
-0.026278723031282425,
-0.08550593256950378,
0.10023369640111923,
0.0801142007112503,
0.010682899504899979,
-0.023190636187791824,
-0.0603938065469265,
-0.008813531138002872,
-0.02451886050403118,
-0.023604972288012505,
-0.06771145761013031,
-0.024374939501285553,
0.044439200311899185,
0.07706229388713837,
0.010289997793734074,
-0.002562678186222911,
-0.002371222712099552,
0.10463669151067734,
-0.018423886969685555,
0.05611355975270271,
-0.047984614968299866,
0.012164382264018059,
0.08143198490142822,
-0.07460284233093262,
0.028730306774377823,
-0.024717917665839195,
0.02795683778822422,
0.008126508444547653,
-0.008389164693653584,
0.044495027512311935,
-0.023013411089777946,
0.002256914973258972,
-0.01708870567381382,
-0.05046819522976875,
-0.03882310912013054,
0.01744106598198414,
0.01769781857728958,
0.006570315454155207,
0.004373353440314531,
-0.00017517291416879743,
0.13994427025318146,
0.008974647149443626,
0.035721760243177414,
0.0006979339523240924,
0.0923110619187355,
-0.049087852239608765,
-0.0014732429990544915,
-0.07367724925279617,
-0.0023304966744035482,
0.015327868051826954,
-0.060519713908433914,
0.09277860075235367,
-0.07541380077600479,
-0.008878304623067379,
0.002713386435061693,
0.026792844757437706,
-0.06587788462638855,
0.042770788073539734,
0.0280813779681921,
0.003937426954507828,
0.12858463823795319,
0.018678387627005577,
0.025564875453710556,
0.017868053168058395,
0.014467825181782246,
0.009854535572230816,
-0.035527218133211136,
0.02958306297659874,
0.054126985371112823,
-0.031037144362926483,
-0.1055523231625557,
0.044755417853593826,
0.056256722658872604,
-0.08103867620229721,
0.07901382446289062,
0.10201143473386765,
-0.04183259978890419,
0.022045182064175606,
0.01843942515552044,
0.014437551610171795,
0.017138423398137093,
0.060516808182001114,
-0.025640640407800674,
0.018082715570926666,
0.05528397858142853,
0.07330062240362167,
0.03308255597949028,
0.05742316320538521,
-0.04671792313456535,
0.053040891885757446,
-0.07712562382221222,
-0.0064813848584890366,
0.1181071475148201,
-0.007075528148561716,
-0.038061097264289856,
-0.039163727313280106,
0.040986258536577225,
0.1010672003030777,
0.030390825122594833,
-0.048625919967889786,
2.4236089059284528e-33,
0.0796767845749855,
-0.05278947576880455,
0.09826469421386719,
0.07067891955375671,
0.03236202523112297,
-0.0038735445123165846,
0.01028688345104456,
-0.02677774615585804,
-0.012077346444129944,
0.009772037155926228,
-0.07167395204305649,
0.01098528504371643,
-0.09137754142284393,
0.03187904134392738,
0.035081665962934494,
0.03821072354912758,
-0.041870418936014175,
0.09088315069675446,
0.033754583448171616,
-0.0027318464126437902,
0.012327934615314007,
-0.062232065945863724,
-0.024738388136029243,
-0.04298451915383339,
-0.013145801611244678,
0.02834034152328968,
0.08352508395910263,
0.023767761886119843,
-0.053393807262182236,
0.005950687453150749,
-0.031510092318058014,
-0.02017321065068245,
-0.12143798917531967,
-0.0010719785932451487,
0.010415088385343552,
-0.0049736835062503815,
0.006432348862290382,
-0.0063686128705739975,
-0.04503307864069939,
0.005021105986088514,
-0.016689948737621307,
-0.020216969773173332,
0.03089098446071148,
0.013624841347336769,
-0.11941298842430115,
-0.11992087215185165,
-0.0008384730899706483,
-0.027715977281332016,
0.07961747795343399,
-0.03351708501577377,
0.029183238744735718,
-0.06999748945236206,
0.07341913133859634,
-0.12859579920768738,
-0.06854431331157684,
0.009854216128587723,
0.02650863490998745,
0.011182681657373905,
0.0020193150267004967,
0.050388745963573456,
-0.03516298532485962,
0.012562279589474201,
-0.052137088030576706,
-0.0029509486630558968,
-0.017863640561699867,
0.06732261180877686,
-0.05581127107143402,
0.040336690843105316,
0.03472381457686424,
-0.015333743765950203,
0.04752359166741371,
0.04453771561384201,
-0.025875046849250793,
-0.05119253322482109,
0.047535110265016556,
0.04830748960375786,
0.08570428192615509,
-0.053719114512205124,
-0.04055250808596611,
0.08037028461694717,
-0.0244695246219635,
-0.05076855793595314,
0.022966885939240456,
0.03565021976828575,
0.05186574161052704,
-0.027711687609553337,
0.06069619953632355,
-0.0004471480206120759,
-0.04494546726346016,
-0.06880168616771698,
-0.06824780255556107,
-0.061179403215646744,
0.06521637737751007,
-0.02738478034734726,
-0.012434974312782288,
-2.9288867037637665e-33,
-0.09228608757257462,
0.023001184687018394,
0.037944238632917404,
0.05167924612760544,
-0.0312909260392189,
-0.09813744574785233,
-0.07318880409002304,
0.08477908372879028,
0.015288098715245724,
-0.10948734730482101,
-0.05390084162354469,
-0.015372421592473984,
0.08278760313987732,
0.002418103627860546,
0.080234594643116,
-0.049454256892204285,
-0.031127316877245903,
-0.027237437665462494,
0.021438583731651306,
-0.034396134316921234,
-0.006066819187253714,
-0.03501792624592781,
-0.036593951284885406,
-0.028502237051725388,
-0.09649847447872162,
0.01981952041387558,
0.04184301942586899,
0.09060745686292648,
-0.0001910781575134024,
-0.039049580693244934,
-0.060966990888118744,
-0.06102865934371948,
-0.016094934195280075,
0.031889595091342926,
-0.05337037518620491,
0.06556964665651321,
0.08350173383951187,
-0.04986386001110077,
-0.01808835379779339,
-0.03298865258693695,
0.031322333961725235,
-0.012163869105279446,
-0.08787190914154053,
-0.0592552050948143,
-0.026164567098021507,
-0.0006597000756300986,
0.03781385347247124,
0.04300689697265625,
0.017335737124085426,
0.0049683027900755405,
0.010717049241065979,
0.08720395714044571,
-0.0721333771944046,
0.015266935341060162,
-0.050131335854530334,
-0.08740837126970291,
-0.05582311749458313,
-0.004025353584438562,
0.01320982538163662,
-0.05435182526707649,
-0.05617574229836464,
-0.030090104788541794,
0.022595353424549103,
0.0069894446060061455,
-0.045954301953315735,
-0.013421325013041496,
0.003894499037414789,
0.005288643296808004,
0.002173434942960739,
0.012878231704235077,
0.02056674100458622,
0.022675510495901108,
0.035507235676050186,
0.05091467127203941,
0.02719831094145775,
-0.04021066427230835,
-0.004093327559530735,
-0.03853986784815788,
0.009490915574133396,
-0.022333243861794472,
0.04843578115105629,
-0.05997157096862793,
0.0157926045358181,
0.021181926131248474,
0.05019199848175049,
0.019503358751535416,
0.040946558117866516,
-0.016931572929024696,
0.04654976353049278,
0.007389952428638935,
-0.08341462910175323,
-0.0017770365811884403,
0.006755176931619644,
0.01118785422295332,
-0.1216205358505249,
-4.5084835420539093e-8,
-0.007046385668218136,
0.08707922697067261,
0.12646502256393433,
0.06195884570479393,
0.04433678835630417,
0.020010938867926598,
0.009821843355894089,
0.015259776264429092,
0.07794853299856186,
-0.05290071666240692,
0.13966718316078186,
-0.0202619731426239,
0.042708270251750946,
-0.10723639279603958,
0.015099829062819481,
0.03434011712670326,
0.0362904891371727,
-0.018453139811754227,
-0.0039794789627194405,
0.03807605430483818,
-0.003673209110274911,
0.08976363390684128,
0.05902455374598503,
-0.03936881572008133,
0.07251350581645966,
-0.0643744170665741,
-0.0517488457262516,
0.07714398205280304,
0.05984862521290779,
0.013680269941687584,
0.0075304098427295685,
-0.01111332792788744,
-0.03435022756457329,
0.049701791256666183,
-0.009893921203911304,
-0.0251507218927145,
-0.007565069943666458,
0.04437143728137016,
-0.0050759087316691875,
-0.06230205297470093,
-0.046491239219903946,
-0.0066797006875276566,
-0.09727310389280319,
0.010270370170474052,
0.056060370057821274,
-0.034696921706199646,
-0.03744412958621979,
-0.03221278265118599,
0.08392488211393356,
-0.015603912062942982,
0.048305656760931015,
-0.03389410674571991,
-0.04249873757362366,
0.011392165906727314,
0.007252133917063475,
0.005937726702541113,
0.030661337077617645,
-0.057798828929662704,
0.0037909471429884434,
0.05872160941362381,
0.03352946415543556,
-0.0013400833122432232,
-0.08419962972402573,
-0.013739857822656631
] | https://github.com/mackelab/markovsbi | 5 | 0 | 0 | 0 |
Bayesian Treatment of the Spectrum of the Empirical Kernel in (Sub)Linear-Width Neural Networks | https://openreview.net/forum?id=O6znYvxC1U | [
"Ouns El Harzli",
"Bernardo Cuenca Grau"
] | Poster | We study Bayesian neural networks (BNNs) in the theoretical limits of infinitely increasing number of training examples, network width and input space dimension. Our findings establish new bridges between kernel-theoretic approaches and techniques derived from statistical mechanics through the correspondence between Mercer's eigenvalues and limiting spectral distributions of covariance matrices studied in random matrix theory.
Our theoretical contributions first consist in novel integral formulas that accurately describe the predictors of BNNs in the asymptotic linear-width and sublinear-width regimes. Moreover, we extend the recently developed renormalisation theory of deep linear neural networks, enabling a rigorous explanation of the mounting empirical evidence that hints at the theory's applicability to nonlinear BNNs with ReLU activations in the linear-width regime.
From a practical standpoint, our results introduce a novel technique for estimating the predictor statistics of a trained BNN that is applicable to the sublinear-width regime where the predictions of the renormalisation theory are inaccurate. | infinite bayesian neural networks, kernel theory, random matrix theory | null | 14,113 | null | [
-0.052384037524461746,
-0.15979386866092682,
0.04929076135158539,
0.026406267657876015,
0.019958358258008957,
0.048993077129125595,
-0.013076044619083405,
-0.01693800278007984,
-0.006711350288242102,
-0.06646770983934402,
-0.008071565069258213,
0.06803420931100845,
-0.015302266925573349,
0.008815565146505833,
-0.0521959587931633,
0.007188228890299797,
0.04131585359573364,
0.1067468523979187,
-0.16275879740715027,
0.023546960204839706,
0.030585192143917084,
0.06714814901351929,
-0.006856249179691076,
-0.004048903007060289,
0.0492095984518528,
-0.041017141193151474,
0.06476882100105286,
-0.04476454108953476,
-0.03824020177125931,
-0.009613718837499619,
0.0483073815703392,
-0.02307775616645813,
-0.059209007769823074,
0.003378089051693678,
0.04611155763268471,
0.01838386245071888,
0.04612729698419571,
0.02308054082095623,
0.030689295381307602,
0.007060266099870205,
0.03068407252430916,
-0.006962204817682505,
0.03493799269199371,
0.12479083240032196,
0.12442632019519806,
-0.013200187124311924,
0.018443036824464798,
-0.13788777589797974,
0.03197789192199707,
0.007624437101185322,
-0.029927588999271393,
0.007032612804323435,
-0.016342561691999435,
0.0809267908334732,
-0.009089079685509205,
-0.03661150112748146,
0.010540595278143883,
0.05742775648832321,
0.04742257297039032,
0.06633088737726212,
0.034659937024116516,
-0.11261168122291565,
-0.002780183684080839,
-0.04058973863720894,
0.005056851543486118,
0.04619862139225006,
-0.002563412068411708,
0.011195994913578033,
-0.03365617245435715,
0.015950407832860947,
-0.039485007524490356,
0.08588895946741104,
-0.14135035872459412,
0.008212760090827942,
0.0029853740707039833,
0.014214945025742054,
0.0003732989134732634,
0.03808070719242096,
-0.00039214815478771925,
-0.05357721820473671,
0.05759704113006592,
0.08234288543462753,
0.058141935616731644,
-0.04614770784974098,
0.08836550265550613,
0.013248933479189873,
-0.03520059213042259,
0.047350671142339706,
0.01473984681069851,
-0.0058242580853402615,
-0.05097150430083275,
-0.0381956547498703,
-0.07410746067762375,
-0.0816405862569809,
-0.013714448548853397,
-0.04695845767855644,
-0.03980034962296486,
-0.06629378348588943,
0.03150797635316849,
0.07127314805984497,
0.009220472536981106,
-0.01589374803006649,
0.03760461136698723,
0.054242826998233795,
0.015732310712337494,
0.038437072187662125,
0.05536933243274689,
0.05761052295565605,
0.10320226848125458,
-0.10452356189489365,
0.038544051349163055,
-0.003404371440410614,
0.014444154687225819,
0.04973771795630455,
0.018350346013903618,
-0.005586341489106417,
0.03884763643145561,
0.03323591500520706,
0.0041756583377718925,
0.046876534819602966,
-0.05912313237786293,
0.052900563925504684,
-0.04448116570711136,
-0.05520104989409447,
0.011026239953935146,
0.0767485722899437,
-0.10321147739887238,
5.120210946472027e-33,
-0.002961181104183197,
0.0400715209543705,
0.057512328028678894,
0.0966118797659874,
0.054555244743824005,
-0.0019904717337340117,
-0.03408277779817581,
-0.03569936752319336,
0.028824089094996452,
-0.005851589608937502,
-0.0012630770215764642,
0.05098153278231621,
-0.05179409310221672,
0.03911230340600014,
0.08243495970964432,
0.0846976488828659,
-0.043176524341106415,
0.01877591758966446,
0.006565462797880173,
-0.08621063083410263,
0.05561651289463043,
-0.034880224615335464,
0.006355578079819679,
0.010419998317956924,
-0.07788630574941635,
-0.09755531698465347,
0.030052727088332176,
0.03319733589887619,
-0.1128900870680809,
0.009903058409690857,
-0.04647795483469963,
0.06250344216823578,
0.0026406077668070793,
0.03758559003472328,
0.03662995994091034,
-0.06936511397361755,
0.03450174257159233,
0.005143114365637302,
0.02242276258766651,
-0.09330974519252777,
-0.05223192274570465,
-0.0046854340471327305,
-0.006394692696630955,
-0.017713109031319618,
-0.061264727264642715,
-0.09862777590751648,
0.07140367478132248,
-0.018395721912384033,
-0.009485277347266674,
-0.1309126913547516,
0.02474154718220234,
0.033158786594867706,
-0.020098213106393814,
-0.008560535497963428,
0.0657503679394722,
0.014002416282892227,
0.02710576541721821,
0.02465428225696087,
0.02005413919687271,
0.04971584677696228,
-0.01141262799501419,
-0.0018593177665024996,
0.01832883059978485,
0.02085834927856922,
-0.007256624288856983,
0.023235119879245758,
-0.014189839363098145,
0.03241686522960663,
0.0754380077123642,
-0.05129034444689751,
0.03965162858366966,
0.04311814159154892,
-0.026071390137076378,
-0.051991257816553116,
-0.015470514073967934,
-0.012914823368191719,
-0.04397046938538551,
-0.04183674976229668,
-0.03610830754041672,
0.08826474845409393,
-0.060265716165304184,
0.012148610316216946,
0.004722015932202339,
-0.0006798849790357053,
-0.07486757636070251,
-0.02834390476346016,
-0.003453463315963745,
-0.08803559839725494,
-0.012426676228642464,
-0.0016418683808296919,
-0.02931543067097664,
-0.052913080900907516,
-0.01877143234014511,
-0.03022206574678421,
-0.060329120606184006,
-4.033647886077231e-33,
-0.052376434206962585,
0.04675406217575073,
-0.04185958579182625,
0.0071197981014847755,
-0.03769683837890625,
-0.06346213817596436,
-0.009921759366989136,
0.0774240642786026,
0.0077039990574121475,
-0.10830948501825333,
0.05984548106789589,
-0.013224576599895954,
-0.01831239089369774,
-0.00568389380350709,
0.08267176896333694,
-0.004018482286483049,
-0.03792119026184082,
0.04781383275985718,
0.019148172810673714,
-0.10841476917266846,
0.039796970784664154,
0.007426622789353132,
0.01807524636387825,
0.02295810729265213,
0.024220304563641548,
-0.018483882769942284,
-0.05516919121146202,
0.10924550145864487,
-0.00705666933208704,
0.012100999243557453,
-0.05869700014591217,
-0.030372783541679382,
-0.08393386751413345,
0.05606747791171074,
-0.02366483397781849,
0.017189383506774902,
0.04565619304776192,
0.03311549872159958,
0.028985852375626564,
0.04781424626708031,
0.01933099329471588,
-0.051561783999204636,
-0.04745188355445862,
-0.07437974959611893,
-0.014617017470300198,
-0.0689052864909172,
0.05245508253574371,
-0.009452777914702892,
0.0013883734354749322,
-0.035730622708797455,
0.07941708713769913,
0.07565927505493164,
-0.08402837812900543,
0.06565142422914505,
-0.0954931303858757,
-0.002135113114491105,
-0.013746325857937336,
-0.004872037097811699,
0.10677235573530197,
0.05642132833600044,
-0.07864443957805634,
-0.03435485437512398,
-0.020135855302214622,
-0.0041581992991268635,
-0.03354312852025032,
-0.0044377828016877174,
-0.03552832081913948,
0.03240477666258812,
0.1072278618812561,
0.05921390280127525,
0.026967501267790794,
-0.018574843183159828,
0.0427829772233963,
0.1376461535692215,
-0.04119746387004852,
-0.021228361874818802,
0.051472555845975876,
-0.12306242436170578,
-0.06277690827846527,
-0.04639948904514313,
0.020481523126363754,
0.000881412997841835,
-0.005980551242828369,
-0.0033445085864514112,
0.08638369292020798,
0.06866303831338882,
0.028846727684140205,
-0.011906751431524754,
0.025266140699386597,
-0.08429205417633057,
0.028352342545986176,
-0.016518982127308846,
-0.016781657934188843,
0.04073431342840195,
-0.02228279411792755,
-4.440096645907943e-8,
-0.02626485750079155,
0.06014018878340721,
0.0016918204491958022,
-0.015567196533083916,
0.13163527846336365,
0.0019553385209292173,
0.08548100292682648,
0.08648551255464554,
-0.02796141430735588,
0.0822034403681755,
0.029402242973446846,
-0.018637700006365776,
0.002461918629705906,
-0.06417196989059448,
-0.06704574823379517,
0.013902070932090282,
-0.03766480088233948,
0.03783218562602997,
0.05150194093585014,
-0.036831147968769073,
0.04115472733974457,
-0.0021737867500633,
0.04689868539571762,
-0.02430555783212185,
0.009219726547598839,
-0.11618281155824661,
-0.06601216644048691,
0.06338275969028473,
0.04696415364742279,
0.02959623746573925,
0.01509738340973854,
0.007383836433291435,
-0.014028471894562244,
0.026913823559880257,
0.03094751201570034,
0.014136960729956627,
-0.026397328823804855,
-0.007763275410979986,
-0.039938412606716156,
0.0034916922450065613,
-0.05436204746365547,
0.03549830988049507,
0.035172395408153534,
0.016929524019360542,
0.016604028642177582,
0.01907893270254135,
0.019750528037548065,
0.028821436688303947,
0.032436467707157135,
0.030352652072906494,
0.07118852436542511,
-0.011709620244801044,
0.0176067091524601,
0.026954680681228638,
0.0036783902905881405,
0.021554868668317795,
0.027263574302196503,
-0.04404401406645775,
-0.08529540151357651,
0.05256614834070206,
-0.020198054611682892,
0.05692287161946297,
-0.11308517307043076,
-0.07334460318088531
] | 0 | 0 | 0 | 0 |
|
When GNNs meet symmetry in ILPs: an orbit-based feature augmentation approach | https://openreview.net/forum?id=wVTJRnZ11Z | [
"Qian Chen",
"Lei Li",
"Qian Li",
"Jianghua Wu",
"Akang Wang",
"Ruoyu Sun",
"Xiaodong Luo",
"Tsung-Hui Chang",
"Qingjiang Shi"
] | Poster | A common characteristic in integer linear programs (ILPs) is symmetry, allowing variables to be permuted without altering the underlying problem structure. Recently, GNNs have emerged as a promising approach for solving ILPs.
However, a significant challenge arises when applying GNNs to ILPs with symmetry: classic GNN architectures struggle to differentiate between symmetric variables, which limits their predictive accuracy. In this work, we investigate the properties of permutation equivalence and invariance in GNNs, particularly in relation to the inherent symmetry of ILP formulations. We reveal that the interaction between these two factors contributes to the difficulty of distinguishing between symmetric variables.
To address this challenge, we explore the potential of feature augmentation and propose several guiding principles for constructing augmented features. Building on these principles, we develop an orbit-based augmentation scheme that first groups symmetric variables and then samples augmented features for each group from a discrete uniform distribution. Empirical results demonstrate that our proposed approach significantly enhances both training efficiency and predictive performance. | integer linear programming, symmetry, machine learning, graph neural networks | null | 14,111 | 2501.14211 | [
-0.0614161379635334,
-0.05787381902337074,
0.07555487751960754,
-0.011961361393332481,
-0.001591941574588418,
-0.00045762001536786556,
0.026770194992423058,
-0.09740152955055237,
-0.0883696898818016,
-0.021004240959882736,
-0.04555531218647957,
0.010496675036847591,
-0.01135406456887722,
0.040840841829776764,
0.0183744877576828,
-0.010234849527478218,
0.12197094410657883,
0.10314559936523438,
-0.05119401961565018,
-0.02911844477057457,
-0.04581232741475105,
-0.08587007969617844,
-0.055408984422683716,
-0.027579214423894882,
0.10315903276205063,
0.022015023976564407,
0.008790873922407627,
-0.050160787999629974,
-0.024248609319329262,
-0.01530710980296135,
0.04723683372139931,
0.0341404564678669,
0.0035553693305701017,
-0.05095713213086128,
-0.0630415603518486,
0.04639650136232376,
-0.06962376087903976,
0.01908356137573719,
0.0018342112889513373,
-0.07093138247728348,
0.0002954177325591445,
-0.021738670766353607,
0.039198897778987885,
0.021398112177848816,
0.0530574657022953,
0.07559937983751297,
0.029546886682510376,
-0.0183313749730587,
-0.024791982024908066,
0.005208482500165701,
0.030205802991986275,
0.013133376836776733,
-0.11126820743083954,
0.031088577583432198,
-0.05453900247812271,
-0.01752379722893238,
0.012496814131736755,
-0.0009502469329163432,
-0.009389995597302914,
-0.058912310749292374,
-0.02099713124334812,
-0.08577275276184082,
-0.022693296894431114,
0.04017407447099686,
-0.07371314615011215,
-0.03703409805893898,
0.08096703141927719,
0.02432240918278694,
-0.0014999499544501305,
-0.019916322082281113,
0.007453397382050753,
0.11313164234161377,
-0.10322581976652145,
0.029034258797764778,
0.05389900505542755,
0.0839640200138092,
0.06962346285581589,
-0.03940930217504501,
0.03349027782678604,
-0.0478050597012043,
-0.04110611230134964,
0.022427914664149284,
0.040838226675987244,
0.06413963437080383,
0.1888817548751831,
-0.05131443589925766,
-0.019248612225055695,
0.029932305216789246,
-0.007758606690913439,
0.04352392628788948,
-0.02750636450946331,
0.029171453788876534,
0.03238491341471672,
-0.038772691041231155,
0.022382616996765137,
0.03294803947210312,
0.051040053367614746,
-0.08102434128522873,
-0.015755554661154747,
0.1190284788608551,
-0.06283381581306458,
-0.0013704225420951843,
0.0475388839840889,
-0.0027571734972298145,
0.02276892587542534,
-0.03252330794930458,
0.034862544387578964,
-0.07772990316152573,
0.0020590913482010365,
-0.12354709208011627,
-0.02176107093691826,
0.0020477520301938057,
0.05560069903731346,
-0.038768112659454346,
-0.019487330690026283,
0.04584002122282982,
-0.04783996567130089,
0.02719699591398239,
0.06351786851882935,
0.03211556375026703,
-0.013651885092258453,
-0.043814387172460556,
-0.0016836854629218578,
0.08913057297468185,
-0.016691889613866806,
0.08183146268129349,
-0.03438032791018486,
3.0205890384523287e-33,
-0.04549519345164299,
0.07399710267782211,
0.07449860870838165,
-0.03641495481133461,
0.04557156190276146,
-0.04556722939014435,
0.04664645344018936,
0.029139526188373566,
-0.006926652975380421,
0.038160908967256546,
-0.06998955458402634,
-0.01746119000017643,
0.030975138768553734,
0.08631853759288788,
0.020107293501496315,
0.0651777982711792,
-0.02202851139008999,
0.05127015709877014,
-0.022690974175930023,
-0.0954650342464447,
0.0399874672293663,
-0.07896025478839874,
-0.05099235475063324,
-0.007126498036086559,
-0.015663854777812958,
0.003148196032270789,
0.029804734513163567,
-0.028537731617689133,
-0.08583774417638779,
0.00743982894346118,
-0.010015523992478848,
0.0780215933918953,
0.01220776978880167,
0.12342908978462219,
0.005445759277790785,
-0.04416763782501221,
0.054057203233242035,
-0.09403714537620544,
0.13395969569683075,
0.013265502639114857,
0.01693216897547245,
0.03414100408554077,
0.020053785294294357,
-0.0497928224503994,
0.0012795208021998405,
0.014157921075820923,
0.045461419969797134,
0.03331272676587105,
-0.05506141111254692,
-0.007561107166111469,
0.025048665702342987,
0.03325280547142029,
-0.058687515556812286,
-0.012356927618384361,
-0.03958841785788536,
0.032689545303583145,
0.017173688858747482,
0.04441147670149803,
-0.02648681215941906,
0.06719513237476349,
0.012917294166982174,
0.10731617361307144,
0.000469831662485376,
-0.015244035050272942,
-0.06298030912876129,
-0.006413796916604042,
-0.01418993342667818,
0.010530831292271614,
0.12674644589424133,
0.021841514855623245,
-0.03578957915306091,
-0.002751748776063323,
-0.062318045645952225,
-0.05384846404194832,
0.05077420547604561,
-0.009687048383057117,
0.053159214556217194,
-0.0781005248427391,
0.012984061613678932,
-0.044380124658346176,
-0.033115487545728683,
0.08831090480089188,
-0.04554446414113045,
-0.07507697492837906,
-0.05683412030339241,
-0.046040263026952744,
-0.02602929063141346,
-0.09151826053857803,
-0.006461430806666613,
-0.07431920617818832,
-0.006775717716664076,
-0.023257754743099213,
-0.0425899401307106,
0.05264963209629059,
0.0033733006566762924,
-3.6337156879107025e-33,
-0.043880436569452286,
0.03533601760864258,
-0.015288139693439007,
0.038060855120420456,
-0.02421874925494194,
-0.04102703928947449,
0.04510262608528137,
-0.05632235109806061,
-0.005955197848379612,
-0.0017084673745557666,
-0.032755330204963684,
-0.030126698315143585,
-0.03359896317124367,
-0.008349472656846046,
0.03931799903512001,
0.014562242664396763,
-0.06564836949110031,
0.06384813040494919,
0.027187323197722435,
0.04210299253463745,
0.011785541661083698,
0.07230163365602493,
-0.03474682196974754,
0.02496490068733692,
0.026124805212020874,
-0.04733777791261673,
-0.0011817299528047442,
-0.006585057359188795,
0.09982626885175705,
0.050698209553956985,
-0.04539228603243828,
0.007914269343018532,
-0.0812816396355629,
-0.011985125951468945,
0.05271439626812935,
-0.007321546785533428,
0.014868554659187794,
0.0017411299049854279,
-0.06429730355739594,
0.056739434599876404,
0.06359998881816864,
-0.08331070840358734,
-0.01956622302532196,
0.10113561898469925,
0.02404870092868805,
0.014151628129184246,
0.031119203194975853,
0.06835122406482697,
-0.024861007928848267,
-0.022027159109711647,
-0.057158153504133224,
-0.006483345292508602,
-0.05910997837781906,
0.012270512990653515,
-0.006918672472238541,
0.031994763761758804,
0.031295374035835266,
-0.05075603350996971,
0.10605939477682114,
-0.01151327881962061,
-0.039661213755607605,
-0.0349319763481617,
0.08510179817676544,
-0.016644509509205818,
-0.008510451763868332,
0.0007235294906422496,
0.011539343744516373,
-0.012172185815870762,
-0.009937703609466553,
0.031752221286296844,
-0.04072580486536026,
-0.04785281792283058,
-0.038877952843904495,
-0.037703629583120346,
-0.03763926029205322,
0.010813300497829914,
0.043525174260139465,
-0.04234025627374649,
-0.05619374290108681,
-0.028477229177951813,
-0.003888799110427499,
-0.014526104554533958,
0.04151919111609459,
0.0890054851770401,
-0.026992660015821457,
0.04820562154054642,
0.06065962091088295,
0.10913906246423721,
0.04373574256896973,
0.00722260307520628,
-0.06140074133872986,
0.11338824033737183,
-0.005009831860661507,
0.06988338381052017,
0.004106826148927212,
-4.34135110083389e-8,
-0.05406356602907181,
0.022425493225455284,
-0.00121497071813792,
0.044175609946250916,
-0.015309911221265793,
-0.000864666944835335,
-0.020560743287205696,
0.058840442448854446,
-0.048084624111652374,
-0.00003929260856239125,
0.06470027565956116,
0.04015471786260605,
-0.01782182976603508,
-0.03282869979739189,
-0.001328478567302227,
0.04896574839949608,
0.017392123118042946,
0.03899947926402092,
-0.03718031197786331,
0.06400159746408463,
0.01548613328486681,
0.018708674237132072,
0.019855177029967308,
-0.09232773631811142,
0.006535971537232399,
-0.11487207561731339,
-0.05919342115521431,
0.0449562668800354,
0.11982167512178421,
0.006960638333112001,
-0.01870555430650711,
-0.036955997347831726,
0.058053769171237946,
-0.027668746188282967,
0.03473205864429474,
0.09587926417589188,
-0.006405873689800501,
-0.04574039205908775,
0.03526948764920235,
-0.15466198325157166,
-0.03496512398123741,
0.05899221450090408,
-0.0030665940139442682,
0.0017447296995669603,
0.025149086490273476,
-0.00401608319953084,
0.028700897470116615,
-0.13018690049648285,
-0.04422399029135704,
-0.017140861600637436,
-0.024624081328511238,
-0.03365369513630867,
-0.0056279501877725124,
0.02813067100942135,
0.03469248488545418,
0.054542917758226395,
-0.04366808384656906,
-0.09076183289289474,
0.05152924731373787,
0.03967588394880295,
-0.036970093846321106,
-0.0390312559902668,
-0.13936761021614075,
-0.057833634316921234
] | https://github.com/netsysopt/gnns_sym_ilps | 0 | 0 | 0 | 0 |
Optimal Transport for Time Series Imputation | https://openreview.net/forum?id=xPTzjpIQNp | [
"Hao Wang",
"zhengnan li",
"Haoxuan Li",
"Xu Chen",
"Mingming Gong",
"BinChen",
"Zhichao Chen"
] | Poster | Missing data imputation through distribution alignment has demonstrated advantages for non-temporal datasets but exhibits suboptimal performance in time-series applications. The primary obstacle is crafting a discrepancy measure that simultaneously (1) captures temporal patterns—accounting for periodicity and temporal dependencies inherent in time-series—and (2) accommodates non-stationarity, ensuring robustness amidst multiple coexisting temporal patterns. In response to these challenges, we introduce the Proximal Spectrum Wasserstein (PSW) discrepancy, a novel discrepancy tailored for comparing two \textit{sets} of time-series based on optimal transport. It incorporates a pairwise spectral distance to encapsulate temporal patterns, and a selective matching regularization to accommodate non-stationarity. Subsequently, we develop the PSW for Imputation (PSW-I) framework, which iteratively refines imputation results by minimizing the PSW discrepancy. Extensive experiments demonstrate that PSW-I effectively accommodates temporal patterns and non-stationarity, outperforming prevailing time-series imputation methods. Code is available at https://github.com/FMLYD/PSW-I. | Time series, Imputation | null | 14,099 | null | [
-0.04911787435412407,
-0.04183187335729599,
0.05583927407860756,
0.05862775817513466,
0.03408418595790863,
0.04090508446097374,
-0.03599229082465172,
-0.04393680393695831,
0.0625348910689354,
-0.058923352509737015,
-0.013192282058298588,
-0.02236662432551384,
-0.022343898192048073,
0.014358923770487309,
-0.05392837896943092,
-0.08292090892791748,
0.06100619584321976,
0.003631248604506254,
-0.05352519080042839,
0.0020668578799813986,
-0.0035961666144430637,
0.05554221197962761,
-0.08163303881883621,
0.06752580404281616,
0.05124392360448837,
-0.009572931565344334,
0.005320917349308729,
-0.010493215173482895,
0.02050497941672802,
0.048467908054590225,
-0.06553598493337631,
0.04026574641466141,
-0.04040507972240448,
0.006387228146195412,
-0.030966652557253838,
0.025539658963680267,
0.015612957067787647,
0.04144562408328056,
0.010586448013782501,
-0.01275944896042347,
0.032629285007715225,
0.054624225944280624,
-0.013979952782392502,
0.06923482567071915,
-0.12567038834095,
-0.025116179138422012,
0.04273642972111702,
-0.08302807807922363,
-0.0018783349078148603,
0.0030654356814920902,
-0.029605146497488022,
0.10339009016752243,
0.0356954000890255,
0.14118990302085876,
0.04939579963684082,
0.04350925609469414,
-0.05175986513495445,
0.01549679134041071,
-0.004076958168298006,
0.04010571166872978,
-0.037031784653663635,
-0.07261360436677933,
-0.04897536709904671,
-0.005553265567868948,
0.00965081062167883,
0.020559359341859818,
0.05071462318301201,
0.032547928392887115,
0.019442468881607056,
0.05684967339038849,
-0.07916639745235443,
0.06032269075512886,
-0.1051207035779953,
0.11986564844846725,
0.023513993248343468,
0.08127619326114655,
0.047334812581539154,
-0.011390326544642448,
0.020389821380376816,
-0.05677666887640953,
-0.03184065595269203,
0.019930580630898476,
-0.02857714146375656,
0.0296433437615633,
0.09692521393299103,
-0.03511793166399002,
0.0010738623095676303,
0.05410582572221756,
0.010244430042803288,
-0.09418575465679169,
-0.04301305487751961,
-0.011311488226056099,
0.030010240152478218,
-0.02733982540667057,
-0.035398099571466446,
0.07251828163862228,
0.06717237085103989,
0.01787695474922657,
0.09735172986984253,
0.10349351912736893,
0.015869367867708206,
0.03993610292673111,
-0.10038015991449356,
0.026523062959313393,
-0.026000069454312325,
-0.06808318197727203,
0.014682608656585217,
0.008136318996548653,
-0.04059061035513878,
0.010092413984239101,
0.058409858494997025,
-0.015880268067121506,
0.04562024772167206,
0.020876817405223846,
0.007585578132420778,
-0.00831406656652689,
-0.010018941015005112,
0.012642585672438145,
0.009085743688046932,
0.04071776941418648,
-0.031104741618037224,
-0.04356003552675247,
0.09684377163648605,
0.05118825286626816,
0.028342029079794884,
0.0244074035435915,
0.03754516318440437,
4.163255689146988e-33,
-0.025461087003350258,
0.011819399893283844,
0.04101071506738663,
-0.07308053970336914,
-0.013230172917246819,
0.00673223240301013,
-0.051455236971378326,
-0.0364064946770668,
0.07690326124429703,
0.005882821511477232,
-0.04082101583480835,
0.03959970921278,
-0.0758635625243187,
-0.01592101715505123,
-0.06556659936904907,
-0.005306396167725325,
0.03840246424078941,
0.031237343326210976,
0.03568785637617111,
-0.005375484935939312,
0.06216368079185486,
-0.09601541608572006,
-0.014755305834114552,
-0.05565962940454483,
0.0525427907705307,
0.007413250394165516,
0.014292202889919281,
-0.06566958874464035,
0.038028791546821594,
0.017309706658124924,
-0.051154252141714096,
-0.007967468351125717,
0.034531883895397186,
0.010143589228391647,
0.02765919454395771,
-0.03684660419821739,
0.0394711047410965,
-0.04838404431939125,
0.015983514487743378,
-0.03428021818399429,
-0.03233484923839569,
0.05241110175848007,
-0.09941327571868896,
-0.07869887351989746,
-0.019469616934657097,
-0.13316698372364044,
0.16258004307746887,
-0.017597677186131477,
-0.018570750951766968,
0.004041650798171759,
-0.013653670437633991,
0.005295623559504747,
0.03748413920402527,
-0.12187574803829193,
-0.11319788545370102,
-0.0005778281483799219,
0.04204225540161133,
-0.08436288684606552,
0.04835314303636551,
0.11401033401489258,
-0.03554026409983635,
0.030726689845323563,
-0.017282437533140182,
0.0012916165869683027,
0.06459355354309082,
-0.011597895994782448,
-0.04254666715860367,
0.02048436366021633,
0.0166661087423563,
-0.01051249634474516,
0.045819178223609924,
-0.015311670489609241,
-0.0036287084221839905,
-0.01728309504687786,
0.07973384112119675,
-0.03858349472284317,
0.04998379573225975,
0.039583273231983185,
-0.05722451210021973,
0.003167321439832449,
-0.02658541314303875,
-0.04401465505361557,
0.012737895362079144,
-0.04021742567420006,
-0.06756867468357086,
-0.03072967939078808,
0.07963530719280243,
-0.06261628121137619,
-0.08499268442392349,
-0.14621654152870178,
0.027847768738865852,
0.052793797105550766,
-0.028850749135017395,
0.039823535829782486,
0.021467342972755432,
-4.358663257871962e-33,
0.01661144569516182,
0.017688557505607605,
0.09402035176753998,
-0.051072586327791214,
0.005039055831730366,
-0.012320801615715027,
0.008807255886495113,
-0.014238618314266205,
0.07681354880332947,
-0.01663319021463394,
0.015531184151768684,
-0.09987978637218475,
0.026805317029356956,
-0.05070945620536804,
-0.06334265321493149,
0.01896790787577629,
0.006200877949595451,
-0.048267021775245667,
0.0009781505214050412,
-0.060688067227602005,
0.060612522065639496,
0.03511970490217209,
-0.09218636900186539,
0.030702244490385056,
-0.007864556275308132,
0.012788335792720318,
0.006286012474447489,
0.031940195709466934,
-0.1012212261557579,
-0.025601588189601898,
-0.03397354483604431,
-0.05756692215800285,
-0.002922767074778676,
-0.10676293820142746,
-0.023995166644454002,
0.10028810799121857,
-0.03178658336400986,
0.10228583961725235,
-0.09347162395715714,
0.06543933600187302,
0.024710575118660927,
0.051596105098724365,
-0.06871508806943893,
-0.03078421950340271,
-0.04121509566903114,
0.03635522723197937,
-0.0906677097082138,
0.08550272136926651,
0.0797639936208725,
-0.024795757606625557,
0.01949341967701912,
0.018197739496827126,
-0.06345388293266296,
0.03746497631072998,
0.02353205345571041,
0.009107483550906181,
0.01690996065735817,
-0.03654293343424797,
0.009739524684846401,
-0.030205458402633667,
-0.07332374155521393,
-0.04248062148690224,
0.010394148528575897,
-0.03194733336567879,
0.004594604484736919,
0.0017748915124684572,
0.08311111479997635,
-0.049960531294345856,
0.05230262130498886,
0.010841554962098598,
-0.003819611156359315,
0.024607840925455093,
-0.020869271829724312,
-0.05217466130852699,
0.05656823888421059,
0.04221965745091438,
0.023675279691815376,
0.023802990093827248,
-0.05727499723434448,
0.03362548351287842,
-0.007207624148577452,
0.03816552087664604,
-0.07635122537612915,
0.053268350660800934,
-0.000910224043764174,
0.11889081448316574,
0.013806209899485111,
-0.05473456159234047,
0.011476404033601284,
-0.026230143383145332,
-0.07027727365493774,
0.07424313575029373,
-0.021241726353764534,
0.06210796907544136,
-0.023338232189416885,
-4.309032775040578e-8,
0.003028532722964883,
0.023548755794763565,
-0.018252268433570862,
-0.046877212822437286,
-0.006767242215573788,
-0.13320192694664001,
0.039690516889095306,
0.02983880788087845,
0.000914323260076344,
-0.018757212907075882,
0.07123611122369766,
-0.1001819595694542,
-0.06124880164861679,
-0.01886824518442154,
-0.008053906261920929,
-0.023912131786346436,
0.02511027082800865,
0.028281280770897865,
-0.048785850405693054,
0.015193609520792961,
-0.014284461736679077,
0.03222635015845299,
-0.02763451263308525,
-0.012241958640515804,
0.0808294266462326,
0.007393166422843933,
-0.0059260535053908825,
0.015069006942212582,
0.057625990360975266,
0.02254701405763626,
0.014480609446763992,
0.037311140447854996,
-0.0039467886090278625,
-0.035998858511447906,
-0.09524864703416824,
0.06396857649087906,
0.00784760806709528,
0.07648483663797379,
-0.18485282361507416,
0.010829833336174488,
-0.01454309094697237,
0.053366027772426605,
-0.08395470678806305,
0.008693101815879345,
0.057283367961645126,
-0.005871349945664406,
-0.018975403159856796,
0.012271001935005188,
0.0769176259636879,
-0.01940695196390152,
0.0435224324464798,
-0.037614427506923676,
0.06520996242761612,
0.016101177781820297,
0.06735775619745255,
0.01606416143476963,
0.02317679300904274,
-0.011150402016937733,
-0.008060656487941742,
-0.012017380446195602,
0.05479983985424042,
-0.0004853823047596961,
-0.09023474156856537,
0.03618396073579788
] | 0 | 0 | 0 | 0 |
|
Video Action Differencing | https://openreview.net/forum?id=3bcN6xlO6f | [
"James Burgess",
"Xiaohan Wang",
"Yuhui Zhang",
"Anita Rau",
"Alejandro Lozano",
"Lisa Dunlap",
"Trevor Darrell",
"Serena Yeung-Levy"
] | Poster | How do two individuals differ when performing the same action? In this work, we introduce Video Action Differencing (VidDiff), the novel task of identifying subtle differences between videos of the same action, which has numerous applications, such as coaching and skill learning. To enable development on this new task, we first create VidDiffBench, a benchmark dataset containing 549 video pairs, with human annotations of 4,469 fine-grained action differences and 2,075 timestamps indicating where these differences occur. Our experiments demonstrate that VidDiffBench poses a significant challenge for state-of-the-art large multimodal models (LMMs), such as GPT-4o and Qwen2-VL. By analyzing the failure cases of LMMs on VidDiffBench, we highlight two key challenges for this task: localizing relevant sub-actions over two videos and fine-grained frame comparison. To overcome these, we propose the VidDiff method, an agentic workflow that breaks the task into three stages: action difference proposal, keyframe localization, and frame differencing, each stage utilizing specialized foundation models. To encourage future research in this new task, we release the benchmark and code. | Video, Actions, Differencing, Zero-shot, benchmark, multimodal, lmm, llm | A new task and benchmark for comparing how an action is performed between two videos, with a zero-shot method | 14,085 | 2503.07860 | [
0.015557090751826763,
-0.1135774776339531,
0.012690183706581593,
-0.09107265621423721,
0.1236705332994461,
0.007071581203490496,
-0.028577517718076706,
-0.040125295519828796,
0.042542606592178345,
-0.09859012812376022,
-0.006003606133162975,
-0.08554716408252716,
-0.04766091704368591,
0.09431102126836777,
-0.01120414026081562,
-0.08995731920003891,
0.03477909415960312,
0.08165355026721954,
-0.12720395624637604,
0.008173086680471897,
0.03394556790590286,
-0.038393255323171616,
0.06572872400283813,
-0.01518980786204338,
0.020708579570055008,
0.04470245912671089,
-0.0062690614722669125,
-0.05952426791191101,
0.0030365565326064825,
-0.043585989624261856,
0.04761333763599396,
0.003991586621850729,
-0.031182045117020607,
0.01600981317460537,
-0.0513596311211586,
0.0786261335015297,
-0.059750452637672424,
0.024650538340210915,
-0.07116570323705673,
-0.07201032340526581,
-0.05589158833026886,
0.02068088762462139,
0.033942122012376785,
-0.01803884096443653,
0.035251934081315994,
0.07243373245000839,
0.019113538786768913,
-0.06540864706039429,
0.016222471371293068,
-0.004342072177678347,
-0.10092393308877945,
-0.05175099894404411,
0.02120385691523552,
0.00920172967016697,
-0.012181788682937622,
0.0527329184114933,
0.053782571107149124,
-0.028068259358406067,
0.037473469972610474,
-0.018071850761771202,
-0.022642526775598526,
-0.05406723916530609,
-0.01727580651640892,
-0.0036568953655660152,
-0.016483217477798462,
0.01842343993484974,
0.05776631459593773,
-0.04580564796924591,
0.023074757307767868,
-0.005866635590791702,
-0.0939333438873291,
0.049133528023958206,
0.01778550073504448,
0.029299406334757805,
-0.05625622719526291,
0.0022332691587507725,
-0.01786264218389988,
-0.00682040024548769,
0.014147739857435226,
-0.1462015062570572,
0.04644177109003067,
-0.08255518972873688,
0.039441805332899094,
-0.02214943990111351,
0.09904800355434418,
-0.024656491354107857,
0.027731284499168396,
0.019107116386294365,
-0.02400263398885727,
0.010812574997544289,
-0.08621660619974136,
0.04456205293536186,
0.04062189906835556,
-0.034022971987724304,
0.014268625527620316,
0.00840120017528534,
-0.014614873565733433,
-0.061157599091529846,
0.044718336313962936,
0.09266757220029831,
0.004380375146865845,
-0.006559000816196203,
-0.042689938098192215,
0.010952572338283062,
0.0455605611205101,
0.05903997644782066,
0.011449442245066166,
-0.00475193839520216,
-0.0019555126782506704,
-0.0503784716129303,
-0.0034969167318195105,
0.01613973081111908,
0.004299534950405359,
0.03476458787918091,
0.0723024532198906,
0.011960994452238083,
0.04485026374459267,
0.02295473963022232,
-0.035262707620859146,
0.018376082181930542,
0.0875258669257164,
-0.03706799075007439,
0.021442314609885216,
-0.05098773539066315,
0.03047543577849865,
-0.008204841986298561,
-0.04490281641483307,
1.7943924938722065e-33,
0.11088154464960098,
-0.010666795074939728,
0.07590124011039734,
0.09401874989271164,
0.06802301853895187,
-0.0018993733683601022,
-0.004548197612166405,
0.015577048063278198,
-0.029768142849206924,
-0.04546424001455307,
0.055864088237285614,
-0.044873613864183426,
-0.040429964661598206,
0.02618444710969925,
0.02343136817216873,
-0.0029827191028743982,
-0.10499133914709091,
-0.05939509719610214,
0.003781061852350831,
0.09592608362436295,
0.03888341784477234,
0.003819105913862586,
-0.026732169091701508,
0.05615566670894623,
-0.022494090721011162,
-0.012693599797785282,
0.06676215678453445,
-0.04605935513973236,
0.023332083597779274,
-0.013334427960216999,
-0.1251409500837326,
-0.05151341110467911,
-0.022641165181994438,
0.04625615105032921,
0.07287699729204178,
-0.021761775016784668,
-0.024943387135863304,
-0.06781842559576035,
0.028081469237804413,
-0.035593558102846146,
0.0013475560117512941,
-0.01715858466923237,
-0.04572319984436035,
-0.09247755259275436,
-0.06602836400270462,
0.021764736622571945,
0.03136177733540535,
0.06997054070234299,
-0.05725088343024254,
0.02651498094201088,
0.08764041215181351,
0.07459556311368942,
0.04300929978489876,
-0.10939741134643555,
-0.051089007407426834,
0.048285190016031265,
0.0794706717133522,
0.02185565046966076,
0.026107173413038254,
-0.019724957644939423,
0.004226564429700375,
0.049185167998075485,
-0.0031262144912034273,
0.10332424193620682,
0.0019106792751699686,
0.023350270465016365,
0.0235932357609272,
0.09230296313762665,
0.04422837123274803,
-0.015042729675769806,
-0.07090558111667633,
0.020283954218029976,
0.07045837491750717,
-0.04135608673095703,
0.08323539048433304,
-0.05344833433628082,
0.02389156073331833,
-0.08573783934116364,
0.015694361180067062,
-0.023268550634384155,
-0.04359237104654312,
0.0033770401496440172,
0.020228398963809013,
-0.0625225380063057,
-0.04915512353181839,
-0.10453974455595016,
0.000020775501980097033,
-0.07250566780567169,
-0.016825729981064796,
0.014497372321784496,
-0.03800921514630318,
-0.01919388398528099,
0.004435470327734947,
0.03778858855366707,
0.0809175968170166,
-1.5996775484592564e-33,
0.022571396082639694,
0.006775686517357826,
-0.01945621520280838,
0.07544343918561935,
0.05083484575152397,
-0.03371685370802879,
0.04847986623644829,
-0.0004526878474280238,
0.08875540643930435,
-0.044371701776981354,
-0.00507567822933197,
-0.04414621740579605,
-0.07237277179956436,
-0.05002089962363243,
0.0027037966065108776,
-0.03312437981367111,
0.02979406714439392,
0.01553219836205244,
-0.012990962713956833,
0.02314796671271324,
0.03774729371070862,
0.07474135607481003,
-0.04749017953872681,
0.006569994147866964,
-0.08146590739488602,
0.05238752439618111,
0.017564404755830765,
0.033432673662900925,
0.015018638223409653,
-0.03949563577771187,
0.021536480635404587,
-0.030632052570581436,
-0.025701897218823433,
-0.029956188052892685,
-0.02183743566274643,
0.028968850150704384,
0.03416579216718674,
-0.03982239216566086,
-0.050430987030267715,
0.12610740959644318,
0.060977138578891754,
-0.013397173956036568,
-0.06663163006305695,
-0.035055454820394516,
-0.009021017700433731,
0.04348296672105789,
-0.09630923718214035,
-0.030016088858246803,
-0.07828263938426971,
-0.018405182287096977,
-0.025946419686079025,
0.05062257871031761,
-0.13428530097007751,
-0.018642479553818703,
-0.024491826072335243,
-0.07022633403539658,
0.05621509626507759,
-0.007207668386399746,
0.03398189693689346,
0.03644891083240509,
-0.028396831825375557,
0.012996427714824677,
-0.05052383989095688,
0.028277475386857986,
0.024324020370841026,
0.012335386127233505,
-0.04331440478563309,
-0.03862495347857475,
0.00368045037612319,
0.02586318552494049,
0.06674859672784805,
-0.047851163893938065,
-0.015224598348140717,
-0.00838090106844902,
0.06049422547221184,
-0.018760502338409424,
0.010568101890385151,
-0.03681524470448494,
0.0311963502317667,
-0.05028489977121353,
0.000996932270936668,
-0.05472155287861824,
0.05225132778286934,
0.041910186409950256,
-0.041924331337213516,
0.20319764316082,
0.007211347110569477,
0.1455085724592209,
0.04122840613126755,
0.029192864894866943,
-0.04091598093509674,
0.04029575362801552,
0.030991366133093834,
0.0019449174869805574,
-0.00726701132953167,
-4.2506975717060413e-8,
-0.09121915698051453,
0.05945070832967758,
-0.018406443297863007,
-0.023441512137651443,
-0.048046935349702835,
-0.0013363079633563757,
-0.007318587973713875,
0.09528729319572449,
0.07332032918930054,
-0.013491760939359665,
0.07049410045146942,
0.0019323145970702171,
0.032691970467567444,
-0.0437973216176033,
-0.004804713185876608,
0.06216741353273392,
-0.04573502764105797,
0.025870829820632935,
-0.03668961673974991,
0.07197894901037216,
0.06422101706266403,
0.046598758548498154,
0.018144916743040085,
-0.032786089926958084,
-0.0516340546309948,
-0.03946983814239502,
-0.029524922370910645,
0.074897401034832,
0.019197512418031693,
0.01516689918935299,
0.03561045601963997,
0.01500082015991211,
0.024009810760617256,
-0.07007995247840881,
0.002912710653617978,
-0.0016526824329048395,
0.02739359624683857,
-0.045430902391672134,
0.04284239187836647,
0.058328863233327866,
-0.029591236263513565,
0.005283513572067022,
-0.005922746378928423,
0.04280295595526695,
-0.0012388895265758038,
0.06779110431671143,
-0.07969886064529419,
-0.1575257033109665,
0.06297054886817932,
-0.020442595705389977,
-0.009726908057928085,
0.010750262066721916,
-0.06390221416950226,
0.09893868118524551,
0.054926253855228424,
0.0409381203353405,
0.03784819692373276,
-0.01206638291478157,
0.10855968296527863,
-0.00556840980425477,
0.007522580213844776,
-0.03736383095383644,
-0.04044506698846817,
0.0545964390039444
] | 0 | 0 | 0 | 0 |
|
GANDALF: Generative AttentioN based Data Augmentation and predictive modeLing Framework for personalized cancer treatment | https://openreview.net/forum?id=WwmtcGr4lP | [
"Aishwarya Jayagopal",
"Yanrong Zhang",
"Robert John Walsh",
"Tuan Zea Tan",
"Anand D Jeyasekharan",
"Vaibhav Rajan"
] | Poster | Effective treatment of cancer is a major challenge faced by healthcare providers, due to the highly individualized nature of patient responses to treatment. This is caused by the heterogeneity seen in cancer-causing alterations (mutations) across patient genomes. Limited availability of response data in patients makes it difficult to train personalized treatment recommendation models on mutations from clinical genomic sequencing reports. Prior methods tackle this by utilising larger, labelled pre-clinical laboratory datasets (‘cell lines’), via transfer learning. These methods augment patient data by learning a shared, domain-invariant representation, between the cell line and patient domains, which is then used to train a downstream drug response prediction (DRP) model. This approach augments data in the shared space but fails to model patient-specific characteristics, which have a strong influence on their drug response. We propose a novel generative attention-based data augmentation and predictive modeling framework, GANDALF, to tackle this crucial shortcoming of prior methods. GANDALF not only augments patient genomic data directly, but also accounts for its domain-specific characteristics. GANDALF outperforms state-of-the-art DRP models on publicly available patient datasets and emerges as the front-runner amongst SOTA cancer DRP models. | personalized drug response prediction, cancer, genomic data augmentation, diffusion model, pseudolabelling | A cancer drug response prediction model that addresses the problem of limited labelled data through a novel genomic data augmentation technique. | 14,072 | null | [
-0.06851978600025177,
-0.0978148877620697,
0.0017997508402913809,
0.0016885852674022317,
0.02244952879846096,
-0.02347583882510662,
-0.028074245899915695,
0.05051426589488983,
0.0354582779109478,
-0.08543175458908081,
-0.04847102612257004,
-0.012860400602221489,
-0.01741068810224533,
-0.015815669670701027,
-0.05981125310063362,
-0.0013165257405489683,
0.029061859473586082,
0.05057520419359207,
-0.05523781105875969,
-0.021765369921922684,
-0.021333234384655952,
0.07816631346940994,
0.015937987715005875,
0.01728152111172676,
0.04354109615087509,
0.02652161568403244,
0.015299906022846699,
0.030851218849420547,
-0.0457751527428627,
-0.02410622127354145,
0.08803941309452057,
0.06869621574878693,
-0.07441437244415283,
0.08065937459468842,
-0.09297293424606323,
0.061560869216918945,
-0.13668902218341827,
0.11577058583498001,
-0.06976016610860825,
-0.07075575739145279,
0.05991010740399361,
-0.030569378286600113,
0.009180999360978603,
0.09556861221790314,
0.03708372265100479,
-0.09141705930233002,
-0.04941318556666374,
-0.07788705080747604,
0.10501126199960709,
0.06936407089233398,
-0.05188927426934242,
-0.026668023318052292,
-0.045371972024440765,
0.1021476536989212,
0.018189314752817154,
-0.019631357863545418,
0.05848076939582825,
-0.057896800339221954,
-0.036356713622808456,
-0.015762533992528915,
-0.05668332055211067,
-0.11637702584266663,
0.025838157162070274,
-0.06038602814078331,
0.006710461340844631,
-0.03419007733464241,
0.035276494920253754,
0.019967155531048775,
-0.08501221984624863,
-0.022890496999025345,
-0.032762277871370316,
0.06673410534858704,
-0.010313172824680805,
0.0027617779560387135,
0.048306435346603394,
0.052395276725292206,
0.013961848802864552,
0.06556079536676407,
0.07627072930335999,
-0.07984749227762222,
0.011573217809200287,
0.023578330874443054,
0.09265515953302383,
0.014710110612213612,
0.09204407781362534,
0.028700342401862144,
0.03363608196377754,
0.03822063282132149,
-0.0033883589785546064,
0.007581684272736311,
-0.011879494413733482,
0.05820133909583092,
0.015748169273138046,
0.035198889672756195,
-0.05847276747226715,
0.024018825963139534,
0.009994523599743843,
-0.08539973944425583,
-0.031184259802103043,
0.044347133487463,
-0.014314403757452965,
0.0035158146638423204,
0.0514991469681263,
0.017530934885144234,
0.025695644319057465,
-0.08031943440437317,
0.04301656410098076,
0.04876856878399849,
-0.017544465139508247,
-0.032033588737249374,
0.05911169573664665,
-0.0012253521708771586,
0.03405357897281647,
-0.018602341413497925,
0.02925804629921913,
-0.07746244221925735,
-0.022419800981879234,
-0.008388286456465721,
-0.051618658006191254,
0.0294394139200449,
-0.07126564532518387,
0.021810555830597878,
-0.0007843943894840777,
-0.003072438295930624,
-0.012980309315025806,
-0.007022003177553415,
-0.047356002032756805,
8.680458237456337e-34,
0.02115066722035408,
0.008547230623662472,
0.08478306978940964,
0.0360422283411026,
0.0148592134937644,
-0.06416475772857666,
-0.09360498189926147,
-0.03496788442134857,
-0.023176951333880424,
-0.0593617744743824,
-0.061589427292346954,
0.0030683951918035746,
-0.0580589734017849,
0.11172866821289062,
-0.035949379205703735,
0.026435989886522293,
-0.04374340921640396,
0.007199516985565424,
-0.032046157866716385,
0.02819930389523506,
-0.06607288867235184,
-0.01577986404299736,
-0.045049428939819336,
-0.0004906179383397102,
-0.008874304592609406,
0.028748568147420883,
-0.01523552741855383,
0.03398134186863899,
-0.01963212341070175,
0.005902666132897139,
-0.14822818338871002,
0.028210986405611038,
0.09332771599292755,
0.011706856079399586,
-0.018007701262831688,
-0.012322054244577885,
0.03967743366956711,
-0.07162583619356155,
0.002864339854568243,
0.013934047892689705,
0.03306039050221443,
0.036413904279470444,
0.003450309857726097,
0.006687133572995663,
-0.04476005211472511,
-0.07174427062273026,
0.03162131458520889,
0.0058205449022352695,
-0.05913468450307846,
-0.0064615788869559765,
0.0012718259822577238,
-0.07269179075956345,
-0.020593538880348206,
-0.018383130431175232,
-0.08821394294500351,
-0.010465758852660656,
-0.06898598372936249,
0.04151028394699097,
0.004790081642568111,
-0.004051268100738525,
0.04265493527054787,
0.033873818814754486,
0.03805312514305115,
0.06935705989599228,
-0.01738402247428894,
-0.006980560719966888,
-0.006982604507356882,
-0.019499707967042923,
0.047669824212789536,
0.07490172237157822,
-0.05089934542775154,
0.045452799648046494,
0.03776608780026436,
0.007248512003570795,
0.027709731832146645,
0.003580582793802023,
-0.008950411342084408,
-0.03549419343471527,
0.02581145614385605,
0.07208201289176941,
0.005249964538961649,
-0.0108466362580657,
-0.13263525068759918,
0.03738013282418251,
-0.07343596965074539,
-0.02310187742114067,
0.05657774582505226,
-0.04049545153975487,
-0.09586238861083984,
-0.035778772085905075,
-0.06304321438074112,
0.011271689087152481,
-0.0036680137272924185,
0.013505165465176105,
0.05552731454372406,
-2.997825039335678e-33,
0.02963336370885372,
0.05404679477214813,
0.053336795419454575,
-0.022055815905332565,
0.06846046447753906,
0.02412470430135727,
-0.04705345258116722,
0.025442419573664665,
0.06986147910356522,
-0.013767257332801819,
0.023299522697925568,
-0.03638507425785065,
0.04716731235384941,
-0.016098173335194588,
0.025067539885640144,
-0.014438392594456673,
-0.048697829246520996,
0.014457897283136845,
-0.09371591359376907,
0.006250862963497639,
0.03154623135924339,
0.07855909317731857,
-0.07944715023040771,
0.017832038924098015,
-0.0742887333035469,
0.007535963319242001,
-0.0059858886525034904,
0.015865739434957504,
-0.001559647498652339,
-0.11860907822847366,
-0.10212293267250061,
0.04290766268968582,
-0.032093487679958344,
-0.004710293374955654,
-0.01405079010874033,
0.07830048352479935,
0.09119812399148941,
-0.003531567519530654,
-0.0604446604847908,
0.03336066007614136,
0.05243147164583206,
-0.009469772689044476,
-0.17067711055278778,
0.029869450256228447,
-0.04154514521360397,
-0.012942738831043243,
0.023374438285827637,
0.04464510455727577,
0.080058753490448,
0.04226820170879364,
-0.02834405191242695,
0.06920801103115082,
-0.07128202170133591,
0.008377717807888985,
-0.05537128448486328,
-0.10479479283094406,
0.07728637009859085,
-0.08760032057762146,
0.043860532343387604,
0.002871155273169279,
-0.0702776312828064,
-0.037552960216999054,
-0.0002940140839200467,
-0.013713860884308815,
-0.04846518486738205,
-0.009541936218738556,
0.04409017041325569,
-0.06467379629611969,
0.0888497456908226,
-0.03283623605966568,
-0.004129808861762285,
-0.01839207112789154,
0.05477842688560486,
0.01674947515130043,
0.023044588044285774,
-0.09855984151363373,
0.019243303686380386,
-0.025154300034046173,
-0.033740025013685226,
-0.03350979834794998,
0.06654530018568039,
-0.045125991106033325,
0.023035217076539993,
0.011081208474934101,
0.04172134771943092,
0.029796214774250984,
0.11180470138788223,
0.04154708981513977,
-0.01677740551531315,
-0.06097855418920517,
-0.03104338049888611,
0.025099873542785645,
-0.05032578483223915,
0.045188773423433304,
-0.09034868329763412,
-3.707809881348112e-8,
-0.022501865401864052,
-0.018361778929829597,
0.041982252150774,
0.062283460050821304,
0.009747829288244247,
-0.016154395416378975,
-0.06715938448905945,
0.11225614696741104,
0.014375521801412106,
0.046910300850868225,
0.054688841104507446,
0.07439687848091125,
-0.052854474633932114,
-0.06290934979915619,
0.03790747746825218,
0.08081120997667313,
0.06217531859874725,
0.020290778949856758,
-0.006431218236684799,
0.026294849812984467,
-0.08400803804397583,
-0.001211046939715743,
0.02676672302186489,
-0.06330285221338272,
0.10303030908107758,
-0.04252748563885689,
-0.023225383833050728,
0.06179560720920563,
0.025136789306998253,
-0.01785336062312126,
0.07195661962032318,
-0.027520863339304924,
0.052010368555784225,
-0.02686186693608761,
-0.026806939393281937,
-0.002843310358002782,
0.08006494492292404,
-0.07130201905965805,
0.00565377390012145,
-0.04543524235486984,
0.06264808028936386,
0.11153038591146469,
-0.06928867846727371,
0.011453265324234962,
0.008753344416618347,
0.03485351428389549,
0.04581353813409805,
-0.06878427416086197,
0.02666102536022663,
-0.0499994158744812,
-0.00908695813268423,
-0.001351136714220047,
0.08854220062494278,
0.004431921523064375,
0.02935328334569931,
0.07298466563224792,
0.019078480079770088,
0.03444990515708923,
0.06370006501674652,
-0.048843249678611755,
0.04519542306661606,
-0.00636768015101552,
-0.03333009034395218,
-0.009352144785225391
] | 0 | 0 | 0 | 0 |
|
RaSA: Rank-Sharing Low-Rank Adaptation | https://openreview.net/forum?id=GdXI5zCoAt | [
"Zhiwei He",
"Zhaopeng Tu",
"Xing Wang",
"Xingyu Chen",
"Zhijie Wang",
"Jiahao Xu",
"Tian Liang",
"Wenxiang Jiao",
"Zhuosheng Zhang",
"Rui Wang"
] | Poster | Low-rank adaptation (LoRA) has been prominently employed for parameter-efficient fine-tuning of large language models (LLMs). However, the limited expressive capacity of LoRA, stemming from the low-rank constraint, has been recognized as a bottleneck, particularly in rigorous tasks like code generation and mathematical reasoning. To address this limitation, we introduce Rank-Sharing Low-Rank Adaptation (RaSA), an innovative extension that enhances the expressive capacity of LoRA by leveraging partial rank sharing across layers. By forming a shared rank pool and applying layer-specific weighting, RaSA effectively increases the number of ranks without augmenting parameter overhead. Our theoretically grounded and empirically validated approach demonstrates that RaSA not only maintains the core advantages of LoRA but also significantly boosts performance in challenging code and math tasks. Code, data and scripts are available at: https://github.com/zwhe99/RaSA. | parameter-efficient fine-tuning, large language model, low-rank adaptation | null | 14,067 | 2503.12576 | [
-0.07732425630092621,
-0.05122330039739609,
-0.015068784356117249,
0.054281458258628845,
0.06881856173276901,
0.04245951399207115,
-0.048394668847322464,
-0.005048288963735104,
-0.04284617677330971,
-0.05360891669988632,
-0.0686449334025383,
-0.046263307332992554,
0.08588909357786179,
0.010244586504995823,
0.015330773778259754,
0.03942490741610527,
0.05932474136352539,
0.07150335609912872,
-0.08281751722097397,
-0.07174087315797806,
-0.04126686602830887,
-0.05829577520489693,
0.0125120235607028,
-0.009980474598705769,
0.11024732887744904,
-0.03213527798652649,
-0.02706921100616455,
-0.030384555459022522,
0.020387724041938782,
-0.07036802172660828,
0.031144538894295692,
0.07151047140359879,
0.0011496309889480472,
0.04871010035276413,
-0.0786866620182991,
0.06021808087825775,
-0.0744004026055336,
-0.017077380791306496,
-0.03221597149968147,
-0.018841033801436424,
0.015543965622782707,
0.10300258547067642,
0.02407539263367653,
-0.015067421831190586,
0.04321878403425217,
-0.01564594730734825,
0.016192184761166573,
-0.05886457860469818,
-0.03611326590180397,
0.02254081331193447,
-0.06457693129777908,
-0.015475481748580933,
-0.07890355587005615,
0.05301092937588692,
-0.0276554673910141,
0.004600824322551489,
0.06229683384299278,
-0.049623146653175354,
0.013870389200747013,
-0.047403037548065186,
-0.08380354195833206,
-0.02216310240328312,
0.0698554590344429,
-0.00513834273442626,
0.056040238589048386,
-0.11562644690275192,
0.04908965155482292,
0.0732269361615181,
-0.06753291934728622,
0.028598211705684662,
-0.053076375275850296,
0.044864222407341,
-0.025769783183932304,
0.0806732326745987,
0.0052774823270738125,
0.041441354900598526,
0.020573129877448082,
0.0018747533904388547,
0.0803363174200058,
-0.032147977501153946,
0.005143773276358843,
0.03398841992020607,
-0.009368752129375935,
0.0523739792406559,
0.12442536652088165,
0.028863999992609024,
0.059414271265268326,
-0.01856791228055954,
-0.01060784887522459,
-0.01989489234983921,
0.06018470227718353,
0.018403155729174614,
0.08428315073251724,
0.07339911162853241,
-0.0035512621980160475,
0.016180885955691338,
-0.04171471670269966,
-0.05328548699617386,
-0.03922882676124573,
0.10456608235836029,
0.006567825097590685,
0.022780418395996094,
-0.002529591554775834,
-0.0562160350382328,
-0.00653240317478776,
-0.021070975810289383,
0.09979017823934555,
0.027296705171465874,
-0.035398952662944794,
-0.09847395867109299,
0.04871591553092003,
0.07563485950231552,
-0.019567886367440224,
0.018983125686645508,
-0.00447558518499136,
-0.02291181869804859,
0.06360768526792526,
-0.03606780245900154,
0.007391121704131365,
-0.03358123451471329,
-0.025468477979302406,
0.010548627004027367,
0.04049862548708916,
-0.006237752735614777,
0.038976237177848816,
0.03485679253935814,
-0.05883605033159256,
2.3113218284470692e-33,
0.09520544111728668,
0.027816468849778175,
-0.013918117620050907,
0.004405105486512184,
0.035488106310367584,
-0.017579268664121628,
0.10309408605098724,
-0.007912442088127136,
-0.11509518325328827,
-0.02835564687848091,
-0.04155182093381882,
0.023421872407197952,
-0.06306225806474686,
0.05807477980852127,
0.05660290643572807,
0.015136568807065487,
0.042314279824495316,
0.04217316210269928,
-0.04587256535887718,
-0.10054367780685425,
0.03968511521816254,
-0.02576967142522335,
-0.01321901474148035,
-0.04753413796424866,
0.026299720630049706,
0.06730393320322037,
0.025873014703392982,
-0.13157059252262115,
0.0043427306227386,
-0.004454864654690027,
-0.03536171093583107,
-0.0708787590265274,
0.04618573933839798,
0.009324433282017708,
0.015302539803087711,
-0.046036865562200546,
-0.009334595873951912,
-0.014248215593397617,
0.01875178888440132,
-0.02159012295305729,
0.0015877772821113467,
-0.0093703493475914,
-0.010450657457113266,
0.006967269815504551,
-0.05262521654367447,
-0.014788314700126648,
-0.0703926607966423,
0.06336989998817444,
-0.06320119649171829,
-0.04219609871506691,
0.013138152658939362,
0.028818542137742043,
-0.005051402375102043,
0.04095914587378502,
-0.0033437423408031464,
-0.04205533489584923,
0.006381544284522533,
0.02196635864675045,
0.04664354771375656,
0.04909941926598549,
-0.04044250026345253,
0.012251810170710087,
-0.07957235723733902,
0.08114653825759888,
0.031050652265548706,
-0.0414939783513546,
-0.030452605336904526,
0.0028681897092610598,
0.1029544547200203,
0.047348346561193466,
-0.03289660066366196,
-0.01957041770219803,
-0.05522751435637474,
0.03129074349999428,
-0.008399535901844501,
-0.0745856985449791,
0.017032215371727943,
-0.11894094944000244,
0.07805817574262619,
-0.009823355823755264,
0.03551728278398514,
0.043618038296699524,
0.06107812002301216,
-0.0528319887816906,
-0.0632709339261055,
-0.09553010761737823,
0.0449284091591835,
-0.04592861235141754,
-0.01723659038543701,
0.01615620031952858,
0.0005288414540700614,
-0.07422490417957306,
0.021645914763212204,
-0.08666477352380753,
0.029308047145605087,
-1.7384771663394674e-33,
0.012346310541033745,
0.007158704102039337,
-0.036016032099723816,
0.06425030529499054,
-0.023286551237106323,
-0.08899956196546555,
0.0012472847010940313,
-0.004332680720835924,
-0.09339317679405212,
0.00011813236051239073,
-0.05851282551884651,
-0.05481841042637825,
0.034040406346321106,
0.013843228109180927,
0.11780326813459396,
-0.06336955726146698,
-0.08121439814567566,
-0.009655621834099293,
-0.021790118888020515,
0.04688616469502449,
-0.0407668799161911,
0.12876366078853607,
0.002458244329318404,
0.05112289637327194,
-0.030350862070918083,
0.025873517617583275,
-0.03894776850938797,
0.036118946969509125,
-0.009991731494665146,
-0.07107697427272797,
-0.049713440239429474,
-0.0029101120308041573,
-0.059203337877988815,
-0.028935961425304413,
-0.033425699919462204,
0.01623034104704857,
0.045423537492752075,
-0.04182155057787895,
-0.07821361720561981,
0.12857559323310852,
-0.010418321006000042,
-0.08209719508886337,
0.00618436885997653,
0.013376203365623951,
0.012472405098378658,
0.09057532995939255,
-0.09870287030935287,
-0.024439387023448944,
-0.014292382635176182,
0.006900646258145571,
-0.026957402005791664,
-0.011212898418307304,
0.042747367173433304,
0.01266886293888092,
0.03603755310177803,
-0.08681686222553253,
-0.07645317167043686,
-0.021312549710273743,
0.014886794611811638,
-0.0013521881774067879,
0.02975866198539734,
-0.063510462641716,
0.03649777173995972,
-0.08586221933364868,
-0.013485508039593697,
0.011294553987681866,
-0.0353844128549099,
-0.047823600471019745,
-0.024977196007966995,
0.023485062643885612,
0.02816171571612358,
-0.06659416109323502,
-0.02502010203897953,
-0.002203765558078885,
-0.06650276482105255,
-0.008662068285048008,
-0.007337464950978756,
0.013806072995066643,
0.005121348891407251,
-0.06262945383787155,
0.005703709553927183,
0.06274139136075974,
0.026848815381526947,
0.009344516322016716,
0.01899283565580845,
0.08722510933876038,
0.11161159723997116,
0.060969650745391846,
0.02712368033826351,
0.014838284812867641,
-0.049518946558237076,
0.0029692689422518015,
0.017430942505598068,
0.0074836150743067265,
-0.08807658404111862,
-3.882082921791152e-8,
-0.08612509816884995,
0.0772760808467865,
-0.03296244889497757,
0.09094928950071335,
-0.003891951171681285,
-0.023918377235531807,
-0.038302868604660034,
0.04812174662947655,
0.04432539641857147,
0.06452001631259918,
0.04441674053668976,
-0.020015763118863106,
0.02988257259130478,
0.0034613064490258694,
0.0459258146584034,
0.08823394030332565,
-0.0244951993227005,
0.04417651891708374,
-0.010089426301419735,
-0.07746212184429169,
-0.022049516439437866,
0.1474437564611435,
-0.021425332874059677,
0.04807332903146744,
0.007140759378671646,
-0.05266588181257248,
-0.006416154559701681,
0.09440279752016068,
0.13596156239509583,
0.03829025849699974,
0.04772203788161278,
-0.0068453517742455006,
0.0699247270822525,
-0.02016182616353035,
0.03336299583315849,
0.057595960795879364,
0.0030375560745596886,
0.017604045569896698,
-0.021376581862568855,
-0.03879731521010399,
0.04307156801223755,
0.09564536809921265,
-0.02598542720079422,
0.047992851585149765,
0.04850668087601662,
-0.07587818056344986,
-0.052522625774145126,
-0.12290632724761963,
0.011598462238907814,
-0.029199901968240738,
0.008930841460824013,
-0.07938797026872635,
-0.06309369206428528,
0.034189607948064804,
0.04986395314335823,
0.0204519871622324,
-0.038007497787475586,
-0.02601572498679161,
0.06259771436452866,
0.026488952338695526,
-0.004184032790362835,
-0.038764696568250656,
-0.03760741651058197,
0.04881174489855766
] | 0 | 0 | 0 | 0 |
|
Scaling Speech-Text Pre-training with Synthetic Interleaved Data | https://openreview.net/forum?id=3tukjsVyrE | [
"Aohan Zeng",
"Zhengxiao Du",
"Mingdao Liu",
"Lei Zhang",
"shengmin jiang",
"Yuxiao Dong",
"Jie Tang"
] | Poster | Speech language models (SpeechLMs) accept speech input and produce speech output, allowing for more natural human-computer interaction compared to text-based large language models (LLMs).
Traditional approaches for developing SpeechLMs are constrained by the limited availability of unsupervised speech data and parallel speech-text data, which are significantly less abundant compared to text pre-training data, thereby limiting their scalability as LLMs.
We propose a novel approach to scaling speech-text pre-training by leveraging large-scale synthetic interleaved data derived from text corpora, eliminating the need for parallel speech-text datasets.
Our method efficiently constructs speech-text interleaved data by sampling text spans from existing text corpora and synthesizing corresponding speech spans using a text-to-token model, bypassing the need to generate actual speech.
We also employ a supervised speech tokenizer derived from an automatic speech recognition (ASR) model by incorporating a vector-quantized bottleneck into the encoder. This supervised training approach results in discrete speech tokens with strong semantic preservation even at lower sampling rates (e.g. 12.5Hz), while still maintaining speech reconstruction quality.
Starting from a pre-trained language model and scaling our pre-training to 1 trillion tokens (with 600B synthetic interleaved speech-text data), we achieve state-of-the-art performance in both speech language modeling and spoken question answering, improving performance on spoken questions tasks from the previous SOTA of 13\% (Moshi) to 31\%.
We further demonstrate that by fine-tuning the pre-trained model with speech dialogue data, we can develop an end-to-end spoken chatbot that achieves competitive performance comparable to existing baselines in both conversational abilities and speech quality, even operating exclusively in the speech domain. | large language models; speech language model; spoken chatbots | null | 14,059 | 2411.17607 | [
-0.09243906289339066,
-0.10310862958431244,
0.0394248440861702,
0.03389880433678627,
0.050975166261196136,
0.014624512754380703,
-0.018072394654154778,
-0.044112224131822586,
-0.02512032352387905,
-0.04073213413357735,
-0.02902974560856819,
-0.057784758508205414,
-0.04243595898151398,
0.004135495983064175,
0.029923690482974052,
-0.0008376362966373563,
0.061046913266181946,
0.12413577735424042,
-0.07456740736961365,
-0.032466016709804535,
0.07282586395740509,
0.07685180753469467,
0.03196179121732712,
0.0661894828081131,
0.06422950327396393,
0.021239375695586205,
-0.015634644776582718,
-0.017113110050559044,
0.07744614034891129,
0.021304721012711525,
0.03922703489661217,
-0.015137890353798866,
0.07859846949577332,
0.016076834872364998,
-0.0557016022503376,
0.0033605818171054125,
-0.0024499192368239164,
0.05804544314742088,
0.005666525103151798,
-0.04281488433480263,
0.03326829895377159,
-0.02585315704345703,
-0.03329857066273689,
0.04003732651472092,
0.02955293469130993,
-0.03970051929354668,
-0.04355068504810333,
-0.0063241021707654,
0.06040947511792183,
0.03698056936264038,
-0.06352747231721878,
-0.010772522538900375,
0.01006573997437954,
0.10474568605422974,
-0.061057012528181076,
-0.03227418661117554,
0.04594426602125168,
0.005133488215506077,
0.04448099061846733,
-0.005070528481155634,
-0.08274906873703003,
-0.11966603994369507,
0.04057246446609497,
-0.012660399079322815,
-0.06867998093366623,
-0.048949647694826126,
0.02070903405547142,
0.09075397998094559,
-0.006188463885337114,
0.00892632082104683,
-0.05914377048611641,
0.07969941198825836,
-0.042388297617435455,
0.08225911110639572,
-0.009558009915053844,
0.05714606121182442,
0.012147684581577778,
-0.027448182925581932,
0.11137652397155762,
0.02165808714926243,
0.03638153150677681,
0.004977551754564047,
0.0970010831952095,
-0.010847619734704494,
0.05929884314537048,
-0.0039709205739200115,
0.0090097701177001,
0.02514197677373886,
-0.02385588549077511,
0.009311927482485771,
-0.10863084346055984,
-0.005155670456588268,
0.022521018981933594,
-0.0078399907797575,
-0.07726380228996277,
0.04562757536768913,
-0.03612897917628288,
-0.04827004671096802,
0.034628450870513916,
0.07491384446620941,
0.030530955642461777,
0.05673391744494438,
0.0016787564381957054,
-0.016446851193904877,
-0.024597525596618652,
-0.0308632031083107,
0.03223292529582977,
-0.008840975351631641,
0.056378740817308426,
0.007039555348455906,
0.032699164003133774,
0.04817977175116539,
-0.04787846654653549,
0.0005467036389745772,
0.06433986872434616,
-0.007766222581267357,
-0.04995585232973099,
-0.046748727560043335,
0.03479968383908272,
0.027119094505906105,
-0.08081884682178497,
0.03884257376194,
-0.02756333164870739,
0.046698786318302155,
-0.09044862538576126,
-0.07663648575544357,
-0.02204078808426857,
-1.2222947298548576e-33,
0.019720086827874184,
0.02680162340402603,
0.06279730051755905,
0.019361436367034912,
0.00826406292617321,
-0.0511963814496994,
-0.02966342493891716,
0.06047937646508217,
-0.0890289694070816,
0.002527741016820073,
0.03206194192171097,
-0.01566459983587265,
-0.007858813740313053,
0.05805399268865585,
-0.025475112721323967,
-0.0044952984899282455,
-0.05444042757153511,
0.014887849800288677,
0.023645363748073578,
-0.06989341974258423,
0.054356228560209274,
-0.011377531103789806,
-0.001235291245393455,
0.05482201650738716,
0.07829286903142929,
-0.0013410112587735057,
0.04823656380176544,
-0.057405635714530945,
0.005536002106964588,
0.02467029169201851,
-0.14427398145198822,
-0.08392393589019775,
0.1004449874162674,
-0.0195013340562582,
0.012675889767706394,
-0.01955111138522625,
0.0414966456592083,
-0.019165826961398125,
-0.10148441046476364,
-0.07348648458719254,
-0.04341844096779823,
0.011655966751277447,
0.08078409731388092,
-0.03984947130084038,
-0.09196540713310242,
-0.07450584322214127,
0.006609618663787842,
0.012453950941562653,
-0.008188679814338684,
-0.0363558754324913,
0.02301952987909317,
0.031484805047512054,
-0.053721919655799866,
-0.017832130193710327,
0.016425134614109993,
-0.011047827079892159,
0.06731782853603363,
0.010253685526549816,
0.04125022143125534,
0.030232910066843033,
-0.04748042672872543,
0.02049105055630207,
-0.004200483672320843,
0.08849795907735825,
0.0025435814168304205,
-0.011200470849871635,
-0.04941451922059059,
0.05410996079444885,
0.01788201928138733,
-0.00017188802303280681,
-0.05062857270240784,
-0.004544151946902275,
-0.021917426958680153,
0.023097550496459007,
0.011406315490603447,
0.034310609102249146,
0.04201886057853699,
-0.04051167890429497,
0.008899329230189323,
0.09589478373527527,
0.012569041922688484,
-0.041025836020708084,
0.010411420837044716,
-0.10818677395582199,
0.028231149539351463,
-0.12771956622600555,
0.005787631031125784,
-0.10779856145381927,
0.02168326824903488,
-0.03580481931567192,
-0.1018795594573021,
-0.02281326800584793,
-0.04917098954319954,
0.040507908910512924,
-0.02382049709558487,
1.5652152683355124e-34,
0.030712850391864777,
0.037339113652706146,
-0.05174592137336731,
0.09987231343984604,
-0.03523727133870125,
-0.007702754344791174,
0.07199868559837341,
0.04278544336557388,
0.0071406252682209015,
-0.07906457036733627,
-0.04922337457537651,
-0.021214429289102554,
0.13443002104759216,
-0.03596869856119156,
0.05263911932706833,
0.021489804610610008,
0.01404075138270855,
0.06070949509739876,
-0.02936660125851631,
0.058652665466070175,
0.013418084941804409,
0.05245613679289818,
-0.027324208989739418,
0.04979574680328369,
-0.028755703940987587,
0.004248473793268204,
-0.09067490696907043,
0.04167952015995979,
-0.01399375032633543,
0.003149748547002673,
-0.043400466442108154,
-0.040569595992565155,
-0.018171586096286774,
-0.05156013369560242,
-0.10088597983121872,
0.04406604543328285,
0.047965969890356064,
-0.01782049797475338,
0.02039610967040062,
0.07350371778011322,
0.12114095687866211,
-0.021528145298361778,
-0.08210142701864243,
-0.040868934243917465,
-0.03841701149940491,
-0.03692122921347618,
-0.09594960510730743,
0.013056284748017788,
0.027497097849845886,
-0.011197848245501518,
-0.022696074098348618,
0.0442984439432621,
-0.008676261641085148,
0.06061103194952011,
0.01770748384296894,
-0.15662333369255066,
-0.04778723418712616,
-0.037116970866918564,
-0.019920848309993744,
-0.05670088902115822,
0.004763070493936539,
-0.07679683715105057,
0.0151450727134943,
-0.10930604487657547,
0.021556515246629715,
-0.05285291373729706,
0.006802670657634735,
0.0055444384925067425,
0.02705749310553074,
-0.03602194786071777,
0.0000353798022842966,
0.0055684675462543964,
-0.0027271986473351717,
0.05580056831240654,
-0.07482267171144485,
-0.034368377178907394,
-0.010137618519365788,
-0.12400888651609421,
-0.055713899433612823,
-0.03647942841053009,
0.08206269890069962,
0.02395843155682087,
0.03071404993534088,
0.03438686579465866,
0.08807794749736786,
0.11608374118804932,
-0.0017734483117237687,
0.03371362015604973,
0.028094930574297905,
0.06473752856254578,
-0.06698457896709442,
0.11861927062273026,
-0.01671750470995903,
0.07281343638896942,
-0.026738526299595833,
-3.150519845007693e-8,
-0.05676419287919998,
0.04367254674434662,
0.02862975373864174,
0.05058158561587334,
-0.008077031932771206,
-0.07415734231472015,
-0.01706228218972683,
-0.007519486825913191,
-0.0030892237555235624,
-0.022394247353076935,
0.02574124187231064,
-0.02310926839709282,
0.0031553790904581547,
-0.04229256883263588,
-0.023142347112298012,
0.012035680003464222,
-0.0021095802076160908,
-0.02870466187596321,
0.036388788372278214,
-0.07086502015590668,
0.02877979353070259,
0.13393986225128174,
0.0019536179024726152,
-0.00939843151718378,
0.0660184845328331,
-0.0460611991584301,
0.03204793483018875,
0.051416996866464615,
-0.004834793973714113,
-0.046398598700761795,
-0.06588058918714523,
-0.010119752958416939,
-0.10315705090761185,
-0.03285277262330055,
0.017003651708364487,
0.039468780159950256,
0.013749520294368267,
0.012224310077726841,
0.005553079303354025,
-0.05504017695784569,
0.029756873846054077,
0.09790133684873581,
-0.10165145993232727,
0.015943506732583046,
0.04544687643647194,
0.01547139510512352,
-0.05404091253876686,
-0.02586205303668976,
0.06354019045829773,
-0.029390176758170128,
-0.007344362325966358,
0.009934462606906891,
0.06320648640394211,
0.03426166996359825,
0.07800600677728653,
0.04835379496216774,
-0.056341491639614105,
0.08367110043764114,
0.05802641436457634,
0.03300828859210014,
0.0006462516612373292,
-0.06282919645309448,
-0.019527863711118698,
0.00690925819799304
] | 0 | 0 | 0 | 0 |
|
Offline Model-Based Optimization by Learning to Rank | https://openreview.net/forum?id=sb1HgVDLjN | [
"Rong-Xi Tan",
"Ke Xue",
"Shen-Huan Lyu",
"Haopu Shang",
"yaowang",
"Yaoyuan Wang",
"Fu Sheng",
"Chao Qian"
] | Poster | Offline model-based optimization (MBO) aims to identify a design that maximizes a black-box function using only a fixed, pre-collected dataset of designs and their corresponding scores. This problem has garnered significant attention from both scientific and industrial domains. A common approach in offline MBO is to train a regression-based surrogate model by minimizing mean squared error (MSE) and then find the best design within this surrogate model by different optimizers (e.g., gradient ascent). However, a critical challenge is the risk of out-of-distribution errors, i.e., the surrogate model may typically overestimate the scores and mislead the optimizers into suboptimal regions. Prior works have attempted to address this issue in various ways, such as using regularization techniques and ensemble learning to enhance the robustness of the model, but it still remains. In this paper, we argue that regression models trained with MSE are not well-aligned with the primary goal of offline MBO, which is to \textit{select} promising designs rather than to predict their scores precisely. Notably, if a surrogate model can maintain the order of candidate designs based on their relative score relationships, it can produce the best designs even without precise predictions. To validate it, we conduct experiments to compare the relationship between the quality of the final designs and MSE, finding that the correlation is really very weak. In contrast, a metric that measures order-maintaining quality shows a significantly stronger correlation. Based on this observation, we propose learning a ranking-based model that leverages learning to rank techniques to prioritize promising designs based on their relative scores. We show that the generalization error on ranking loss can be well bounded. Empirical results across diverse tasks demonstrate the superior performance of our proposed ranking-based method than twenty existing methods. Our implementation is available at \url{https://github.com/lamda-bbo/Offline-RaM}. | Offline model-based optimization, black-box optimization, learning to rank, learning to optimize | null | 14,057 | 2410.11502 | [
-0.06370458751916885,
-0.04153262451291084,
0.009572752751410007,
0.11670489609241486,
0.03596040606498718,
0.03878246620297432,
-0.06614837050437927,
0.06833809614181519,
-0.0489153228700161,
-0.026121776551008224,
-0.08893529325723648,
0.028299428522586823,
0.05942188575863838,
-0.00011046968575101346,
-0.09605327248573303,
-0.014738189056515694,
0.11840387433767319,
0.04966248571872711,
-0.06175397336483002,
0.007038688287138939,
0.050551071763038635,
0.03490322083234787,
-0.00901222787797451,
-0.016274048015475273,
0.033510033041238785,
-0.02388041839003563,
-0.0037870502565056086,
0.06418145447969437,
-0.043856553733348846,
-0.01835227571427822,
0.021228253841400146,
-0.022426022216677666,
0.0305862445384264,
-0.0411965548992157,
-0.01587953045964241,
0.026342177763581276,
-0.04340169206261635,
0.04741622135043144,
0.04271590709686279,
-0.0345865823328495,
-0.07103104144334793,
-0.04206858575344086,
-0.01574893295764923,
0.0010586811695247889,
0.10319170355796814,
-0.07855653017759323,
0.05813299119472504,
-0.0874311700463295,
-0.025186477228999138,
0.03912350907921791,
-0.06507790833711624,
-0.049024809151887894,
-0.07747237384319305,
0.028160054236650467,
0.02302796021103859,
0.06722988188266754,
0.020113492384552956,
-0.04296278581023216,
0.038971591740846634,
-0.02520967833697796,
-0.03920196741819382,
-0.12664343416690826,
-0.012663505040109158,
-0.03305275738239288,
0.01826571114361286,
-0.00817052274942398,
-0.00972860585898161,
0.033721674233675,
-0.0245269313454628,
0.0349116176366806,
0.006495852489024401,
0.08127909153699875,
-0.01897241733968258,
-0.008127408102154732,
0.09235676378011703,
0.04101303592324257,
0.07605790346860886,
0.005290469154715538,
0.031802285462617874,
-0.029830019921064377,
0.0058077373541891575,
0.028462208807468414,
-0.014284072443842888,
-0.007750279735773802,
0.0837024450302124,
0.042528942227363586,
-0.002525834133848548,
0.022506406530737877,
0.12512263655662537,
-0.027348726987838745,
0.020503295585513115,
0.003637331072241068,
0.01112578809261322,
0.042908910661935806,
-0.04139789193868637,
-0.059460125863552094,
0.021335745230317116,
-0.11365054547786713,
0.040778737515211105,
0.07275541871786118,
-0.05718809366226196,
0.06640727818012238,
0.03548264503479004,
-0.015556748025119305,
0.03703431040048599,
-0.09327855706214905,
0.0789065733551979,
0.0847616121172905,
0.07120265066623688,
-0.07730234414339066,
0.05544101074337959,
0.03818390890955925,
-0.009971514344215393,
0.07334752380847931,
-0.010088048875331879,
-0.0034201662056148052,
0.03170749917626381,
-0.012732915580272675,
-0.009502057917416096,
0.08280838280916214,
-0.03002977930009365,
0.00037128746043890715,
0.06661543250083923,
-0.0014293596614152193,
-0.017835553735494614,
0.04604937508702278,
-0.06684844195842743,
2.640116781921679e-33,
0.011766094714403152,
-0.00836392492055893,
0.08406076580286026,
-0.008346112444996834,
0.07814832776784897,
-0.08726575970649719,
0.01781243272125721,
-0.010744642466306686,
-0.0037167128175497055,
0.06270162761211395,
-0.07024912536144257,
-0.05100187286734581,
-0.023027272894978523,
0.11418446898460388,
0.12136686593294144,
0.021603934466838837,
-0.024598456919193268,
0.019874705001711845,
-0.034935202449560165,
-0.05711529403924942,
0.07734818011522293,
-0.07536128908395767,
0.014284754171967506,
-0.08630762994289398,
0.03659147396683693,
-0.002887740731239319,
0.05607481673359871,
-0.033411093056201935,
-0.06889105588197708,
0.02798444777727127,
-0.029687300324440002,
-0.0526701882481575,
0.04449262470006943,
0.0205070823431015,
-0.002658957615494728,
0.021684570237994194,
-0.01634870283305645,
-0.07097230106592178,
-0.01900288090109825,
0.042307741940021515,
-0.049334220588207245,
0.0016371870879083872,
0.03012688085436821,
0.000005075632998341462,
-0.013808006420731544,
-0.08499379456043243,
0.024203801527619362,
-0.021503111347556114,
-0.024965841323137283,
-0.0549100860953331,
-0.06978492438793182,
0.04636649042367935,
0.05669669434428215,
-0.03332056105136871,
-0.08942342549562454,
0.006011681631207466,
0.01845521852374077,
-0.00958738848567009,
0.0626695454120636,
0.01329602301120758,
-0.0773279145359993,
-0.0027350580785423517,
-0.009953818283975124,
-0.0272101741284132,
0.03397928550839424,
0.01383049227297306,
0.03925538435578346,
-0.013221260160207748,
-0.0174294151365757,
-0.0042064255103468895,
-0.012515166774392128,
0.026173822581768036,
0.049147527664899826,
-0.06902715563774109,
0.03049389459192753,
-0.009514335542917252,
0.08774259686470032,
0.027178412303328514,
0.08276604861021042,
-0.08396042138338089,
0.09085102379322052,
0.07683591544628143,
-0.03561026230454445,
-0.12657220661640167,
0.018912846222519875,
-0.019922886043787003,
0.0006888404604978859,
-0.042249105870723724,
-0.07476256787776947,
-0.011449883691966534,
-0.11869972944259644,
0.01969374716281891,
-0.09387350082397461,
-0.045435886830091476,
-0.01259509939700365,
-8.406169487661418e-34,
-0.05555832013487816,
-0.0066713085398077965,
0.013207188807427883,
-0.011125626042485237,
-0.01097326260060072,
-0.02117064595222473,
-0.036406293511390686,
-0.023607872426509857,
-0.05729225277900696,
-0.07214634865522385,
-0.018748760223388672,
-0.04982515424489975,
0.06260374188423157,
0.0044650062918663025,
0.06035951152443886,
0.0064656357280910015,
-0.059605129063129425,
-0.018476538360118866,
0.02328917756676674,
-0.005563091021031141,
0.046003587543964386,
0.026443079113960266,
-0.03420274704694748,
0.050091229379177094,
-0.06995969265699387,
0.0037529345136135817,
-0.07548550516366959,
0.1470930278301239,
-0.025216272100806236,
-0.00718087051063776,
0.03547714650630951,
0.033610712736845016,
-0.04817036911845207,
-0.031194938346743584,
-0.039313435554504395,
0.12021391093730927,
-0.08451851457357407,
-0.006441113073378801,
0.011280673556029797,
0.1396038681268692,
0.01751871034502983,
0.025240225717425346,
-0.09404397755861282,
-0.049158692359924316,
0.05258903279900551,
-0.022466428577899933,
-0.0020705233328044415,
-0.03384875878691673,
0.04192318394780159,
-0.017758123576641083,
-0.04909621179103851,
0.06959160417318344,
-0.03576848655939102,
0.054176799952983856,
-0.07049579918384552,
-0.045795198529958725,
-0.020505059510469437,
-0.0785214826464653,
0.06001976877450943,
0.01089413370937109,
0.05384085327386856,
-0.029029831290245056,
-0.008114808239042759,
0.002070683753117919,
-0.04463468864560127,
-0.005003616213798523,
0.035549502819776535,
0.07530810683965683,
0.011431981809437275,
0.043413661420345306,
-0.08190669864416122,
-0.08630359172821045,
0.06696797907352448,
0.022910278290510178,
-0.0651354268193245,
-0.01298573985695839,
-0.02947874553501606,
0.051883067935705185,
-0.028249941766262054,
0.05643921345472336,
0.06766276806592941,
0.015957951545715332,
0.04654694348573685,
0.0757487565279007,
0.07795872539281845,
0.019845284521579742,
0.0500897616147995,
0.09084679186344147,
0.050086408853530884,
-0.09498844295740128,
-0.04224478453397751,
-0.026011310517787933,
0.047590818256139755,
0.049248106777668,
-0.005979268811643124,
-4.7478554421331864e-8,
-0.04701827093958855,
0.02677621878683567,
0.10565454512834549,
0.03054625354707241,
0.021027637645602226,
0.002049580682069063,
-0.02976410463452339,
0.007160846143960953,
-0.03845364600419998,
-0.05882655829191208,
0.030230050906538963,
-0.05106157436966896,
-0.046543240547180176,
-0.047406360507011414,
-0.010537335649132729,
0.08229269087314606,
0.000441273907199502,
0.07708856463432312,
-0.007845453917980194,
-0.03114590421319008,
0.02610704116523266,
0.04570581763982773,
0.04864099994301796,
-0.05220945179462433,
0.09119155257940292,
-0.056968897581100464,
-0.023629747331142426,
0.012453692965209484,
0.059839483350515366,
0.06803430616855621,
-0.048775654286146164,
0.00473390007391572,
0.07423058152198792,
-0.009660504758358002,
0.014502931386232376,
0.0830349549651146,
0.10808981955051422,
-0.01655830256640911,
-0.06034862622618675,
0.022804610431194305,
0.03586835041642189,
0.017538627609610558,
-0.05902661383152008,
0.0005024217534810305,
0.025951122865080833,
-0.01508492510765791,
-0.031449414789676666,
-0.035052377730607986,
0.07726908475160599,
-0.050412192940711975,
0.054664984345436096,
-0.04731956496834755,
-0.014663321897387505,
0.0688854306936264,
0.06474290788173676,
0.004952073097229004,
-0.03219185024499893,
-0.03734888508915901,
-0.013396567665040493,
0.04654461517930031,
0.037326324731111526,
-0.03435195982456207,
-0.06241806969046593,
0.03411421179771423
] | 0 | 0 | 0 | 0 |
|
From Search to Sampling: Generative Models for Robust Algorithmic Recourse | https://openreview.net/forum?id=NtwFghsJne | [
"Prateek Garg",
"Lokesh Nagalapatti",
"Sunita Sarawagi"
] | Poster | Algorithmic Recourse provides recommendations to individuals who are adversely impacted by automated model decisions, on how to alter their profiles to achieve a favorable outcome. Effective recourse methods must balance three conflicting goals: proximity to the original profile to minimize cost, plausibility for realistic recourse, and validity to ensure the desired outcome. We show that existing methods train for these objectives separately and then search for recourse through a joint optimization over the recourse goals during inference, leading to poor recourse recommendations. We introduce GenRe, a generative recourse model designed to train the three recourse objectives jointly. Training such generative models is non-trivial due to lack of direct recourse supervision. We propose efficient ways to synthesize such supervision and further show that GenRe's training leads to a consistent estimator. Unlike most prior methods, that employ non-robust gradient descent based search during inference, GenRe simply performs a forward sampling over the generative model to produce minimum cost recourse, leading to superior performance across multiple metrics. We also demonstrate GenRe provides the best trade-off between cost, plausibility and validity, compared to state-of-art baselines. We release anonymized code at: https://anonymous.4open.science/r/GenRe-BD71 | Algorithmic recourse, explainability, generative modelling | We propose a generative model for recourse that outputs a distribution over likely recourse instances. | 14,050 | null | [
-0.09622137248516083,
-0.0785875916481018,
-0.00538701331242919,
0.08917233347892761,
0.05182258412241936,
0.03542345389723778,
0.01931905746459961,
0.017593298107385635,
-0.01559013593941927,
-0.06211647018790245,
-0.07752317935228348,
-0.06142212450504303,
0.06321799010038376,
0.02270493283867836,
-0.07174474000930786,
-0.04169641062617302,
0.14118853211402893,
0.12214428931474686,
-0.0213531032204628,
-0.026120886206626892,
0.018322736024856567,
0.07020304352045059,
0.042983267456293106,
-0.026364648714661598,
0.023055382072925568,
-0.05644439160823822,
-0.005858022253960371,
-0.018408501520752907,
-0.055049650371074677,
-0.012487932108342648,
0.03407948836684227,
0.06830044835805893,
-0.09120110422372818,
-0.015798380598425865,
-0.017681745812296867,
0.07239571958780289,
-0.07591322064399719,
0.053341131657361984,
-0.019647106528282166,
-0.017388956621289253,
-0.042944151908159256,
0.04509315267205238,
-0.008970484137535095,
-0.001548286760225892,
0.03530653566122055,
-0.04853378236293793,
0.0495772548019886,
-0.021794043481349945,
-0.044429223984479904,
0.06699956208467484,
-0.0301430094987154,
-0.040993932634592056,
-0.02263687551021576,
0.05287104472517967,
-0.00046887045027688146,
-0.004332948941737413,
0.07593211531639099,
-0.002216078108176589,
-0.00462747598066926,
-0.06442448496818542,
-0.04547692835330963,
-0.13378697633743286,
-0.09320275485515594,
-0.08619461208581924,
-0.04107782617211342,
-0.08665953576564789,
-0.010516863316297531,
0.032473016530275345,
-0.05879092216491699,
0.05048497021198273,
0.06585259735584259,
0.047415509819984436,
-0.08437369018793106,
0.04284900054335594,
0.05904504284262657,
0.0781397596001625,
0.0174128245562315,
0.015334034338593483,
0.04534412920475006,
-0.11276715248823166,
-0.00954470969736576,
0.01312398724257946,
0.0246350709348917,
0.032006796449422836,
0.1188773587346077,
-0.025206493213772774,
0.00550722936168313,
0.05202944576740265,
0.1258203536272049,
0.02965731918811798,
-0.06747038662433624,
0.03950474411249161,
0.05266549810767174,
-0.04427153244614601,
-0.003672374412417412,
0.03368381783366203,
-0.00549441808834672,
-0.050456900149583817,
0.03564348816871643,
0.07935479283332825,
-0.02732343226671219,
0.12013527005910873,
-0.013967400416731834,
-0.1283615082502365,
0.0014120112173259258,
-0.008321668021380901,
0.05932353064417839,
0.06124819442629814,
0.0036946297623217106,
-0.061658911406993866,
-0.03175916150212288,
0.049037836492061615,
0.03347061201930046,
-0.0201876163482666,
0.06074683740735054,
0.01712741144001484,
-0.030424784868955612,
0.011136962100863457,
0.0059238458052277565,
0.04334423318505287,
-0.02818809635937214,
0.07098845392465591,
0.024328123778104782,
-0.011416459456086159,
-0.016439057886600494,
-0.04026489332318306,
-0.11870338767766953,
3.1120459886951566e-33,
0.038505297154188156,
0.07233116030693054,
0.017144938930869102,
0.03284454718232155,
0.04812394455075264,
-0.019438546150922775,
-0.08108143508434296,
0.09489041566848755,
0.008054299280047417,
-0.00983144249767065,
-0.012545263394713402,
0.008460781536996365,
-0.08618990331888199,
0.003926437348127365,
0.06354027986526489,
-0.0019797352142632008,
-0.03261084109544754,
0.14831450581550598,
-0.05289533734321594,
-0.02176707796752453,
-0.0023144667502492666,
-0.029583357274532318,
-0.02514410763978958,
0.0032565444707870483,
0.024743199348449707,
-0.006873530335724354,
0.03247757628560066,
-0.03305259719491005,
0.04494517296552658,
0.0011063569691032171,
-0.051520053297281265,
-0.0209334883838892,
0.05669442564249039,
0.011966710910201073,
0.04409022256731987,
0.005257847253233194,
-0.05535870045423508,
-0.0032046043779700994,
-0.025296766310930252,
0.05257321149110794,
-0.09090518206357956,
0.01780756004154682,
0.018892159685492516,
0.0029912872705608606,
-0.06600107997655869,
-0.08679025620222092,
0.033780574798583984,
-0.08575023710727692,
-0.04732813686132431,
0.023687440901994705,
0.043613333255052567,
0.02963121607899666,
0.08333459496498108,
-0.004578512627631426,
-0.10050737857818604,
0.001103110029362142,
-0.04839794710278511,
-0.0031144318636506796,
-0.020009085536003113,
-0.008098560385406017,
0.058933936059474945,
0.08321990072727203,
-0.011632693000137806,
0.009510543197393417,
-0.04724416136741638,
0.016177047044038773,
0.048356492072343826,
0.043361928313970566,
0.08153757452964783,
0.009647526778280735,
-0.020134689286351204,
0.01931336522102356,
0.00717851473018527,
-0.04873467981815338,
0.07210134714841843,
-0.0021961601451039314,
0.04795364663004875,
-0.028103649616241455,
-0.018495166674256325,
-0.03211623430252075,
-0.02260187454521656,
0.03685900196433067,
0.0015389438485726714,
0.0028425503987818956,
-0.019879044964909554,
-0.058704447001218796,
0.02141578681766987,
-0.026395419612526894,
0.0027808432932943106,
-0.0031634292099624872,
-0.06832811236381531,
-0.018418148159980774,
-0.14263887703418732,
0.02653416618704796,
0.06651035696268082,
-2.686947925845462e-33,
-0.016752738505601883,
-0.01818760298192501,
0.03616412356495857,
-0.024486973881721497,
-0.0163761917501688,
-0.05189061909914017,
-0.049030907452106476,
-0.03377266228199005,
0.08215329796075821,
-0.11705124378204346,
-0.12293031066656113,
-0.11533734202384949,
0.07640746980905533,
-0.00806781742721796,
-0.02394345961511135,
-0.0027716271579265594,
-0.08205090463161469,
0.014850995503365993,
-0.06262481212615967,
0.017055610194802284,
0.036011338233947754,
0.0439622588455677,
-0.06871673464775085,
0.002229056553915143,
-0.056668493896722794,
0.08992898464202881,
-0.02146216109395027,
0.06535226106643677,
0.007895640097558498,
0.022436968982219696,
-0.02770419418811798,
-0.03979182615876198,
-0.04833373427391052,
-0.03390785679221153,
-0.07108715176582336,
0.07200466096401215,
0.009680950082838535,
0.02643447183072567,
-0.09224158525466919,
0.0903848260641098,
0.07377728819847107,
0.04609246924519539,
-0.014827512204647064,
0.034748394042253494,
0.009495413862168789,
-0.021962886676192284,
-0.016991576179862022,
-0.038543082773685455,
0.07619358599185944,
-0.018799230456352234,
-0.02315027080476284,
-0.006211407016962767,
-0.07950811088085175,
0.04717407003045082,
-0.0505228117108345,
-0.11952698975801468,
0.008498681709170341,
-0.06977076828479767,
0.07541295886039734,
0.026997512206435204,
-0.037519220262765884,
0.02116946503520012,
-0.006570377387106419,
0.023620007559657097,
-0.006091600749641657,
-0.05382921174168587,
0.03224702179431915,
-0.056426580995321274,
0.05553240329027176,
-0.012871774844825268,
0.0024092532694339752,
0.01701928675174713,
0.03076106496155262,
0.006501303985714912,
0.021290315315127373,
0.03282390534877777,
0.06223159655928612,
-0.03866906464099884,
-0.04257757589221001,
-0.03587906062602997,
0.08637221157550812,
-0.017321715131402016,
0.026841120794415474,
-0.026523027569055557,
0.0699215680360794,
0.005154434125870466,
0.0023728033993393183,
0.07568518817424774,
-0.044872939586639404,
0.021814420819282532,
-0.02329433523118496,
-0.0738447904586792,
0.05811643600463867,
0.04123874753713608,
-0.05029907822608948,
-4.321975666243816e-8,
-0.0057283309288322926,
0.06449376046657562,
0.05278144031763077,
0.08221792429685593,
0.04180696979165077,
-0.02383696660399437,
-0.05988365411758423,
0.11513902246952057,
0.011850986629724503,
-0.035636045038700104,
0.01743173412978649,
-0.05853378772735596,
0.0009612506255507469,
-0.01908957213163376,
-0.0034799613058567047,
0.019741928204894066,
-0.01376539096236229,
0.021514423191547394,
-0.05824714154005051,
0.03926083445549011,
0.013450601138174534,
0.058071259409189224,
-0.03163030743598938,
-0.04467374086380005,
0.08464256674051285,
-0.048772357404232025,
0.02302752062678337,
0.047889504581689835,
0.059675443917512894,
0.06692430377006531,
-0.03339211642742157,
0.03815706819295883,
0.0021271556615829468,
-0.01914912648499012,
-0.037107814103364944,
-0.002955760806798935,
0.01847529038786888,
-0.03677542880177498,
-0.02005932480096817,
-0.07148406654596329,
0.03787505626678467,
0.13045364618301392,
-0.032734014093875885,
0.056912172585725784,
0.08635060489177704,
-0.012933566235005856,
0.035820215940475464,
-0.04089909419417381,
0.0581679567694664,
-0.004256790969520807,
-0.033424682915210724,
-0.07817790657281876,
0.07571431994438171,
-0.00825216993689537,
0.0076530249789357185,
-0.013615312986075878,
0.007946177385747433,
-0.048621766269207,
0.011405757628381252,
0.054054826498031616,
-0.0006297621876001358,
-0.05576767772436142,
-0.0011211445089429617,
-0.002461864147335291
] | 0 | 0 | 0 | 0 |
|
Neural Wave Equation for Irregularly Sampled Sequence Data | https://openreview.net/forum?id=kbeX97jExm | [
"Arkaprava Majumdar",
"M Anand Krishna",
"P. K. Srijith"
] | Poster | Sequence labeling problems arise in several real-world applications such as healthcare and robotics. In many such applications, sequence data are irregularly sampled and are of varying complexities. Recently, efforts have been made to develop neural ODE-based architectures to model the evolution of hidden states continuously in time, to address irregularly sampled sequence data. However, they assume a fixed architectural depth and limit their flexibility to adapt to data sets with varying complexities. We propose the neural wave equation, a novel deep learning method inspired by the wave equation, to address this through continuous modeling of depth. Neural Wave Equation models the evolution of hidden states continuously across time as well as depth by using a non-homogeneous wave equation parameterized by a neural network. Through d'Alembert's analytical solution of the wave equation, we also show that the neural wave equation provides denser connections across the hidden states, allowing for better modeling capability. We conduct experiments on several sequence labeling problems involving irregularly sampled sequence data and demonstrate the superior performance of the proposed neural wave equation model. | Wave Equation, Neural ODE, Sequence Labelling | Partial Differential Equations parameterised by a Neural Network (like Neural ODE) can be used to solve sequence modeling problems. We hypothesize why this might be the case and demonstrate that it outpeforms many known continuous RNN models. | 14,044 | null | [
-0.20106205344200134,
-0.02374332956969738,
0.029639944434165955,
-0.008521290495991707,
-0.04502597451210022,
0.047096751630306244,
-0.061424244195222855,
-0.04216437414288521,
0.03271659091114998,
-0.10851266235113144,
0.004713914357125759,
-0.043219972401857376,
-0.06871851533651352,
0.01880830153822899,
-0.01760910078883171,
0.01919190213084221,
-0.026213768869638443,
0.006172658409923315,
-0.03046979196369648,
-0.08128679543733597,
0.038967300206422806,
0.10755617916584015,
-0.024188391864299774,
0.04378403723239899,
0.05855648219585419,
0.07611197978258133,
0.1099793016910553,
-0.036701977252960205,
0.01841793954372406,
0.03424302116036415,
0.11658620089292526,
-0.004244234878569841,
-0.028818432241678238,
-0.031245464459061623,
-0.0047389366663992405,
0.022034289315342903,
-0.10331934690475464,
-0.004862534813582897,
-0.06811964511871338,
-0.05789093300700188,
0.08546572178602219,
0.09128613024950027,
-0.003662589704617858,
0.09877264499664307,
0.08874388784170151,
0.023950349539518356,
0.057238537818193436,
-0.04509168490767479,
0.04867778718471527,
0.03424295783042908,
-0.05404435098171234,
0.016794830560684204,
0.008915270678699017,
0.08037351071834564,
0.03269504755735397,
0.035919737070798874,
-0.026993410661816597,
0.008282328955829144,
-0.05405304208397865,
0.015102701261639595,
0.0538189522922039,
-0.002845321549102664,
0.07583154737949371,
-0.0651664137840271,
-0.025972409173846245,
0.05665237084031105,
-0.0012503566686064005,
0.08771286904811859,
0.0008410607115365565,
0.07153588533401489,
-0.023023927584290504,
0.10409146547317505,
0.02909398265182972,
-0.009938652627170086,
-0.044402070343494415,
-0.04630938917398453,
0.016002656891942024,
0.0010456395102664828,
0.060633979737758636,
-0.017092574387788773,
-0.02108212374150753,
-0.07095944136381149,
-0.03590107709169388,
0.047804344445466995,
0.003786373883485794,
0.04048856720328331,
-0.029908621683716774,
0.005913466215133667,
-0.04554366320371628,
-0.025489140301942825,
-0.038501590490341187,
-0.07190313935279846,
0.04166155681014061,
-0.07749758660793304,
0.01882029138505459,
0.02086544968187809,
0.004440797958523035,
-0.021898243576288223,
0.014970014803111553,
0.084200419485569,
-0.025535475462675095,
0.054146114736795425,
0.01896793209016323,
0.07325920462608337,
0.001509851193986833,
-0.0021278567146509886,
0.016926873475313187,
0.04520978406071663,
0.012173260562121868,
0.018487079069018364,
0.034237299114465714,
0.025575660169124603,
0.058057013899087906,
0.061082594096660614,
0.05232017859816551,
-0.09344398975372314,
0.037483666092157364,
-0.11160096526145935,
-0.024071136489510536,
0.07948369532823563,
-0.07006293535232544,
-0.0553152859210968,
-0.09261655807495117,
0.06419841945171356,
0.02124442532658577,
0.017004281282424927,
-0.002436666516587138,
-2.483668723485054e-34,
0.04234657436609268,
-0.017072517424821854,
0.06267514824867249,
-0.06271857023239136,
0.015562169253826141,
-0.035598400980234146,
0.011099901050329208,
0.0036850585602223873,
0.011344937607645988,
0.07783559709787369,
-0.138087660074234,
-0.03406142070889473,
-0.07905387133359909,
0.10506634414196014,
-0.024460257962346077,
-0.04039886221289635,
-0.04908943176269531,
0.06701740622520447,
-0.00528626749292016,
-0.03826424479484558,
0.004297195002436638,
-0.012661416083574295,
-0.029066983610391617,
-0.027772720903158188,
-0.008660199120640755,
0.05160694569349289,
0.0038176008965820074,
0.02121402509510517,
0.04469041898846626,
-0.020924942567944527,
-0.06635976582765579,
0.056352049112319946,
0.006153289694339037,
0.03371883183717728,
0.08439905196428299,
0.0607483945786953,
0.09274053573608398,
0.0024168461095541716,
0.019506419077515602,
0.0008633656543679535,
0.015581732615828514,
0.004470186773687601,
0.054752595722675323,
0.037174392491579056,
-0.023228948935866356,
-0.08320341259241104,
0.024769825860857964,
-0.012408697046339512,
-0.021757351234555244,
-0.026890847831964493,
-0.013865353539586067,
-0.03178665041923523,
-0.04371195286512375,
-0.07031917572021484,
-0.0348936952650547,
-0.028697675094008446,
0.011382847093045712,
0.028281254693865776,
-0.03295678645372391,
0.07696019858121872,
-0.008718556724488735,
-0.021511346101760864,
0.08261974155902863,
0.05314410477876663,
0.03962098807096481,
0.054121002554893494,
-0.015195130370557308,
0.06281440705060959,
0.10546315461397171,
0.015162110328674316,
0.03766394779086113,
0.028385940939188004,
-0.07733224332332611,
-0.03547998517751694,
0.028936175629496574,
0.0225522480905056,
0.0033960796426981688,
-0.04041478410363197,
-0.06638950109481812,
0.07334736734628677,
0.006500582210719585,
-0.043021634221076965,
-0.03253067284822464,
0.04836353659629822,
-0.0052169510163366795,
0.011227406561374664,
0.06905058771371841,
-0.07703649252653122,
-0.06157271936535835,
-0.08110137283802032,
-0.12367959320545197,
-0.0018621174385771155,
0.03156536817550659,
-0.027454547584056854,
0.03562304750084877,
-2.331370251942328e-33,
0.006323764566332102,
0.021486438810825348,
-0.04235706850886345,
-0.009582296013832092,
0.008308464661240578,
-0.03316008299589157,
-0.04832484945654869,
0.08521416038274765,
0.02518465928733349,
-0.0036973776295781136,
-0.00531771220266819,
-0.025029335170984268,
0.05566450208425522,
-0.004879688378423452,
0.019786175340414047,
0.017912283539772034,
-0.019120609387755394,
-0.03502807393670082,
0.032213807106018066,
-0.0087993573397398,
0.014170468784868717,
0.025703230872750282,
-0.05006341263651848,
-0.011800804175436497,
-0.06867498904466629,
-0.02442854456603527,
-0.04316244274377823,
0.07116905599832535,
0.027652347460389137,
-0.02617134526371956,
-0.0917210802435875,
-0.09093853831291199,
0.004935234319418669,
-0.04621611535549164,
-0.0790758952498436,
0.09059745818376541,
0.07777344435453415,
-0.030716288834810257,
-0.032065730541944504,
-0.07805898785591125,
0.017273709177970886,
-0.06688360124826431,
-0.03745963051915169,
0.018745234236121178,
-0.04564310982823372,
-0.007500598207116127,
-0.058025162667036057,
0.002097661606967449,
-0.04290841519832611,
0.06812961399555206,
-0.04390563815832138,
0.03554105386137962,
0.0028912019915878773,
-0.008892027661204338,
-0.07826240360736847,
0.026753446087241173,
-0.05550415813922882,
0.06164862588047981,
0.027513621374964714,
0.01861572451889515,
-0.03722483664751053,
-0.07870548963546753,
0.01937578059732914,
-0.0634155124425888,
-0.016088955104351044,
0.004813321866095066,
-0.027462104335427284,
0.0023145321756601334,
0.02788075990974903,
-0.05834126099944115,
0.04608545079827309,
-0.042705927044153214,
0.0517345629632473,
0.08758141845464706,
-0.027518466114997864,
-0.13247838616371155,
-0.05102726072072983,
-0.031226862221956253,
-0.0046828133054077625,
-0.012826803140342236,
0.045293960720300674,
0.0038638904225081205,
0.009978548623621464,
0.08358260244131088,
0.01708727516233921,
-0.015317603014409542,
0.025914287194609642,
-0.041640330106019974,
-0.035661667585372925,
0.055354952812194824,
-0.011461354792118073,
0.056417230516672134,
-0.09778496623039246,
-0.027034282684326172,
-0.05272456631064415,
-4.4339550697714e-8,
-0.03363099694252014,
0.08498910814523697,
0.0586073212325573,
0.006606339477002621,
0.07662534713745117,
-0.07772480696439743,
0.0220379326492548,
-0.007631738670170307,
-0.0034259986132383347,
-0.10078044980764389,
0.08865683525800705,
-0.05233520269393921,
-0.06485787779092789,
-0.10700751096010208,
-0.06793589144945145,
0.08456189185380936,
-0.00858706422150135,
-0.049352504312992096,
0.015269724652171135,
0.02188393473625183,
-0.03158888593316078,
0.023889467120170593,
-0.018395593389868736,
0.011824694462120533,
0.06591016054153442,
-0.019156312569975853,
-0.0468483492732048,
0.008351150900125504,
0.06856035441160202,
0.06130877137184143,
0.02931511588394642,
0.021086370572447777,
0.034273821860551834,
-0.004297428298741579,
-0.05291871353983879,
0.012019675225019455,
0.005931483116000891,
-0.05953162536025047,
-0.03985123336315155,
-0.07086920738220215,
-0.09515108168125153,
0.08526013046503067,
-0.01879054866731167,
-0.032377153635025024,
0.025318004190921783,
-0.020153973251581192,
0.059744544327259064,
-0.07508745789527893,
0.07427459955215454,
0.02198784239590168,
0.001851817942224443,
0.026382779702544212,
0.01845036819577217,
-0.028813326731324196,
0.10403837263584137,
0.014077754691243172,
0.002048743888735771,
-0.06363724917173386,
0.004998665302991867,
0.1289711892604828,
-0.042845066636800766,
0.07609497010707855,
0.00022128946147859097,
0.010758252814412117
] | 0 | 0 | 0 | 0 |
|
ComaDICE: Offline Cooperative Multi-Agent Reinforcement Learning with Stationary Distribution Shift Regularization | https://openreview.net/forum?id=5o9JJJPPm6 | [
"The Viet Bui",
"Thanh Hong Nguyen",
"Tien Anh Mai"
] | Poster | Offline reinforcement learning (RL) has garnered significant attention for its ability to learn effective policies from pre-collected datasets without the need for further environmental interactions. While promising results have been demonstrated in single-agent settings, offline multi-agent reinforcement learning (MARL) presents additional challenges due to the large joint state-action space and the complexity of multi-agent behaviors. A key issue in offline RL is the distributional shift, which arises when the target policy being optimized deviates from the behavior policy that generated the data. This problem is exacerbated in MARL due to the interdependence between agents' local policies and the expansive joint state-action space. Prior approaches have primarily addressed this challenge by incorporating regularization in the space of either Q-functions or policies. In this work, we propose a novel type of regularizer in the space of stationary distributions to address the distributional shift more effectively. Our algorithm, ComaDICE, provides a principled framework for offline cooperative MARL to correct the stationary distribution of the global policy, which is then leveraged to derive local policies for individual agents. Through extensive experiments on the offline multi-agent MuJoCo and StarCraft II benchmarks, we demonstrate that ComaDICE achieves superior performance compared to state-of-the-art offline MARL methods across nearly all tasks. | Offline Reinforcement Learning, Multi-Agent Reinforcement Learning, Stationary Distribution Correction Estimation | This paper introduces ComaDICE, a novel offline cooperative multi-agent reinforcement learning algorithm that uses stationary distribution shift regularization to improve performance in complex environments like MuJoCo and StarCraft II. | 14,042 | 2410.01954 | [
-0.054578810930252075,
-0.029379965737462044,
0.06914859265089035,
0.009914066642522812,
0.016493501141667366,
0.07033123821020126,
-0.010535956360399723,
-0.0881618782877922,
-0.01692916825413704,
0.035737767815589905,
0.02241639420390129,
0.02750624343752861,
0.013756731525063515,
0.07074234634637833,
0.020684421062469482,
0.04274198040366173,
0.03851446881890297,
0.018833402544260025,
-0.12763218581676483,
-0.020183498039841652,
0.029283622279763222,
0.03700366988778114,
0.037256527692079544,
-0.016945794224739075,
0.060094621032476425,
0.0009743389091454446,
0.012077207677066326,
0.0037280831020325422,
0.03020448610186577,
-0.013125384226441383,
0.07265952974557877,
-0.0059614041820168495,
-0.002942007500678301,
-0.09899328649044037,
-0.04709591716527939,
0.06841201335191727,
-0.12208002805709839,
0.017299972474575043,
-0.005159823223948479,
0.03697893023490906,
0.03968709707260132,
0.05180622637271881,
-0.0082081975415349,
-0.015892749652266502,
0.024065151810646057,
0.007297842297703028,
-0.03367559239268303,
-0.024726230651140213,
0.07656815648078918,
-0.023104552179574966,
0.020303593948483467,
-0.018602248281240463,
-0.010257731191813946,
0.12236195057630539,
0.06405854225158691,
0.02268184721469879,
0.015178793109953403,
0.013771158643066883,
0.05688301473855972,
-0.09837224334478378,
-0.02837996557354927,
-0.1012548878788948,
-0.0013165822019800544,
-0.019150884822010994,
-0.08843787759542465,
-0.038672707974910736,
-0.05807020142674446,
0.09393338114023209,
0.0000037638037611031905,
-0.02986324019730091,
-0.10542229562997818,
0.04574711248278618,
0.008734382688999176,
0.00358933606185019,
0.056351371109485626,
-0.05222977325320244,
-0.034986335784196854,
-0.027796944603323936,
0.024915169924497604,
-0.05787328630685806,
0.012832493521273136,
-0.0276290625333786,
0.004774917848408222,
0.01658305712044239,
0.08754056692123413,
-0.0015593424905091524,
-0.07876630127429962,
0.046870820224285126,
0.13384521007537842,
-0.028333520516753197,
-0.01949562132358551,
0.060863010585308075,
0.03918261453509331,
-0.020151155069470406,
-0.07587947696447372,
0.0206927303224802,
-0.007029741071164608,
-0.10995322465896606,
0.010167155414819717,
0.14089784026145935,
0.04326477646827698,
0.029450418427586555,
0.013707920908927917,
0.02533169463276863,
0.03783430904150009,
-0.0753406286239624,
-0.030860496684908867,
0.07096612453460693,
0.06489403545856476,
-0.06671693176031113,
0.023294828832149506,
0.024342915043234825,
0.017579669132828712,
0.01691235601902008,
-0.08302626013755798,
0.04196628928184509,
0.04124696925282478,
0.005139744840562344,
-0.04444417357444763,
-0.03840531408786774,
0.008896145038306713,
-0.13485582172870636,
0.012046216987073421,
0.04113684222102165,
0.03347410261631012,
-0.010015607811510563,
0.02065904065966606,
1.5595683955122537e-33,
0.05598285794258118,
-0.01619088090956211,
-0.0027744779363274574,
-0.011304824613034725,
0.02278897911310196,
-0.009944183751940727,
0.0763901099562645,
-0.007933124899864197,
-0.03333047032356262,
-0.040102384984493256,
0.048582158982753754,
0.02658803202211857,
-0.010669243521988392,
0.076301209628582,
-0.020805470645427704,
-0.0025328483898192644,
-0.01928122527897358,
0.032704759389162064,
0.0643291026353836,
-0.0484604574739933,
0.028318703174591064,
-0.009186692535877228,
0.011933044530451298,
-0.025240203365683556,
0.03347482532262802,
-0.031015079468488693,
0.02219417318701744,
-0.00510646915063262,
0.039015788584947586,
0.013467160984873772,
0.037948139011859894,
0.0317719392478466,
-0.03884723782539368,
0.04013487696647644,
0.01563878543674946,
0.01875760406255722,
-0.10202095657587051,
-0.04984419047832489,
-0.060620471835136414,
0.01586868055164814,
0.0013498851330950856,
-0.013617481105029583,
0.05508485063910484,
-0.026952167972922325,
-0.05148555338382721,
-0.14014437794685364,
0.034079354256391525,
0.004167499486356974,
0.003353128209710121,
-0.07595695555210114,
0.0036952905356884003,
0.019993243739008904,
0.008927281945943832,
-0.1102459505200386,
-0.05098888650536537,
0.08367477357387543,
-0.0019247748423367739,
0.06789973378181458,
-0.033410150557756424,
0.010078280232846737,
-0.019431160762906075,
0.013092641718685627,
0.016635462641716003,
0.12775208055973053,
0.06714295595884323,
0.06925255805253983,
-0.019279902800917625,
0.03972191736102104,
0.05707576125860214,
-0.01953401044011116,
0.08016372472047806,
0.06651785224676132,
0.04073093459010124,
-0.06847868114709854,
0.11612503975629807,
-0.022942576557397842,
0.0603218749165535,
-0.05909694731235504,
0.018781714141368866,
-0.005898309405893087,
-0.01247442327439785,
0.005499827675521374,
-0.03290409594774246,
0.006905448157340288,
-0.017750369384884834,
-0.09884029626846313,
0.005487757734954357,
-0.06038945913314819,
-0.02500026859343052,
0.030820898711681366,
-0.08693782985210419,
-0.020050140097737312,
-0.06828755140304565,
0.03721778839826584,
0.03119811601936817,
-1.7588073411369388e-33,
-0.006590855773538351,
-0.02894744463264942,
0.01485980674624443,
-0.013632428832352161,
-0.016300059854984283,
0.009522447362542152,
-0.054243892431259155,
-0.03199740871787071,
0.00952333677560091,
-0.014684691093862057,
-0.09137047827243805,
0.0688869059085846,
0.026167605072259903,
-0.033716265112161636,
-0.028553776443004608,
-0.0007550175068899989,
0.029980596154928207,
0.01139744557440281,
0.01937396451830864,
0.03170326352119446,
0.026835530996322632,
-0.02891801856458187,
-0.052308689802885056,
0.12042272835969925,
0.029356345534324646,
0.023425159975886345,
-0.010568223893642426,
0.09014071524143219,
-0.003335577202960849,
0.09960021823644638,
-0.06939435750246048,
-0.09497921168804169,
-0.06274872273206711,
-0.04169677197933197,
-0.0809934213757515,
0.0921931266784668,
-0.07004411518573761,
0.09178656339645386,
-0.07681769132614136,
0.11571729928255081,
0.01805294305086136,
-0.08098629117012024,
-0.07964751124382019,
-0.031192859634757042,
0.0018113666446879506,
0.0015164531068876386,
-0.013484979048371315,
0.018724488094449043,
-0.05505092069506645,
0.04133934527635574,
-0.052815742790699005,
0.07807264477014542,
-0.08322376757860184,
0.01919649913907051,
-0.05238060653209686,
-0.0060117002576589584,
0.020233791321516037,
0.03306204453110695,
-0.00928801391273737,
-0.05510196089744568,
0.01196873839944601,
-0.03502175211906433,
-0.012100161984562874,
0.06795437633991241,
-0.007259685080498457,
0.0072010173462331295,
-0.005014568101614714,
0.030250202864408493,
0.04248658940196037,
-0.04775845259428024,
0.008557941764593124,
-0.029762187972664833,
0.0016576503403484821,
-0.0763961523771286,
0.04703044146299362,
-0.0034109128173440695,
-0.011340401135385036,
-0.03553764149546623,
-0.037325046956539154,
0.06107268109917641,
0.02424553409218788,
-0.004770561587065458,
0.041668545454740524,
0.08658620715141296,
0.0033204781357198954,
0.05687493458390236,
-0.006291520781815052,
0.13348795473575592,
0.039323270320892334,
-0.02289082109928131,
0.020559115335345268,
0.03429821506142616,
0.0328613743185997,
-0.021272096782922745,
-0.05731629952788353,
-4.559043631502391e-8,
-0.07892265170812607,
0.0047686719335615635,
0.10306094586849213,
-0.02223101258277893,
0.0826486125588417,
-0.003228231566026807,
0.0304370429366827,
-0.021789491176605225,
-0.05015537887811661,
-0.04624220356345177,
0.00034373393282294273,
0.07667936384677887,
0.031188447028398514,
-0.14206533133983612,
-0.060922250151634216,
0.04951270669698715,
0.037085793912410736,
-0.050702422857284546,
-0.03214966505765915,
0.03170301392674446,
0.03137675300240517,
-0.00900280848145485,
0.012350953184068203,
-0.0813642144203186,
0.011510215699672699,
-0.05063168704509735,
-0.0034703293349593878,
0.040910691022872925,
0.08249723166227341,
-0.0011970981722697616,
0.0382365882396698,
-0.06166389584541321,
0.03466891497373581,
-0.006626231595873833,
-0.02582777664065361,
0.05192321911454201,
0.06291370093822479,
-0.0821085199713707,
-0.009096266701817513,
0.003376634791493416,
-0.026961715891957283,
0.02674521692097187,
-0.029260525479912758,
-0.023713594302535057,
-0.005348678678274155,
-0.00445739459246397,
-0.0768079161643982,
-0.1230739876627922,
0.09176357835531235,
-0.06748758256435394,
0.07841324061155319,
-0.05887339264154434,
-0.014368738047778606,
0.051175326108932495,
0.08207400143146515,
-0.0063688973896205425,
0.007455684710294008,
-0.09088949859142303,
0.05421844497323036,
0.0569286048412323,
-0.032855525612831116,
0.0037656985223293304,
-0.10081000626087189,
0.012492779642343521
] | 0 | 0 | 0 | 0 |
|
Probabilistic Conformal Prediction with Approximate Conditional Validity | https://openreview.net/forum?id=Nfd7z9d6Bb | [
"Vincent Plassier",
"Alexander Fishkov",
"Mohsen Guizani",
"Maxim Panov",
"Eric Moulines"
] | Poster | We develop a new method for generating prediction sets that combines the flexibility of conformal methods with an estimate of the conditional distribution $\textup{P}_{Y \mid X}$. Existing methods, such as conformalized quantile regression and probabilistic conformal prediction, usually provide only a marginal coverage guarantee. In contrast, our approach extends these frameworks to achieve approximately conditional coverage, which is crucial for many practical applications. Our prediction sets adapt to the behavior of the predictive distribution, making them effective even under high heteroscedasticity. While exact conditional guarantees are infeasible without assumptions on the underlying data distribution, we derive non-asymptotic bounds that depend on the total variation distance of the conditional distribution and its estimate. Using extensive simulations, we show that our method consistently outperforms existing approaches in terms of conditional coverage, leading to more reliable statistical inference in a variety of applications. | Conformal Prediction, Conditional coverage, Probabilistic method, Uncertainty Quantification | We introduce a method that effectively integrates conformal approaches with an estimate of the conditional distribution to ensure the approximate conditional validity. | 14,036 | 2407.01794 | [
-0.1030174195766449,
-0.018623843789100647,
0.04819517582654953,
0.024979181587696075,
0.08148664981126785,
-0.06530765444040298,
-0.02068263106048107,
-0.004946210887283087,
0.020565779879689217,
-0.0008063911809585989,
-0.01447442639619112,
-0.04426368698477745,
0.015845810994505882,
0.0021268618293106556,
0.037330009043216705,
-0.03635578602552414,
0.07997286319732666,
-0.011236295104026794,
-0.09192900359630585,
-0.01307462528347969,
0.01264625322073698,
0.06325966119766235,
-0.02500280551612377,
0.013869193382561207,
0.03502638265490532,
-0.05175139755010605,
0.028740637004375458,
-0.018021713942289352,
-0.008223260752856731,
-0.013373138383030891,
-0.03078318014740944,
0.021428275853395462,
-0.026786230504512787,
-0.006520447321236134,
0.0011073179775848985,
0.013138731010258198,
-0.03186026215553284,
0.015260443091392517,
0.03496214747428894,
0.005578238051384687,
-0.007599956355988979,
0.031141657382249832,
0.05304815620183945,
0.11313255876302719,
0.03418346866965294,
0.01531996950507164,
0.02311312034726143,
0.026822248473763466,
-0.0010977122001349926,
0.03463960811495781,
-0.0009614360751584172,
-0.017750419676303864,
0.06969816982746124,
-0.028405673801898956,
-0.007868744432926178,
-0.07052437216043472,
0.06354863196611404,
-0.029686208814382553,
-0.0036266024690121412,
0.02245740406215191,
-0.03684012591838837,
-0.14378169178962708,
-0.010730093345046043,
-0.02033241093158722,
-0.04579980671405792,
-0.002233209554105997,
0.05345004051923752,
0.053608011454343796,
-0.02024146355688572,
0.03423375263810158,
-0.07853599637746811,
0.06399087607860565,
-0.1345100700855255,
0.038052141666412354,
0.05887353792786598,
0.09224572032690048,
0.11480829864740372,
-0.052428457885980606,
0.0046806056052446365,
-0.11062053591012955,
-0.04530911520123482,
0.037103548645973206,
0.0585535503923893,
-0.004853896331042051,
0.10714525729417801,
0.006077900063246489,
0.030310293659567833,
0.0564592070877552,
0.08397071808576584,
0.010953445918858051,
-0.03596113249659538,
0.024721138179302216,
-0.059712715446949005,
-0.009069850668311119,
-0.07322700321674347,
0.03900455683469772,
0.015719251707196236,
-0.11096002906560898,
0.06920633465051651,
0.1139751523733139,
-0.020352331921458244,
0.09239944815635681,
0.06962747126817703,
-0.06724322587251663,
0.03233621269464493,
-0.12141449004411697,
0.09172985702753067,
-0.0014808808919042349,
-0.0075939311645925045,
-0.060294341295957565,
0.02423088066279888,
0.03586587682366371,
0.05715705081820488,
0.0048401555977761745,
-0.02177717350423336,
-0.0722113624215126,
0.04191116616129875,
0.06326793879270554,
-0.05473842844367027,
0.01915907673537731,
-0.026511037722229958,
0.060749225318431854,
0.040757425129413605,
0.04937979206442833,
-0.018544549122452736,
-0.04056774079799652,
-0.09200718253850937,
3.117608281026454e-33,
-0.0053103165701031685,
-0.028293892741203308,
0.07198584079742432,
0.008550165221095085,
0.026155972853302956,
-0.02830379270017147,
-0.016417259350419044,
-0.046811506152153015,
-0.016815707087516785,
0.049151867628097534,
0.036332789808511734,
-0.01590954139828682,
-0.05302552506327629,
-0.0440223291516304,
0.08320793509483337,
0.028786472976207733,
-0.07528992742300034,
0.051429204642772675,
-0.07623524218797684,
0.014825970865786076,
0.016096780076622963,
-0.12501440942287445,
-0.036707330495119095,
0.028197571635246277,
-0.051703233271837234,
-0.07345060259103775,
-0.013725114054977894,
0.03669925779104233,
-0.0009048072970472276,
0.020050426945090294,
-0.10878023505210876,
0.0297038946300745,
0.07753005623817444,
-0.0006628656992688775,
0.0031590403523296118,
-0.008263264782726765,
-0.06511950492858887,
-0.058669038116931915,
-0.04087230563163757,
0.06111975759267807,
-0.08131231367588043,
-0.0004812362021766603,
0.05187515169382095,
-0.026852238923311234,
-0.003992403391748667,
-0.09790641814470291,
0.001172341755591333,
0.05094851180911064,
0.0247541144490242,
-0.10181158035993576,
0.02147505059838295,
0.014508895576000214,
-0.0880800411105156,
-0.07968221604824066,
-0.03156622126698494,
0.05221438780426979,
-0.043064989149570465,
-0.044613029807806015,
-0.03490309789776802,
-0.008112771436572075,
-0.037924110889434814,
0.014230752363801003,
-0.028242284432053566,
-0.019841602072119713,
-0.016248617321252823,
0.04784667119383812,
-0.027152225375175476,
-0.0631115660071373,
0.003792896866798401,
0.03823775053024292,
0.08258481323719025,
0.00782198365777731,
-0.041646577417850494,
-0.02805682271718979,
0.0023521671537309885,
0.02188711054623127,
0.0019311289070174098,
0.07463954389095306,
0.07828948646783829,
0.0028748956974595785,
0.0060808383859694,
0.004913566168397665,
-0.05082101374864578,
-0.011599570512771606,
0.005611006636172533,
-0.07354718446731567,
0.0994773581624031,
0.011809450574219227,
-0.12275109440088272,
-0.04490082338452339,
-0.06577587127685547,
-0.0572180412709713,
-0.057703565806150436,
-0.03210802748799324,
0.0005988276680000126,
-4.3833412923995374e-33,
-0.08374883979558945,
0.058453552424907684,
0.08243519067764282,
-0.01079039741307497,
-0.02422451041638851,
-0.06672714650630951,
-0.01804661937057972,
0.032545480877161026,
0.021852370351552963,
-0.07062792778015137,
-0.03767148777842522,
-0.003602791577577591,
0.0843328982591629,
-0.028681883588433266,
-0.10871631652116776,
0.048246800899505615,
0.03679332509636879,
0.04742375388741493,
-0.06939035654067993,
-0.02084801159799099,
0.08142988383769989,
-0.015642832964658737,
0.03957533463835716,
0.09668790549039841,
-0.03009285032749176,
-0.05642857030034065,
-0.061728429049253464,
0.14790959656238556,
-0.09417928010225296,
-0.00025599109358154237,
-0.047826506197452545,
-0.039757054299116135,
-0.017902912572026253,
-0.051980044692754745,
-0.028964858502149582,
-0.041872538626194,
0.010301748290657997,
0.07124732434749603,
-0.009591459296643734,
0.013804980553686619,
0.02125043235719204,
0.00186021882109344,
-0.05811570957303047,
-0.06511449813842773,
-0.046292077749967575,
0.0033154375851154327,
0.09314954280853271,
0.009144299663603306,
0.060910772532224655,
-0.0044072591699659824,
0.003272899892181158,
0.12216304242610931,
-0.023552928119897842,
0.07970904558897018,
-0.040025267750024796,
-0.028207553550601006,
-0.1041889488697052,
-0.0571109838783741,
0.06585024297237396,
0.02879558503627777,
-0.07842136919498444,
0.0036541398148983717,
0.05688253790140152,
0.08554959297180176,
-0.034919340163469315,
-0.04970289021730423,
-0.03446789085865021,
0.07372290641069412,
0.11253099888563156,
0.09887390583753586,
0.004581545013934374,
0.008568569086492062,
0.02767820656299591,
-0.027470577508211136,
-0.03399538993835449,
0.022698476910591125,
0.07455368340015411,
-0.057553019374608994,
0.00557613093405962,
0.035411279648542404,
0.05180398374795914,
0.05360540375113487,
0.05792085826396942,
-0.03819606825709343,
0.08433640748262405,
-0.051621440798044205,
0.08543159812688828,
-0.0033110573422163725,
0.09759818017482758,
0.033189237117767334,
-0.057780440896749496,
0.013162162154912949,
0.004506255034357309,
-0.004943584091961384,
-0.0602475069463253,
-4.5451628238879493e-8,
0.002118056872859597,
0.07970795035362244,
0.027780376374721527,
0.0021820676047354937,
0.04049909487366676,
0.02020309306681156,
0.007714568171650171,
-0.0632273256778717,
-0.00540619483217597,
0.027716392651200294,
0.04445608705282211,
0.02805786021053791,
0.005403976887464523,
-0.03588057681918144,
-0.030791187658905983,
0.016085561364889145,
0.016345292329788208,
0.0685223788022995,
-0.024284670129418373,
-0.03614848107099533,
-0.03644832596182823,
0.005171025171875954,
-0.009215972386300564,
-0.049855493009090424,
0.09059450775384903,
-0.04447821527719498,
-0.09913351386785507,
-0.03758798539638519,
0.004305844195187092,
0.0677829161286354,
0.029313810169696808,
-0.00962329376488924,
0.009349397383630276,
-0.026411257684230804,
-0.025463704019784927,
0.04161560907959938,
-0.03293868899345398,
-0.05770282819867134,
0.02235286310315132,
0.009702995419502258,
0.02620253711938858,
-0.016749171540141106,
-0.017934877425432205,
0.06302712857723236,
0.05641280859708786,
-0.026070399209856987,
0.03965744376182556,
0.0026935823261737823,
0.025151589885354042,
-0.043648041784763336,
0.04572651535272598,
-0.012159053236246109,
0.06062278896570206,
-0.0008692730334587395,
0.08950062096118927,
0.03743438050150871,
-0.007559121120721102,
0.021459057927131653,
-0.015820680186152458,
0.012796028517186642,
0.002820685738697648,
-0.030125651508569717,
0.030149761587381363,
0.004924633540213108
] | 0 | 0 | 0 | 0 |
|
Rethinking Neural Multi-Objective Combinatorial Optimization via Neat Weight Embedding | https://openreview.net/forum?id=GM7cmQfk2F | [
"Jinbiao Chen",
"Zhiguang Cao",
"Jiahai Wang",
"Yaoxin Wu",
"Hanzhang Qin",
"Zizhen Zhang",
"Yue-Jiao Gong"
] | Poster | Recent decomposition-based neural multi-objective combinatorial optimization (MOCO) methods struggle to achieve desirable performance. Even equipped with complex learning techniques, they often suffer from significant optimality gaps in weight-specific subproblems. To address this challenge, we propose a neat weight embedding method to learn weight-specific representations, which captures weight-instance interaction for the subproblems and was overlooked by most current methods. We demonstrate the potentials of our method in two instantiations. First, we introduce a succinct addition model to learn weight-specific node embeddings, which surpassed most existing neural methods. Second, we design an enhanced conditional attention model to simultaneously learn the weight embedding and node embeddings, which yielded new state-of-the-art performance. Experimental results on classic MOCO problems verified the superiority of our method. Remarkably, our method also exhibits favorable generalization performance across problem sizes, even outperforming the neural method specialized for boosting size generalization. | Neural Multi-Objective Combinatorial Optimization, Weight Embedding, Conditional Attention | We propose a neat weight embedding method for neural multi-objective combinatorial optimization | 14,028 | null | [
-0.06390910595655441,
-0.06703484058380127,
0.034322407096624374,
0.09051225334405899,
0.02522895485162735,
-0.01299201138317585,
0.009892882779240608,
-0.05013313889503479,
-0.029184142127633095,
-0.06836359202861786,
-0.046143896877765656,
-0.07356847822666168,
0.016750063747167587,
0.08233821392059326,
-0.06385166198015213,
0.07944023609161377,
0.1335039883852005,
0.09463604539632797,
-0.06654734164476395,
-0.02817116677761078,
0.052148256450891495,
-0.04694092273712158,
0.04516223818063736,
-0.026494499295949936,
0.06276889145374298,
0.004269698169082403,
-0.08589598536491394,
-0.02210353873670101,
0.08146192133426666,
0.04370098561048508,
0.0862204059958458,
-0.005665733944624662,
-0.0054609645158052444,
0.012119892984628677,
-0.051401395350694656,
0.03642536327242851,
-0.08012544363737106,
-0.006378329358994961,
-0.03447197377681732,
-0.02407362125813961,
0.016956398263573647,
0.0751027837395668,
-0.017717428505420685,
0.04286414757370949,
0.04484903812408447,
-0.021001413464546204,
-0.011867758817970753,
-0.04627435281872749,
-0.0028613286558538675,
0.0023123787250369787,
0.028668208047747612,
-0.06754672527313232,
-0.14042124152183533,
0.07199390977621078,
0.0691949874162674,
-0.016852078959345818,
0.008235941641032696,
-0.005959604401141405,
-0.027781430631875992,
0.01691460981965065,
-0.03806035593152046,
-0.06952957063913345,
0.04484524950385094,
-0.025196075439453125,
0.08331150561571121,
-0.05573650822043419,
-0.019116338342428207,
0.06488627940416336,
-0.05017338693141937,
0.0665641650557518,
0.0659109577536583,
0.03960726410150528,
0.00007675699453102425,
0.0376800112426281,
0.03487597405910492,
0.15516987442970276,
0.09065660089254379,
-0.06340478360652924,
-0.01627521589398384,
0.00995513703674078,
-0.06235026940703392,
0.01818058080971241,
0.01783273182809353,
0.010146764107048512,
0.18335294723510742,
0.037476878613233566,
0.013667329214513302,
-0.002636419842019677,
-0.028974123299121857,
0.0522017627954483,
-0.029269011691212654,
0.00198167422786355,
0.012818152084946632,
0.03804641216993332,
0.02776164375245571,
0.010140297934412956,
-0.017725566402077675,
-0.0016218326054513454,
-0.04695459455251694,
0.06789373606443405,
-0.027467112988233566,
0.060607973486185074,
0.1507183164358139,
-0.03340180590748787,
0.025725165382027626,
-0.04706179350614548,
0.011923671700060368,
0.030882764607667923,
0.008776557631790638,
-0.09285473823547363,
0.005970004480332136,
0.08161827176809311,
0.028791479766368866,
0.06481920927762985,
-0.06031137332320213,
-0.0001059672940755263,
0.044742900878190994,
-0.011492413468658924,
0.044113706797361374,
0.09579786658287048,
-0.04341091588139534,
0.003599717514589429,
-0.05851121246814728,
-0.0008263371419161558,
-0.0016884789802134037,
-0.015061705373227596,
-0.050847481936216354,
2.7625419471659277e-33,
-0.013805098831653595,
-0.056467216461896896,
0.03238430991768837,
0.05667494982481003,
0.057158734649419785,
-0.01873517781496048,
-0.0002208609803346917,
0.07853073626756668,
-0.054318636655807495,
-0.006573815830051899,
-0.018644975498318672,
0.009159760549664497,
-0.037482887506484985,
0.04868371784687042,
0.04911937937140465,
-0.07279084622859955,
-0.05506421998143196,
-0.008608736097812653,
-0.012286433018743992,
-0.07687833160161972,
-0.03460235148668289,
-0.04285811260342598,
0.01110811810940504,
-0.005151545163244009,
0.022848723456263542,
-0.026115940883755684,
0.08604300767183304,
-0.07380124926567078,
-0.03544770926237106,
0.02110547572374344,
-0.1306656301021576,
0.024451542645692825,
0.002966653322800994,
0.05775575712323189,
-0.04174429923295975,
-0.055698901414871216,
0.06398780643939972,
-0.03601226955652237,
0.0033142240718007088,
0.014297242276370525,
0.007371202576905489,
0.0006791832856833935,
0.029460757970809937,
-0.017504548653960228,
-0.06680583208799362,
-0.0744989812374115,
0.016388392075896263,
0.04304314777255058,
-0.010203520767390728,
-0.021221285685896873,
-0.02742641605436802,
-0.01745464652776718,
-0.026978595182299614,
-0.028884997591376305,
-0.02429068833589554,
-0.014884650707244873,
-0.01542768906801939,
0.016013268381357193,
0.07625435292720795,
-0.0032050556037575006,
-0.07443159818649292,
0.07964029908180237,
-0.01717534475028515,
0.04934273660182953,
-0.013888874091207981,
0.041153933852910995,
-0.0491703562438488,
-0.0022588460706174374,
0.053712550550699234,
0.04468771815299988,
0.017473502084612846,
0.08316515386104584,
0.005722654052078724,
-0.04288066551089287,
0.04667928069829941,
0.009879260323941708,
0.0648835077881813,
-0.13377243280410767,
-0.005313099827617407,
-0.02356041595339775,
0.040981490164995193,
0.08370883762836456,
0.021686851978302002,
-0.019732430577278137,
-0.024458203464746475,
-0.04920411854982376,
0.03210313618183136,
-0.026577631011605263,
-0.014636349864304066,
0.056444235146045685,
-0.02033674158155918,
-0.014594918116927147,
-0.08709884434938431,
-0.061892639845609665,
-0.024565652012825012,
-1.3983898447507827e-33,
-0.02461383491754532,
-0.006060255691409111,
-0.04377445951104164,
-0.017178134992718697,
0.04524807631969452,
-0.035142213106155396,
-0.08733166754245758,
-0.07170429825782776,
-0.0036344481632113457,
0.006986253894865513,
0.004021016415208578,
-0.04789365828037262,
0.004747522529214621,
-0.03148338943719864,
-0.020947737619280815,
-0.0007133047329261899,
-0.14437492191791534,
0.004020646680146456,
0.020039407536387444,
0.025415731593966484,
0.04091278091073036,
0.07182707637548447,
-0.0029743413906544447,
0.020278844982385635,
-0.04749403893947601,
0.030345825478434563,
0.02120523527264595,
0.018850337713956833,
0.01669645868241787,
0.014630372636020184,
-0.05337030440568924,
-0.0785895511507988,
-0.0684368759393692,
0.047825172543525696,
-0.03420696035027504,
0.053321417421102524,
0.05151618272066116,
0.06080959737300873,
0.0026896833442151546,
0.038617655634880066,
0.0534164123237133,
-0.06096683815121651,
-0.0736808031797409,
0.02397375926375389,
-0.02559249848127365,
-0.013418378308415413,
-0.012803019024431705,
0.011515401303768158,
0.07571125030517578,
0.006248460616916418,
-0.02557278797030449,
-0.02217271365225315,
-0.08046936243772507,
0.051605649292469025,
-0.055358294397592545,
-0.06503584235906601,
-0.014399834908545017,
-0.042031094431877136,
0.11312360316514969,
-0.08510611951351166,
-0.0010472933063283563,
-0.028731824830174446,
0.04332392290234566,
-0.0919494777917862,
-0.04225626215338707,
-0.046105463057756424,
0.0636206641793251,
-0.03007570095360279,
-0.037715766578912735,
0.006319352891296148,
-0.06770069897174835,
0.05317985266447067,
0.01669357717037201,
-0.025570861995220184,
-0.07794210314750671,
0.012146025896072388,
0.05362725630402565,
0.010250190272927284,
-0.00025208701845258474,
-0.055358000099658966,
0.02155338041484356,
0.019867539405822754,
0.045163411647081375,
0.000926233537029475,
0.125589519739151,
0.06966561824083328,
0.047523383051157,
0.1532381772994995,
0.05138338729739189,
0.03201870247721672,
-0.027622926980257034,
0.04716577008366585,
0.008875450119376183,
0.01128746010363102,
-0.0016435052966699004,
-4.3553061601642185e-8,
-0.09933122992515564,
0.022483067587018013,
0.013062254525721073,
0.044230856001377106,
0.05610935389995575,
-0.03416572883725166,
-0.023328926414251328,
0.12004237622022629,
-0.024087024852633476,
0.013126946985721588,
0.10725388675928116,
0.06255824118852615,
-0.01595490053296089,
-0.057600077241659164,
-0.04720676690340042,
0.009570715948939323,
-0.025565993040800095,
-0.057910285890102386,
0.032490916550159454,
-0.044467467814683914,
-0.019303899258375168,
0.013201018795371056,
0.06848263740539551,
0.07003717124462128,
0.006795474793761969,
-0.08565651625394821,
-0.047110188752412796,
0.029035214334726334,
0.04708154499530792,
0.01083306409418583,
-0.019510384649038315,
0.06480274349451065,
0.0025855216663330793,
-0.0045813899487257,
0.01284310594201088,
0.09540128707885742,
-0.014929492957890034,
0.032037779688835144,
-0.09077480435371399,
-0.04529942199587822,
0.03902755305171013,
0.04360748827457428,
-0.00401724549010396,
0.0038630252238363028,
0.07942813634872437,
-0.012746980413794518,
-0.0434616357088089,
-0.02025831863284111,
0.0261212270706892,
0.002239170717075467,
0.0736679881811142,
-0.06988336890935898,
-0.041357673704624176,
0.0683215782046318,
0.010812022723257542,
0.006350746378302574,
0.008234414272010326,
-0.1096254289150238,
0.04112394154071808,
0.09632808715105057,
0.01944892294704914,
-0.03220638632774353,
-0.05317729339003563,
-0.0018564192578196526
] | 0 | 0 | 0 | 0 |
|
Robust Root Cause Diagnosis using In-Distribution Interventions | https://openreview.net/forum?id=l11DZY5Nxu | [
"Lokesh Nagalapatti",
"Ashutosh Srivastava",
"Sunita Sarawagi",
"Amit Sharma"
] | Poster | Diagnosing the root cause of an anomaly in a complex interconnected system is
a pressing problem in today’s cloud services and industrial operations. We propose In-Distribution Interventions (IDI), a novel algorithm that predicts root cause
as nodes that meet two criteria: 1) Anomaly: root cause nodes should take on
anomalous values; 2) Fix: had the root cause nodes assumed usual values, the
target node would not have been anomalous. Prior methods of assessing the fix
condition rely on counterfactuals inferred from a Structural Causal Model (SCM)
trained on historical data. But since anomalies are rare and fall outside the training distribution, the fitted SCMs yield unreliable counterfactual estimates. IDI
overcomes this by relying on interventional estimates obtained by solely probing the fitted SCM at in-distribution inputs. We present a theoretical analysis
comparing and bounding the errors in assessing the fix condition using interventional and counterfactual estimates. We then conduct experiments by systematically varying the SCM’s complexity to demonstrate the cases where IDI’s interventional approach outperforms the counterfactual approach and vice versa.
Experiments on both synthetic and PetShop RCD benchmark datasets demonstrate that IDI consistently identifies true root causes more accurately and robustly than nine existing state-of-the-art RCD baselines. Code will be released
at https://github.com/nlokeshiisc/IDI_release. | Root Cause Diagnosis, Causal Inference, Interventional RCD | Identifying root cause of anomalies using interventions rather than counterfactuals estimated from a learned SCM | 14,022 | null | [
-0.029215624555945396,
-0.033945176750421524,
0.08289376646280289,
0.08738648891448975,
0.08656441420316696,
-0.012524103745818138,
-0.06615493446588516,
0.03202996030449867,
-0.028221361339092255,
0.020876528695225716,
-0.015025299973785877,
0.003543939208611846,
0.09435046464204788,
-0.012988364323973656,
-0.08892040699720383,
-0.0707843154668808,
0.05734704062342644,
-0.05267149955034256,
-0.036077894270420074,
-0.0525071956217289,
-0.0540301688015461,
0.0302719846367836,
-0.06800717115402222,
0.05181913077831268,
-0.04565570130944252,
0.0056815361604094505,
-0.02677985467016697,
-0.01772979646921158,
-0.04167378693819046,
-0.01743619330227375,
0.01059070136398077,
0.07544555515050888,
-0.09106457233428955,
-0.0037331993225961924,
0.034210558980703354,
0.013578802347183228,
0.03489513322710991,
-0.013527452945709229,
0.048593174666166306,
-0.03254501521587372,
0.08706334978342056,
-0.10842565447092056,
0.018738292157649994,
-0.04772019758820534,
0.04270760342478752,
-0.06954038888216019,
-0.008151146583259106,
-0.023049896582961082,
-0.05688115209341049,
0.036717917770147324,
-0.050758447498083115,
-0.06208343803882599,
0.05953935533761978,
0.06522444635629654,
0.03424748778343201,
-0.014146995730698109,
0.10837501287460327,
0.022119930014014244,
0.01349295862019062,
0.006119619123637676,
-0.0007579409284517169,
-0.09025231748819351,
-0.08330874145030975,
0.015089304186403751,
0.012651994824409485,
0.10595859587192535,
0.016111724078655243,
0.02663199044764042,
-0.047631192952394485,
0.028915954753756523,
0.035692550241947174,
0.012288103811442852,
-0.08855550736188889,
0.04342516139149666,
0.04261239990592003,
0.02468263730406761,
0.009089606814086437,
0.05379374325275421,
-0.002443054923787713,
-0.030159546062350273,
0.021347418427467346,
0.0890161469578743,
-0.021962903439998627,
0.01169871911406517,
0.07966077327728271,
0.013121451251208782,
0.033755432814359665,
0.06747458875179291,
0.07447443157434464,
0.08288854360580444,
0.052155502140522,
-0.019723976030945778,
0.04809235408902168,
0.027831923216581345,
0.04092935100197792,
0.05373935028910637,
0.01845516264438629,
-0.0007537505589425564,
0.06710166484117508,
0.04827430844306946,
0.05852678045630455,
0.0343913659453392,
-0.03366219624876976,
-0.04143794998526573,
0.0017056213691830635,
-0.04109877347946167,
-0.04273190349340439,
0.0417327843606472,
0.026269948109984398,
-0.026461008936166763,
0.01829160377383232,
-0.04649755731225014,
0.038749851286411285,
-0.034635573625564575,
-0.0171232707798481,
-0.0624845065176487,
0.05878675729036331,
-0.026154901832342148,
-0.05052671954035759,
-0.026076558977365494,
0.06563034653663635,
-0.00645477557554841,
0.060468703508377075,
0.010287344455718994,
-0.016466543078422546,
0.022449608892202377,
-0.1185225173830986,
2.357315799987876e-33,
0.02152455598115921,
0.011913487687706947,
0.02334379218518734,
-0.04290218651294708,
0.056015677750110626,
-0.06524576991796494,
-0.05900023877620697,
0.04788105934858322,
0.05719434842467308,
0.06097826734185219,
-0.022306112572550774,
0.08259434998035431,
-0.01963304542005062,
-0.10051761567592621,
0.07052607089281082,
0.036025192588567734,
-0.006343600805848837,
0.05256003141403198,
-0.049877043813467026,
-0.002429444808512926,
-0.001649100217036903,
-0.20631848275661469,
-0.04272088780999184,
-0.03452536091208458,
0.06368240714073181,
-0.0376414880156517,
0.03952905535697937,
0.03251943364739418,
0.06847270578145981,
0.007435960229486227,
0.0749286487698555,
-0.016538184136152267,
0.12230885028839111,
0.051666583865880966,
-0.03183400258421898,
-0.0674210786819458,
-0.07629628479480743,
-0.07685329020023346,
-0.08826432377099991,
-0.04465845599770546,
-0.03452334553003311,
0.004637612495571375,
0.0026770723052322865,
0.028679179027676582,
0.0021261468064039946,
-0.06444753706455231,
0.01256624050438404,
-0.13028010725975037,
-0.006130469962954521,
-0.09678738564252853,
0.010505713522434235,
0.0922548845410347,
0.05896841734647751,
-0.039921149611473083,
-0.03562815487384796,
0.08491889387369156,
0.005854510236531496,
-0.035347893834114075,
-0.050720006227493286,
0.02247459627687931,
-0.03547348082065582,
-0.07954487204551697,
0.012029445730149746,
0.04680715128779411,
0.001414767699316144,
0.00880490243434906,
-0.017082251608371735,
0.008764873258769512,
-0.06312468647956848,
0.03586664795875549,
-0.03738655894994736,
-0.00032415433088317513,
-0.05311106517910957,
0.031263068318367004,
0.012054326012730598,
-0.10988583415746689,
-0.009311102330684662,
0.029966315254569054,
-0.05262216925621033,
0.008989471010863781,
-0.05347267910838127,
-0.05352931097149849,
0.011471832171082497,
0.05380641669034958,
-0.05218082293868065,
0.0032595398370176554,
-0.03697161376476288,
0.07462941855192184,
-0.05389125272631645,
-0.027355890721082687,
0.006567737087607384,
0.01393324974924326,
-0.03218727186322212,
0.050935063511133194,
0.016641870141029358,
-3.4110190178313894e-33,
-0.09727459400892258,
0.0365603007376194,
0.016182152554392815,
-0.050893329083919525,
0.018152691423892975,
-0.016339793801307678,
0.0029704065527766943,
-0.00038061500526964664,
0.08188660442829132,
-0.04302811995148659,
-0.007871147245168686,
0.008955755271017551,
0.010098799131810665,
0.005938543472439051,
0.002239940455183387,
0.013189215213060379,
-0.030383778735995293,
-0.006128106266260147,
0.016334349289536476,
0.04272361472249031,
0.04676186293363571,
0.015814701095223427,
0.052171945571899414,
-0.039135824888944626,
-0.06093261018395424,
0.05716349929571152,
-0.05366458743810654,
0.09812945127487183,
-0.004615440499037504,
0.029370488598942757,
0.005573741160333157,
0.014647957868874073,
-0.02838725410401821,
0.06342315673828125,
-0.040353916585445404,
0.06404697149991989,
-0.005078727379441261,
-0.07105226069688797,
-0.03275180608034134,
0.03254958987236023,
0.05355529114603996,
0.05399058759212494,
-0.06471854448318481,
0.028928693383932114,
0.01535548735409975,
-0.010241648182272911,
0.01478827279061079,
0.05125047266483307,
0.0009524135966785252,
-0.04741537943482399,
-0.10223928093910217,
0.05302903801202774,
0.022887079045176506,
0.11048683524131775,
-0.024760644882917404,
-0.04051505774259567,
0.03269294276833534,
-0.060717783868312836,
-0.09347876161336899,
-0.025457706302404404,
-0.059097953140735626,
-0.0011729066027328372,
0.042134687304496765,
0.02824416570365429,
-0.023905837908387184,
-0.01617441326379776,
-0.004107964690774679,
-0.04069601744413376,
0.02278514765202999,
0.011926078237593174,
0.012872511520981789,
0.04995892569422722,
-0.10835485905408859,
-0.04302378371357918,
0.019730104133486748,
-0.049078069627285004,
-0.0866735577583313,
-0.05522771179676056,
-0.07555307447910309,
-0.015015590004622936,
0.019093265756964684,
0.03834884613752365,
0.013241467997431755,
0.042710453271865845,
-0.030533889308571815,
0.05772128328680992,
0.11873438209295273,
0.06204390898346901,
0.07778806239366531,
0.02334725484251976,
-0.12912507355213165,
0.05225702002644539,
-0.10020516067743301,
0.028415238484740257,
-0.03784013167023659,
-4.908600814701458e-8,
-0.09615703672170639,
0.042795393615961075,
0.017048722133040428,
0.03592650964856148,
0.10699258744716644,
-0.039259202778339386,
-0.05289016291499138,
0.0627836138010025,
-0.03807908669114113,
-0.035142067819833755,
-0.0396842360496521,
-0.06882264465093613,
-0.006287764757871628,
-0.05112600699067116,
-0.022369058802723885,
-0.026443835347890854,
-0.0029204655438661575,
0.0719105675816536,
-0.06529530137777328,
0.014524870552122593,
0.08635266870260239,
0.04035862535238266,
0.024776538833975792,
-0.008356858044862747,
0.03730697184801102,
-0.04993526637554169,
0.0029809949919581413,
0.04856470972299576,
0.03018774278461933,
0.015508361160755157,
-0.054688312113285065,
-0.04213569313287735,
0.03240890055894852,
-0.010486205108463764,
-0.0804799497127533,
0.07219536602497101,
0.06653402000665665,
-0.10033759474754333,
-0.006697335280478001,
-0.04032641649246216,
0.0401812382042408,
0.06777475029230118,
-0.016254745423793793,
0.06567592918872833,
-0.031663935631513596,
-0.02360420487821102,
-0.038996510207653046,
0.01109697762876749,
0.07205387204885483,
-0.05724087357521057,
0.025074385106563568,
0.005667433608323336,
-0.0002625680644996464,
0.09370646625757217,
0.07822343707084656,
-0.05416769161820412,
0.054190218448638916,
-0.053107116371393204,
0.008129412308335304,
0.07693447172641754,
0.03716855123639107,
-0.01007570419460535,
-0.037722278386354446,
-0.015670927241444588
] | 0 | 0 | 0 | 0 |
|
Boosting Neural Combinatorial Optimization for Large-Scale Vehicle Routing Problems | https://openreview.net/forum?id=TbTJJNjumY | [
"Fu Luo",
"Xi Lin",
"Yaoxin Wu",
"Zhenkun Wang",
"Tong Xialiang",
"Mingxuan Yuan",
"Qingfu Zhang"
] | Poster | Neural Combinatorial Optimization (NCO) methods have exhibited promising performance in solving Vehicle Routing Problems (VRPs). However, most NCO methods rely on the conventional self-attention mechanism that induces excessive computational complexity, thereby struggling to contend with large-scale VRPs and hindering their practical applicability. In this paper, we propose a lightweight cross-attention mechanism with linear complexity, by which a Transformer network is developed to learn efficient and favorable solutions for large-scale VRPs. We also propose a Self-Improved Training (SIT) algorithm that enables direct model training on large-scale VRP instances, bypassing extensive computational overhead for attaining labels. By iterating solution reconstruction, the Transformer network itself can generate improved partial solutions as pseudo-labels to guide the model training. Experimental results on the Travelling Salesman Problem (TSP) and the Capacitated Vehicle Routing Problem (CVRP) with up to 100K nodes indicate that our method consistently achieves superior performance for synthetic and real-world benchmarks, significantly boosting the scalability of NCO methods. | Neural Combinatorial Optimization, Large-Scale Vehicle Routing Problem | null | 14,013 | null | [
-0.060357242822647095,
-0.023602426052093506,
0.012713476084172726,
0.02295442298054695,
0.01278163306415081,
0.009097940288484097,
-0.01710052229464054,
0.05271463841199875,
-0.00864288117736578,
-0.03410368412733078,
-0.060103800147771835,
0.01701524294912815,
-0.016113899648189545,
0.05506105720996857,
-0.034536246210336685,
0.05377523973584175,
0.07923657447099686,
0.08360148966312408,
-0.022269032895565033,
-0.06285475194454193,
-0.019272107630968094,
0.008852103725075722,
0.003838792908936739,
0.0027833206113427877,
0.03705686703324318,
0.011485666036605835,
-0.01063354779034853,
-0.04048515483736992,
-0.010811920277774334,
0.04060877114534378,
0.02103056013584137,
-0.06295301020145416,
-0.08140328526496887,
0.04867524281144142,
-0.04184029996395111,
0.0376715250313282,
-0.14385315775871277,
-0.021902622655034065,
-0.004905147477984428,
-0.07120195031166077,
0.037456560879945755,
0.004071539733558893,
-0.042087092995643616,
0.010554441250860691,
0.08841404318809509,
-0.002238566754385829,
0.05156247317790985,
-0.010594830848276615,
0.0708826556801796,
-0.07932797074317932,
-0.034995418041944504,
-0.05097917094826698,
-0.05885791778564453,
0.013452314771711826,
-0.05144805833697319,
0.042834606021642685,
0.011059854179620743,
-0.009531861171126366,
-0.03338111564517021,
0.05252976343035698,
-0.03796079009771347,
-0.07653728872537613,
0.01597873494029045,
-0.043633028864860535,
-0.034755054861307144,
0.015346841886639595,
0.09484680742025375,
0.019222378730773926,
-0.06288666278123856,
0.022388769313693047,
-0.006672626826912165,
0.1170666366815567,
0.009354369714856148,
0.06745046377182007,
0.08793123066425323,
0.05305345728993416,
0.10646845400333405,
0.024846740067005157,
0.02373158186674118,
0.00581860588863492,
-0.026955438777804375,
0.02080351486802101,
-0.001451254589483142,
0.06747054308652878,
0.12159384787082672,
-0.01633104309439659,
-0.044695038348436356,
0.03440579026937485,
-0.03148790821433067,
0.0045519256964325905,
-0.08184880018234253,
-0.060905326157808304,
0.010571353137493134,
-0.03551071137189865,
-0.04567105323076248,
0.021383928135037422,
-0.04833607003092766,
0.013767232187092304,
-0.0009969071252271533,
0.05093616992235184,
-0.026173941791057587,
0.020105890929698944,
0.10989960283041,
-0.04420476406812668,
-0.03177705034613609,
0.038622867316007614,
0.05673639103770256,
0.04224153608083725,
0.0460045263171196,
-0.09637787938117981,
0.005958316847681999,
0.04355590045452118,
0.02157197892665863,
-0.036288004368543625,
-0.009147525765001774,
0.022629329934716225,
-0.06803715974092484,
-0.04084443300962448,
0.029010994359850883,
0.09802401065826416,
-0.07865339517593384,
-0.07409945130348206,
0.026342865079641342,
0.06339247524738312,
0.006697864271700382,
0.010469435714185238,
-0.010073625482618809,
2.3590344094629766e-33,
-0.07762303948402405,
0.11907825618982315,
0.05818876251578331,
-0.0025566951371729374,
0.05934476852416992,
0.006070779170840979,
0.016773013398051262,
0.04659537598490715,
-0.004649891052395105,
0.031102454289793968,
-0.0229008998721838,
0.011114971712231636,
0.014176539145410061,
0.0032137171365320683,
0.014174056239426136,
-0.041751421988010406,
-0.06641717255115509,
0.02546335756778717,
-0.06389148533344269,
-0.13008490204811096,
0.037258442491292953,
-0.05415097624063492,
0.0805821418762207,
-0.02808007039129734,
0.07713105529546738,
-0.02509562484920025,
-0.0030404215212911367,
-0.03304862231016159,
0.003964085131883621,
-0.010843183845281601,
-0.061430126428604126,
0.06880905479192734,
0.017836986109614372,
0.061401017010211945,
0.03396463021636009,
-0.05000544339418411,
0.0021264180541038513,
-0.06795664876699448,
0.005596504081040621,
0.07397352159023285,
-0.0013429314130917192,
0.056733712553977966,
-0.002248316304758191,
0.04613104835152626,
-0.09450718015432358,
-0.046168021857738495,
-0.0062899054028093815,
0.010083777830004692,
-0.05783306062221527,
0.02516186609864235,
-0.05402412265539169,
-0.01786719635128975,
-0.050493355840444565,
-0.037551529705524445,
-0.014894584193825722,
-0.0027231655549257994,
0.08970523625612259,
0.07826098799705505,
0.05649850144982338,
0.06047236919403076,
0.01581788994371891,
-0.013383075594902039,
-0.019902732223272324,
0.08922221511602402,
0.01100231148302555,
-0.008632348850369453,
-0.029257766902446747,
0.07774823158979416,
0.13370977342128754,
-0.019839925691485405,
0.006975595373660326,
0.026181915774941444,
-0.005489921197295189,
-0.033641912043094635,
0.14334514737129211,
-0.028302393853664398,
0.0420946441590786,
-0.04279264062643051,
-0.04333841800689697,
-0.04549838602542877,
-0.05228244140744209,
0.06883172690868378,
-0.010637911967933178,
-0.06296145915985107,
0.045559048652648926,
0.0004283021262381226,
0.019652968272566795,
-0.062161240726709366,
0.04012858122587204,
-0.022180553525686264,
-0.07027639448642731,
0.026574159041047096,
-0.061638858169317245,
-0.015283661894500256,
-0.009561916813254356,
-9.22015157114785e-34,
0.0011283489875495434,
0.03786277770996094,
-0.06997237354516983,
0.018787022680044174,
0.03629225119948387,
-0.05360805615782738,
-0.025724537670612335,
-0.013583035208284855,
0.01716926135122776,
-0.01694377511739731,
-0.08218959718942642,
-0.03524548187851906,
0.09732627123594284,
-0.018615009263157845,
0.04235710948705673,
-0.046312786638736725,
-0.005641985218971968,
-0.00404310179874301,
-0.01753154955804348,
0.004829069133847952,
-0.06015528738498688,
0.12149820476770401,
-0.07290514558553696,
-0.04207303375005722,
-0.04575372859835625,
-0.016238056123256683,
-0.060249265283346176,
0.05525096505880356,
0.046939667314291,
0.015799056738615036,
-0.015404440462589264,
-0.011887706816196442,
-0.002383030718192458,
-0.04581490904092789,
0.00393782090395689,
0.11180644482374191,
0.008025070652365685,
0.03672395646572113,
0.06101872771978378,
0.08027422428131104,
0.024476120248436928,
-0.12858982384204865,
-0.05145406723022461,
0.005271230824291706,
-0.0814615786075592,
0.0008656197460368276,
-0.04889988154172897,
-0.06583242118358612,
-0.01820426993072033,
0.08365800231695175,
-0.0009217344922944903,
-0.024243980646133423,
-0.08725536614656448,
0.03748341277241707,
-0.04133220389485359,
-0.020656852051615715,
-0.021934429183602333,
0.03471921384334564,
0.11290293186903,
-0.03840043395757675,
0.01628606766462326,
-0.007878417149186134,
0.07466094940900803,
-0.06119702011346817,
0.006116587668657303,
-0.10745025426149368,
0.009440179914236069,
-0.050978679209947586,
0.029362089931964874,
0.020703811198472977,
-0.06024304777383804,
0.019114168360829353,
0.05371256545186043,
0.027918843552470207,
-0.10146410018205643,
-0.025011269375681877,
0.03889385238289833,
0.02149185724556446,
0.022383669391274452,
-0.08782905340194702,
-0.002351683797314763,
-0.02785121090710163,
0.00995126273483038,
0.04147391393780708,
0.01911202073097229,
0.04995777830481529,
0.09324859827756882,
0.05965961515903473,
0.0278481375426054,
0.010290747508406639,
-0.015786083415150642,
0.02688969112932682,
0.017805000767111778,
0.053765956312417984,
-0.10092025995254517,
-4.3402305749395964e-8,
-0.07225140929222107,
0.02827959507703781,
-0.007345003075897694,
0.05023220181465149,
0.08548363298177719,
-0.004497733898460865,
0.05543274059891701,
0.13382108509540558,
-0.0992654338479042,
0.03618527576327324,
0.03319158777594566,
-0.006858249194920063,
0.00991829577833414,
-0.017141068354249,
-0.0466279461979866,
0.012800486758351326,
0.047169167548418045,
0.009661941789090633,
-0.006792252417653799,
-0.012354712933301926,
-0.10806868970394135,
0.053390730172395706,
0.04797881841659546,
0.0069954427890479565,
0.07806091010570526,
-0.033965103328228,
-0.04684120789170265,
0.0003086464130319655,
0.10370668023824692,
-0.031055232509970665,
-0.05185858905315399,
0.03353617340326309,
0.01824222505092621,
-0.004192451946437359,
0.030855640769004822,
0.10348988324403763,
0.007728703320026398,
0.026859648525714874,
-0.05043691396713257,
-0.06176337972283363,
0.06710554659366608,
0.032673854380846024,
-0.016522200778126717,
0.03971664980053902,
0.024264486506581306,
0.03458849713206291,
-0.006382764782756567,
-0.0574873611330986,
-0.010934887453913689,
-0.06423383206129074,
0.04446117952466011,
-0.04255092144012451,
-0.07311426103115082,
0.04862087219953537,
0.026395615190267563,
-0.04972340539097786,
0.04417189955711365,
-0.11664742231369019,
0.04215696081519127,
0.11524678021669388,
-0.07285745441913605,
0.03528615087270737,
-0.05818551778793335,
-0.011742757633328438
] | 0 | 0 | 0 | 0 |
|
Sensitivity Verification for Additive Decision Tree Ensembles | https://openreview.net/forum?id=h0vC0fm1q7 | [
"Arhaan Ahmad",
"Tanay Vineet Tayal",
"Ashutosh Gupta",
"S. Akshay"
] | Poster | Tree ensemble models, such as Gradient Boosted Decision Trees (GBDTs) and random forests, are widely popular models for a variety of machine learning tasks. The power of these models comes from the ensemble of decision trees, which makes analysis of such models significantly harder than for single trees. As a result, recent work has focused on developing exact and approximate techniques for questions such as robustness verification, fairness and explainability for such models of tree ensembles.
In this paper, we focus on a specific problem of feature sensitivity for additive decision tree ensembles and build a formal verification framework for a parametrized variant of it, where we also take into account the confidence of the tree ensemble in its output. We start by showing theoretical (NP-)hardness of the problem and explain how it relates to other verification problems. Next, we provide a novel encoding of the problem using pseudo-Boolean constraints. Based on this encoding, we develop a tunable algorithm to perform sensitivity analysis, which can trade off precision for running time. We implement our algorithm and study its performance on a suite of GBDT benchmarks from the literature. Our experiments show the practical utility of our approach and its improved performance compared to existing approaches. | Robustness verification, Sensitivity analysis, SAT solvers, efficient encodings, NP-hardness, fairness, confidence | We ask if an (additive) decision tree ensemble is sensitive to (potentially small) changes to a given feature or set of features. We show theoretical NP-hardness results, and provide a pseudo-Boolean encoding to solve the problem. | 14,006 | null | [
-0.060755278915166855,
-0.02234133705496788,
0.010120495222508907,
0.08445439487695694,
0.10668370872735977,
-0.02893611043691635,
0.02829894609749317,
-0.05992498621344566,
-0.07938750088214874,
0.04084811359643936,
-0.0976313129067421,
-0.12663474678993225,
0.037993136793375015,
0.03213372454047203,
-0.07056288421154022,
0.02954607456922531,
0.08080402761697769,
0.05232466012239456,
-0.017025252804160118,
-0.008443278260529041,
-0.010033845901489258,
0.025093164294958115,
-0.03105047717690468,
0.07758577913045883,
0.03567073866724968,
-0.113451287150383,
0.006758538074791431,
0.025661759078502655,
-0.04303845390677452,
0.03276391699910164,
-0.057058144360780716,
-0.05845487117767334,
0.0010007427772507071,
-0.05152884125709534,
-0.06433650106191635,
-0.008779231458902359,
0.0014031142927706242,
-0.027645820751786232,
-0.01338883489370346,
-0.013204600661993027,
0.0007167799049057066,
0.00899715255945921,
-0.023844553157687187,
-0.04249478504061699,
0.03754787519574165,
0.05780690908432007,
0.022094253450632095,
0.010843364521861076,
-0.05996357649564743,
-0.05605573207139969,
0.022854169830679893,
-0.029269438236951828,
-0.0862550437450409,
-0.05397515371441841,
-0.057272911071777344,
0.021334150806069374,
-0.009475947357714176,
-0.026258982717990875,
0.0020050269085913897,
-0.002720103831961751,
-0.01796826347708702,
-0.1199098527431488,
-0.007702982984483242,
0.005612480919808149,
0.09389837086200714,
0.030091572552919388,
-0.009136691689491272,
0.035984478890895844,
0.08418993651866913,
0.0173603817820549,
-0.017006555572152138,
0.09207909554243088,
-0.05420543998479843,
0.016990697011351585,
0.045515164732933044,
0.023114755749702454,
0.05570320785045624,
0.045035459101200104,
0.05988164246082306,
-0.00840232614427805,
-0.0886177346110344,
0.024087773635983467,
-0.035749562084674835,
0.008428605273365974,
0.1045163944363594,
-0.030931342393159866,
-0.02270338498055935,
0.10977304726839066,
-0.0018771992763504386,
0.07206538319587708,
0.016205448657274246,
-0.021240340545773506,
0.03785141184926033,
0.03866901248693466,
-0.07197374850511551,
0.037856392562389374,
-0.0017177161062136292,
-0.060439690947532654,
-0.007770598400384188,
0.05452091991901398,
-0.11054927855730057,
0.024784639477729797,
0.01900319568812847,
-0.11750028282403946,
0.0680001750588417,
-0.013310014270246029,
0.007893568836152554,
-0.06925813108682632,
0.011751585640013218,
-0.06001228094100952,
0.07646707445383072,
-0.00027008558390662074,
-0.0010945458197966218,
0.053202759474515915,
-0.04595478996634483,
0.0388600192964077,
-0.014801436103880405,
-0.01603628695011139,
0.02445572055876255,
0.019531605765223503,
0.008557518012821674,
-0.014162580482661724,
0.020473400130867958,
0.04226846992969513,
0.0747496485710144,
-0.02511734701693058,
-0.042208604514598846,
1.6830485669363655e-33,
0.05976015701889992,
0.038699518889188766,
0.03231404349207878,
-0.002961015561595559,
0.04664088785648346,
-0.03536207601428032,
-0.0420568510890007,
0.03203333169221878,
-0.04573652520775795,
0.06979479640722275,
-0.03612057864665985,
0.020183419808745384,
-0.040807876735925674,
0.002040150109678507,
0.08391241729259491,
0.05281674861907959,
-0.011982456780970097,
0.008997677825391293,
-0.05742311850190163,
-0.057511620223522186,
0.10247806459665298,
-0.07780468463897705,
-0.06493844836950302,
-0.010840261355042458,
-0.0007001860649324954,
-0.056817103177309036,
-0.010429702699184418,
0.01781369559466839,
0.02374058961868286,
0.0036332104355096817,
-0.10960806161165237,
-0.1064034029841423,
0.012544525787234306,
0.06238972768187523,
0.0212855227291584,
-0.015215273015201092,
-0.02213349938392639,
-0.027144040912389755,
-0.052611250430345535,
0.04433031007647514,
0.0005025190766900778,
-0.03411572426557541,
0.009391557425260544,
0.03296273201704025,
-0.022843163460493088,
0.03624894842505455,
-0.01945379003882408,
-0.03150080516934395,
0.0027505182661116123,
0.05490951985120773,
-0.003493041265755892,
0.042459964752197266,
0.011298009194433689,
-0.019104721024632454,
-0.0935620442032814,
0.034779857844114304,
0.027111336588859558,
0.0513593927025795,
0.035361647605895996,
0.11264403909444809,
-0.04301907494664192,
0.007168538402765989,
0.016585756093263626,
-0.04429792985320091,
-0.06938032805919647,
0.08333344012498856,
0.006900130771100521,
0.07717914879322052,
0.036027126014232635,
-0.052228894084692,
-0.03883129358291626,
-0.05152473226189613,
0.024190440773963928,
-0.0474071241915226,
0.046517886221408844,
0.032913658767938614,
0.12510354816913605,
0.0027276277542114258,
-0.00027624311042018235,
-0.06793850660324097,
-0.0556856133043766,
0.03817541524767876,
0.015002780593931675,
-0.001112853060476482,
-0.026202470064163208,
-0.08587634563446045,
-0.01558450423181057,
0.03527311235666275,
-0.013514726422727108,
-0.012679556384682655,
-0.12106659263372421,
0.028206717222929,
0.02930917590856552,
-0.0009422298753634095,
0.04033041372895241,
-1.5965435703173688e-33,
-0.03315316140651703,
-0.03665710613131523,
0.03985120728611946,
0.04137727618217468,
-0.02877400442957878,
-0.04715275764465332,
-0.12748579680919647,
-0.022431956604123116,
0.004402319435030222,
-0.06160585954785347,
-0.018727872520685196,
-0.004947233479470015,
0.08469432592391968,
0.017478186637163162,
-0.016721054911613464,
-0.027877330780029297,
-0.11506009101867676,
0.07391240447759628,
-0.013028237968683243,
0.0646621435880661,
0.05211503058671951,
0.11124107986688614,
-0.0666498914361,
0.0034045870415866375,
-0.066212497651577,
0.052013494074344635,
-0.07033885270357132,
0.027455177158117294,
0.01871548406779766,
0.034715332090854645,
-0.012745619751513004,
0.04945453628897667,
-0.10702871531248093,
-0.07952211797237396,
0.009043690748512745,
-0.08837693184614182,
0.006475054658949375,
-0.022714203223586082,
0.04144764319062233,
0.17215344309806824,
0.02300674095749855,
0.009806351736187935,
-0.10240352898836136,
0.03147688880562782,
0.022525077685713768,
0.00574076734483242,
0.02573457919061184,
-0.04811817780137062,
0.08223080635070801,
-0.038790952414274216,
0.02849057875573635,
0.0007245151791721582,
0.026391291990876198,
0.13743557035923004,
-0.0009447718621231616,
-0.03352465108036995,
-0.03368080034852028,
0.01589781977236271,
0.1552976369857788,
0.006010182201862335,
-0.06179433315992355,
-0.030681921169161797,
-0.04102429747581482,
0.01270389836281538,
-0.007106883451342583,
-0.037609510123729706,
0.029687801375985146,
-0.04611613601446152,
0.05684706196188927,
0.05768539011478424,
0.02652849815785885,
-0.0021803344134241343,
-0.0021562634501606226,
0.0367455817759037,
-0.0413484089076519,
-0.0346870981156826,
0.08712570369243622,
-0.041741080582141876,
0.004808281548321247,
0.051401279866695404,
0.03970229998230934,
0.029812578111886978,
0.10523855686187744,
0.03741716220974922,
0.021084781736135483,
0.03717169165611267,
0.05317578837275505,
0.056229833513498306,
0.03783603012561798,
-0.004280619323253632,
0.002639860613271594,
0.06962864845991135,
-0.015143993310630322,
0.0640951618552208,
-0.05666528269648552,
-4.691783317412046e-8,
-0.03779137507081032,
0.05344365909695625,
-0.0810716301202774,
0.007502230349928141,
0.043233729898929596,
-0.01300083938986063,
-0.08023443818092346,
0.009698213078081608,
-0.027419978752732277,
-0.03310485556721687,
0.10154221951961517,
-0.014395326375961304,
-0.037767812609672546,
-0.039341773837804794,
-0.00460560480132699,
0.04496338590979576,
-0.04957836493849754,
0.008669794537127018,
-0.043663863092660904,
0.0683775395154953,
0.049622781574726105,
0.01016190554946661,
0.044360969215631485,
0.03879505395889282,
0.02765583246946335,
-0.041712090373039246,
0.022426392883062363,
0.06452676653862,
0.07762810587882996,
0.0979669839143753,
-0.06960676610469818,
0.03315583989024162,
0.040431562811136246,
0.018834082409739494,
0.029181303456425667,
0.13834598660469055,
-0.009615004062652588,
0.013766215182840824,
-0.0731026828289032,
0.008544168435037136,
0.008572970516979694,
-0.04076545685529709,
-0.039025068283081055,
-0.02220373973250389,
0.0005136268446221948,
-0.047061510384082794,
0.03222109004855156,
0.020308857783675194,
-0.030325544998049736,
0.012491058558225632,
-0.010734450072050095,
-0.04694937542080879,
-0.006751806475222111,
0.0610598586499691,
-0.03394399583339691,
0.029375243932008743,
-0.09335620701313019,
-0.08549298346042633,
0.003397272899746895,
0.0018386748852208257,
0.07096634060144424,
-0.040229614824056625,
0.00012394716031849384,
-0.028333326801657677
] | 0 | 0 | 0 | 0 |
|
Monte Carlo Planning with Large Language Model for Text-Based Game Agents | https://openreview.net/forum?id=r1KcapkzCt | [
"Zijing Shi",
"Meng Fang",
"Ling Chen"
] | Poster | Text-based games provide valuable environments for language-based autonomous agents. However, planning-then-learning paradigms, such as those combining Monte Carlo Tree Search (MCTS) and reinforcement learning (RL), are notably time-consuming due to extensive iterations. Additionally, these algorithms perform uncertainty-driven exploration but lack language understanding and reasoning abilities.
In this paper, we introduce the Monte Carlo planning with Dynamic Memory-guided Large language model (MC-DML) algorithm. MC-DML leverages the language understanding and reasoning capabilities of Large Language Models (LLMs) alongside the exploratory advantages of tree search algorithms. Specifically, we enhance LLMs with in-trial and cross-trial memory mechanisms, enabling them to learn from past experiences and dynamically adjust action evaluations during planning. We conduct experiments on a series of text-based games from the Jericho benchmark. Our results demonstrate that the MC-DML algorithm significantly enhances performance across various games at the initial planning phase, outperforming strong contemporary methods that require multiple iterations. This demonstrates the effectiveness of our algorithm, paving the way for more efficient language-grounded planning in complex environments. | Large language model, Monte Carlo tree search, Text-based games | null | 14,005 | null | [
0.01615368388593197,
-0.007313721347600222,
0.05709249526262283,
0.0396546870470047,
-0.006849152967333794,
-0.008789172396063805,
-0.02234756574034691,
0.012984871864318848,
0.0402841717004776,
0.000045966546167619526,
-0.09347783029079437,
-0.014552033506333828,
0.049772582948207855,
0.042900200933218,
0.046120766550302505,
-0.01906878501176834,
0.08904269337654114,
0.08249165117740631,
-0.0038206533063203096,
-0.09185368567705154,
0.025065038353204727,
0.020404702052474022,
0.08726079761981964,
-0.039029259234666824,
0.015373834408819675,
0.01410854421555996,
0.008619139902293682,
-0.015663620084524155,
-0.009711110964417458,
0.0499599426984787,
0.06420442461967468,
-0.01633729413151741,
0.016417866572737694,
-0.0155254565179348,
-0.04642508178949356,
0.091245636343956,
-0.15517468750476837,
-0.05373430624604225,
-0.03175589442253113,
-0.055445823818445206,
-0.04014650732278824,
0.05157078057527542,
0.06718611717224121,
-0.0007730554789304733,
0.1173306331038475,
0.010388635098934174,
-0.0341404527425766,
0.00034671532921493053,
0.012602314352989197,
0.02720615826547146,
-0.0759323313832283,
-0.10165045410394669,
0.012477410025894642,
-0.06239299848675728,
-0.010133149102330208,
0.0894659236073494,
0.03350507467985153,
-0.007749485317617655,
0.02735736593604088,
-0.047815438359975815,
-0.007136327214539051,
-0.13184620440006256,
0.010867252014577389,
-0.01840074174106121,
-0.0911407545208931,
-0.0416727252304554,
0.04285794869065285,
0.0777767226099968,
0.012042870745062828,
0.004630976356565952,
0.028855936601758003,
0.05840396508574486,
0.030241746455430984,
0.04774811118841171,
-0.008762259036302567,
0.036645855754613876,
-0.057766035199165344,
-0.10979603230953217,
0.05093012750148773,
-0.005929374601691961,
-0.02304605394601822,
-0.017200369387865067,
-0.007349569816142321,
0.036118246614933014,
0.06924713402986526,
0.014411781914532185,
0.039564620703458786,
0.07820472866296768,
0.06819947063922882,
0.0259340051561594,
0.007518844213336706,
-0.062314458191394806,
0.002648664638400078,
0.012088914401829243,
-0.0482926219701767,
0.06342846155166626,
-0.048360247164964676,
-0.11086270958185196,
-0.027224870398640633,
0.04583435878157616,
0.04415123909711838,
0.09786026179790497,
0.021187124773859978,
-0.03049575351178646,
-0.03544538468122482,
-0.009935997426509857,
0.0440252423286438,
0.0034268852323293686,
-0.023503819480538368,
-0.06076585501432419,
0.004961901344358921,
-0.03121516853570938,
0.06084743142127991,
0.061996348202228546,
-0.023941688239574432,
0.024695776402950287,
0.06672680377960205,
-0.013810412958264351,
-0.02644723653793335,
0.10913610458374023,
0.03783874586224556,
-0.025196155533194542,
0.01936248131096363,
0.027991142123937607,
0.003357460955157876,
-0.020447982475161552,
0.042375072836875916,
-1.1771090025157373e-33,
0.0812116339802742,
0.010614389553666115,
0.08499789237976074,
0.10375939309597015,
0.08711110800504684,
-0.015515446662902832,
0.09256453812122345,
0.0076545607298612595,
-0.03908379375934601,
-0.004032476339489222,
0.007579846773296595,
0.00012983029591850936,
-0.07678186893463135,
0.012625022791326046,
0.05939166620373726,
-0.02250347100198269,
-0.07200375944375992,
0.037122633308172226,
-0.02651250548660755,
-0.01958656497299671,
0.05262066796422005,
0.019047239795327187,
0.03957611694931984,
-0.10386989265680313,
0.05738065764307976,
0.08294812589883804,
-0.011741410940885544,
-0.030773039907217026,
0.032635997980833054,
-0.02790968120098114,
-0.1016295924782753,
-0.012910827063024044,
-0.08323559910058975,
0.06102677807211876,
0.05683550238609314,
0.006974017713218927,
-0.05644889920949936,
-0.09290359169244766,
-0.04429912939667702,
0.020194150507450104,
-0.03116009570658207,
-0.01876918599009514,
0.05790161341428757,
0.008180917240679264,
-0.07294095307588577,
-0.05785943940281868,
-0.036293335258960724,
-0.04838324338197708,
-0.024763664230704308,
0.0009356523514725268,
0.03990812599658966,
0.04017452523112297,
-0.017386427149176598,
-0.018749769777059555,
0.025586768984794617,
0.017653366550803185,
0.010325650684535503,
0.022698724642395973,
-0.023223984986543655,
0.05888315662741661,
0.04438186064362526,
-0.021233731880784035,
0.015739714726805687,
0.11612874269485474,
0.030670491978526115,
0.11248398572206497,
-0.017889009788632393,
0.09628627449274063,
0.06394637376070023,
-0.06057346239686012,
0.02763265371322632,
-0.00983873475342989,
0.05974024534225464,
-0.04515474662184715,
0.016588499769568443,
0.0010141503298655152,
0.040461231023073196,
-0.07522562891244888,
-0.013599024154245853,
-0.021995404735207558,
0.0019199901726096869,
-0.06321202218532562,
-0.048790156841278076,
0.02306313067674637,
0.04315357655286789,
-0.03340267017483711,
0.03233267739415169,
-0.07608900964260101,
0.03907201811671257,
-0.035740140825510025,
-0.07766713947057724,
-0.029267137870192528,
-0.02726602926850319,
0.0011855581542477012,
0.008231861516833305,
3.097434789064865e-36,
-0.017691494897007942,
-0.13177622854709625,
-0.00018818836542777717,
0.026220137253403664,
-0.005111404228955507,
-0.07881078869104385,
-0.06722614169120789,
-0.07485078275203705,
0.08512377738952637,
-0.08690736442804337,
-0.14753368496894836,
0.06385000050067902,
0.04065697267651558,
-0.0072192903608083725,
0.013295737095177174,
-0.06540773063898087,
0.04007257521152496,
-0.00363946077413857,
0.011332273483276367,
0.08275400102138519,
-0.004003074485808611,
0.05303102359175682,
-0.06775003671646118,
-0.020003413781523705,
-0.002070758957415819,
-0.0018120151944458485,
0.023677997291088104,
-0.0007723476155661047,
0.014290370978415012,
0.040165238082408905,
-0.007745176553726196,
0.004432212561368942,
-0.026721978560090065,
0.022345498204231262,
-0.08322090655565262,
0.031684763729572296,
0.0845436304807663,
0.004219272639602423,
-0.07650608569383621,
0.10771914571523666,
0.0658128559589386,
-0.07031500339508057,
-0.0915825292468071,
-0.0398755818605423,
-0.047311000525951385,
0.1113910973072052,
-0.05554654449224472,
0.03657326474785805,
0.008370543830096722,
0.006258401088416576,
-0.005446293391287327,
0.007515974808484316,
-0.08264317363500595,
-0.03872419893741608,
-0.04993061348795891,
-0.08084828406572342,
-0.08111724257469177,
0.024035414680838585,
0.07440894842147827,
0.006434041541069746,
-0.07150827348232269,
0.03423815220594406,
-0.008736967109143734,
-0.0016672804486006498,
-0.04509645327925682,
-0.009970840997993946,
-0.07299169152975082,
-0.028704071417450905,
0.004685858264565468,
-0.047790590673685074,
-0.014397499151527882,
0.007794620934873819,
-0.024632293730974197,
0.005805513355880976,
0.027305684983730316,
0.037704650312662125,
-0.02537628822028637,
-0.006797716487199068,
-0.011927604675292969,
-0.039795272052288055,
0.05264031141996384,
0.05997003614902496,
0.0037030777893960476,
0.0708857849240303,
0.05922473594546318,
0.05595102533698082,
-0.04282071068882942,
0.11488863825798035,
-0.01838039979338646,
-0.007722699549049139,
0.0022711181081831455,
0.012469245120882988,
0.06482476741075516,
0.0423174612224102,
-0.12919175624847412,
-3.518981017691658e-8,
-0.0667487382888794,
0.08363880217075348,
0.010006885975599289,
0.017852913588285446,
0.04794948548078537,
-0.019249841570854187,
-0.03260017931461334,
0.036511894315481186,
0.012863028794527054,
-0.014763180166482925,
0.07102885842323303,
0.018307490274310112,
0.014698879793286324,
-0.07392850518226624,
-0.051804330199956894,
0.047454942017793655,
0.015415209345519543,
-0.07699023187160492,
-0.004959221463650465,
0.03391363099217415,
0.04761829227209091,
0.02629353664815426,
-0.026073120534420013,
-0.01777944713830948,
-0.02438283897936344,
-0.026914823800325394,
-0.02960965596139431,
0.07571094483137131,
0.09167352318763733,
-0.004319123458117247,
0.050886139273643494,
-0.03061101958155632,
0.0032826976384967566,
0.06898865848779678,
0.05985866114497185,
0.02305757626891136,
0.04107886180281639,
-0.05138009414076805,
0.04133158549666405,
-0.06710649281740189,
0.005136060994118452,
0.03752314671874046,
-0.020457660779356956,
0.00038232762017287314,
-0.024420369416475296,
0.04073844477534294,
-0.1127162054181099,
-0.08176247775554657,
0.05778355151414871,
-0.08758900314569473,
-0.02405434474349022,
0.006293910089880228,
-0.09017516672611237,
0.001961946254596114,
0.09149445593357086,
0.06019040197134018,
0.01852208562195301,
-0.01284125167876482,
0.02975441701710224,
0.03535151854157448,
-0.014034206978976727,
-0.05006447434425354,
-0.05036197975277901,
-0.0016443402273580432
] | 0 | 0 | 0 | 0 |
|
Actions Speak Louder Than Words: Rate-Reward Trade-off in Markov Decision Processes | https://openreview.net/forum?id=Za3M6OZuCU | [
"Haotian Wu",
"Gongpu Chen",
"Deniz Gunduz"
] | Poster | The impact of communication on decision-making systems has been extensively studied under the assumption of dedicated communication channels. We instead consider communicating through actions, where the message is embedded into the actions of an agent which interacts with the environment in a Markov decision process (MDP) framework. We conceptualize the MDP environment as a finite-state channel (FSC), where the actions of the agent serve as the channel input, while the states of the MDP observed by another agent (i.e., receiver) serve as the channel output. Here, we treat the environment as a communication channel over which the agent communicates through its actions, while at the same time, trying to maximize its reward. We first characterize the optimal information theoretic trade-off between the average reward and the rate of reliable communication in the infinite-horizon regime. Then, we propose a novel framework to design a joint control/coding policy, termed Act2Comm, which seamlessly embeds messages into actions. From a communication perspective, Act2Comm functions as a learning-based channel coding scheme for non-differentiable FSCs under input-output constraints. From a control standpoint, Act2Comm learns an MDP policy that incorporates communication capabilities, though at the cost of some control performance. Overall, Act2Comm effectively balances the dual objectives of control and communication in this environment. Experimental results validate Act2Comm's capability to enable reliable communication while maintaining a certain level of control performance. | Markov Decision Process, Channel coding, Rate-Reward Trade-off, Finite state channel | null | 13,994 | 2502.03335 | [
0.0009496163111180067,
-0.0895436629652977,
-0.05622238293290138,
-0.0355072021484375,
0.018721846863627434,
0.011141943745315075,
0.05927722528576851,
-0.02444884553551674,
0.03985751420259476,
0.09860855340957642,
-0.05839930847287178,
0.0010247582104057074,
0.05386575311422348,
0.06375686824321747,
0.015144675970077515,
0.0032177302055060863,
0.09339959174394608,
-0.008942755870521069,
-0.03616034612059593,
0.021834485232830048,
0.07279371470212936,
0.03031950443983078,
0.04246049374341965,
-0.007255189120769501,
-0.07218462228775024,
-0.0012281331000849605,
0.01672634296119213,
-0.0026033855974674225,
-0.06664030253887177,
-0.015583417378365993,
0.05351680517196655,
-0.03026728890836239,
0.07776623219251633,
-0.08414749056100845,
-0.03285069391131401,
0.09719640761613846,
-0.09059058874845505,
-0.05052363499999046,
-0.10894930362701416,
-0.04560261592268944,
-0.001368449884466827,
0.00022515590535476804,
-0.038797102868556976,
0.0383170060813427,
-0.03931894898414612,
-0.017591994255781174,
0.06986600160598755,
-0.012387542054057121,
-0.016125351190567017,
-0.0007725441828370094,
-0.031674791127443314,
0.04143302142620087,
0.011432794854044914,
0.03457803279161453,
0.026575524359941483,
0.03966166079044342,
-0.02908845618367195,
0.022423816844820976,
-0.11213760077953339,
-0.06733674556016922,
-0.023170966655015945,
-0.04463043436408043,
-0.02346871607005596,
-0.015962230041623116,
-0.0632026419043541,
0.01087989378720522,
-0.017072713002562523,
0.0775805190205574,
-0.03992585837841034,
0.016816407442092896,
-0.05592671409249306,
-0.022449728101491928,
-0.0041447971016168594,
-0.04731369391083717,
0.05821243301033974,
-0.05925567075610161,
-0.026208557188510895,
-0.018719784915447235,
-0.0008698115707375109,
0.0035928317811340094,
0.0011163150193169713,
0.01642792485654354,
-0.0026241918094456196,
-0.02400088869035244,
0.028235098347067833,
-0.038784533739089966,
-0.028645846992731094,
0.04752045497298241,
0.1422087401151657,
-0.011178920976817608,
-0.0792231634259224,
0.008870170451700687,
0.024300990626215935,
0.03934650495648384,
-0.01453564502298832,
-0.0220137070864439,
-0.021947702392935753,
-0.03561297059059143,
-0.03553101420402527,
0.08024076372385025,
-0.028159668669104576,
-0.004687929060310125,
-0.026208842173218727,
-0.06323890388011932,
0.048240337520837784,
-0.050710417330265045,
-0.06964973360300064,
0.049359019845724106,
0.04157352074980736,
-0.05236008018255234,
0.020574314519762993,
0.020478518679738045,
0.037409961223602295,
0.05237075313925743,
0.03973245248198509,
0.06879788637161255,
0.059233229607343674,
-0.03998205065727234,
0.021973488852381706,
0.034718774259090424,
-0.018265213817358017,
-0.10426541417837143,
0.04269326478242874,
0.030193427577614784,
-0.00895354151725769,
-0.022329937666654587,
-0.015198164619505405,
-2.7830236505320598e-33,
-0.014624680392444134,
-0.031015541404485703,
0.0085288155823946,
0.008957047015428543,
0.08012547343969345,
0.0289266649633646,
0.0545518733561039,
-0.038825396448373795,
0.0009184361551888287,
-0.031936775892972946,
0.0033468473702669144,
0.009001986123621464,
-0.007891707122325897,
0.09591049700975418,
0.10238204896450043,
-0.03552475944161415,
-0.08810954540967941,
0.04354926198720932,
0.08487487584352493,
-0.10204149782657623,
0.04805571213364601,
0.020954186096787453,
-0.050181180238723755,
-0.05182907357811928,
0.10659801214933395,
-0.02531224489212036,
0.0011488551972433925,
-0.008044072426855564,
0.0590544231235981,
-0.011983385309576988,
-0.04607928916811943,
0.00946787465363741,
0.016228247433900833,
-0.006437632255256176,
0.011474560014903545,
0.015465852804481983,
-0.10014808177947998,
-0.024490509182214737,
-0.0126648535951972,
-0.05589607357978821,
-0.01149135921150446,
-0.022078603506088257,
0.02791299670934677,
0.03349470719695091,
-0.06789269298315048,
-0.11220593750476837,
0.024416295811533928,
-0.009647964499890804,
-0.07680893689393997,
-0.05595114827156067,
0.09090924263000488,
-0.0021070996299386024,
0.058644216507673264,
-0.043533939868211746,
-0.025453925132751465,
-0.025791268795728683,
-0.036856915801763535,
0.03301255777478218,
-0.028218887746334076,
0.07941040396690369,
-0.019187496975064278,
-0.04869650676846504,
0.06528220325708389,
0.016445057466626167,
0.02627202682197094,
0.08748764544725418,
-0.05158389359712601,
0.07573751360177994,
0.10054157674312592,
-0.1008315160870552,
-0.00045993688399903476,
0.09128999710083008,
0.08114892989397049,
-0.07385610044002533,
0.06469990313053131,
0.020411260426044464,
0.08896150439977646,
-0.0795324519276619,
-0.009263165295124054,
0.09019263088703156,
-0.010673673823475838,
-0.02649816684424877,
-0.1225326731801033,
0.05762244388461113,
0.02734208293259144,
-0.01318725012242794,
-0.03554050251841545,
-0.09544112533330917,
-0.030598880723118782,
0.0523914210498333,
-0.10598794370889664,
-0.01357241626828909,
0.044849347323179245,
0.014921436086297035,
0.07630939781665802,
2.3269750969317027e-34,
0.005485422443598509,
-0.023483457043766975,
-0.03344465792179108,
0.0023360021878033876,
-0.016858596354722977,
-0.033483460545539856,
-0.030236246064305305,
-0.04358578845858574,
-0.04014156758785248,
-0.03651196509599686,
-0.07821625471115112,
0.008888727985322475,
0.030120011419057846,
0.02025902271270752,
-0.006961326580494642,
-0.08181430399417877,
0.017650043591856956,
0.025889942422509193,
0.030877748504281044,
0.021549373865127563,
0.07600037753582001,
0.10488080233335495,
-0.09852297604084015,
0.06179635599255562,
-0.0039664669893682,
0.042039670050144196,
0.03141357749700546,
0.09186238795518875,
0.027782119810581207,
0.06206635385751724,
-0.06079377979040146,
-0.039714790880680084,
-0.07050656527280807,
-0.0036011035554111004,
-0.015243359841406345,
0.019924208521842957,
0.054440852254629135,
0.008298642933368683,
-0.12045355886220932,
0.044866081327199936,
0.08884631097316742,
-0.08106725662946701,
-0.021334482356905937,
0.05274346098303795,
0.02647964283823967,
-0.049335379153490067,
-0.026045318692922592,
-0.019819185137748718,
-0.04258869215846062,
-0.02231035754084587,
-0.008959607221186161,
-0.0011386480182409286,
-0.06273211538791656,
0.011694594286382198,
-0.08147149533033371,
-0.016289034858345985,
-0.01813141442835331,
0.09188460558652878,
0.07152940332889557,
0.037246666848659515,
0.002116457326337695,
-0.005591851659119129,
-0.020962456241250038,
0.002287303563207388,
0.03235376998782158,
0.02504490315914154,
0.006955181714147329,
-0.011325621977448463,
0.073799729347229,
0.027180301025509834,
0.09636256843805313,
0.016473010182380676,
-0.02383497916162014,
0.042617812752723694,
0.15546302497386932,
0.022769639268517494,
-0.04175753891468048,
-0.012908168137073517,
-0.06280338019132614,
0.050499677658081055,
-0.0056025246158242226,
0.05399170145392418,
0.04163571819663048,
0.07087091356515884,
-0.0006487623322755098,
0.0006646770634688437,
-0.01263307686895132,
0.009121949784457684,
-0.0089395921677351,
-0.11073547601699829,
0.023956557735800743,
0.016053179278969765,
0.00917583703994751,
0.031744226813316345,
-0.10842611640691757,
-4.429504585345967e-8,
-0.005264568608254194,
0.016224339604377747,
0.0990205630660057,
-0.040636513382196426,
0.018205376341938972,
0.012367747724056244,
-0.016330759972333908,
-0.07540157437324524,
0.03954413905739784,
-0.04446832090616226,
0.08158696442842484,
0.06437855213880539,
0.030710678547620773,
-0.08955079317092896,
-0.021836763247847557,
0.0537748783826828,
-0.026288514956831932,
-0.11143410205841064,
-0.0030848912429064512,
0.0040140096098184586,
0.02591150812804699,
-0.012350478209555149,
-0.02209421619772911,
0.06776183098554611,
-0.04972146451473236,
-0.04475288465619087,
0.03873644024133682,
0.10138069838285446,
0.06200224906206131,
0.022107653319835663,
0.06029294803738594,
-0.0008085235604085028,
0.06512768566608429,
0.09170901775360107,
-0.018962642177939415,
-0.08523508906364441,
0.0005298355245031416,
-0.07209254056215286,
0.006696355063468218,
-0.006912516430020332,
-0.008465510793030262,
0.04102419689297676,
0.028427522629499435,
0.0029826832469552755,
-0.020162194967269897,
0.07453101873397827,
-0.08737976104021072,
-0.06298613548278809,
0.0612330324947834,
-0.051363252103328705,
-0.001840866869315505,
0.059157099574804306,
-0.009977810084819794,
0.05488232150673866,
0.04061178117990494,
-0.021528448909521103,
0.010406074114143848,
-0.06645181030035019,
0.016873696818947792,
0.056561775505542755,
-0.019821031019091606,
0.05825171247124672,
-0.10076597332954407,
-0.01564931310713291
] | 0 | 0 | 0 | 0 |
|
A Statistical Framework for Ranking LLM-based Chatbots | https://openreview.net/forum?id=rAoEub6Nw2 | [
"Siavash Ameli",
"Siyuan Zhuang",
"Ion Stoica",
"Michael W. Mahoney"
] | Poster | Large language models (LLMs) have transformed natural language processing, with frameworks like Chatbot Arena providing pioneering platforms for evaluating these models. By facilitating millions of pairwise comparisons based on human judgments, Chatbot Arena has become a cornerstone in LLM evaluation, offering rich datasets for ranking models in open-ended conversational tasks. Building upon this foundation, we propose a statistical framework that incorporates key advancements to address specific challenges in pairwise comparison analysis. First, we introduce a factored tie model that enhances the ability to handle ties—an integral aspect of human-judged comparisons—significantly improving the model's fit to observed data. Second, we extend the framework to model covariance between competitors, enabling deeper insights into performance relationships and facilitating intuitive groupings into performance tiers. Third, we resolve optimization challenges arising from parameter non-uniqueness by introducing novel constraints, ensuring stable and interpretable parameter estimation. Through rigorous evaluation and extensive experimentation, our framework demonstrates substantial improvements over existing methods in modeling pairwise comparison data. To support reproducibility and practical adoption, we release leaderbot, an open-source Python package implementing our models and analyses. | Large Language Models (LLMs), Paired Comparison, Statistical Ranking, Human Preferences, Chatbot Arena, Logistic Regression | We introduce a rigorous statistical framework for ranking large language models (LLMs) using crowdsourced comparisons, improving accuracy for ties, wins, and losses beyond current methods like Elo. | 13,986 | 2412.18407 | [
-0.00652981037274003,
-0.09933985769748688,
0.018007267266511917,
0.016728738322854042,
0.012886695563793182,
-0.035615697503089905,
-0.04775211215019226,
0.006177576258778572,
0.016250688582658768,
-0.0791257694363594,
-0.04139481112360954,
-0.0516752228140831,
0.039048776030540466,
0.037494346499443054,
0.05535706877708435,
0.010023069567978382,
0.12558095157146454,
-0.006116761360317469,
-0.07591750472784042,
-0.05530163273215294,
-0.0005933829816058278,
0.053657084703445435,
0.0868704542517662,
-0.024487480521202087,
0.03412456810474396,
-0.08231417834758759,
-0.03424989804625511,
0.012810682877898216,
-0.017921168357133865,
-0.010459387674927711,
-0.01689225807785988,
0.018444614484906197,
-0.005645080469548702,
0.09018214046955109,
-0.008951080031692982,
0.04534709081053734,
-0.07398560643196106,
-0.04828372225165367,
-0.005199408158659935,
-0.03806741163134575,
-0.07156948745250702,
-0.028122596442699432,
0.008653055876493454,
-0.01914605125784874,
0.04289352521300316,
0.02789941616356373,
-0.0835740715265274,
0.007163607515394688,
-0.007718206848949194,
0.07015913724899292,
-0.12277735769748688,
0.036496538668870926,
0.04598332569003105,
0.09795531630516052,
-0.05451526120305061,
0.09579025954008102,
-0.01604405790567398,
0.021799862384796143,
0.024304015561938286,
-0.08200615644454956,
-0.04446481168270111,
-0.12472319602966309,
-0.026222795248031616,
-0.00348073267377913,
-0.026226412504911423,
-0.04163169860839844,
-0.01937170699238777,
0.026352960616350174,
-0.06400227546691895,
0.05795089155435562,
0.008232548832893372,
0.02987772598862648,
0.026056738570332527,
0.08492963761091232,
-0.017789876088500023,
0.04909421503543854,
0.016382934525609016,
-0.051386602222919464,
0.04668044298887253,
-0.011186148039996624,
-0.049181390553712845,
-0.03588311746716499,
0.02524256519973278,
0.05837085098028183,
0.0876476988196373,
-0.0019463594071567059,
0.06470947712659836,
0.006426190957427025,
-0.04599867761135101,
0.022923458367586136,
-0.05832921713590622,
-0.041276153177022934,
0.11207693815231323,
0.06534025073051453,
0.02486344613134861,
0.03208573907613754,
0.016026481986045837,
-0.01274377666413784,
0.009568310342729092,
0.07824673503637314,
0.015621841885149479,
0.01054575014859438,
-0.02498469315469265,
-0.1033131554722786,
0.06901313364505768,
0.023912372067570686,
0.04240340366959572,
0.016456151381134987,
0.010172334499657154,
-0.08487109839916229,
-0.017000852152705193,
-0.03313842788338661,
0.07819465547800064,
0.04120398312807083,
0.09773075580596924,
-0.03235713765025139,
0.057126350700855255,
0.031928062438964844,
0.05316397175192833,
0.007420721463859081,
0.08586833626031876,
0.0071274335496127605,
0.05018800497055054,
-0.038375839591026306,
0.014767440035939217,
0.06148332729935646,
0.03030477836728096,
1.0644251038554551e-33,
0.10113869607448578,
-0.0339905209839344,
0.05255577340722084,
0.06313319504261017,
0.03957902267575264,
0.014403335750102997,
-0.04727384075522423,
0.024390554055571556,
-0.11512025445699692,
-0.027464143931865692,
-0.09891045838594437,
0.051900576800107956,
-0.030412375926971436,
0.027362613007426262,
0.01897990144789219,
-0.028788568452000618,
-0.05186650902032852,
0.04908808320760727,
0.0023601048160344362,
-0.013765895739197731,
0.06701331585645676,
0.039730679243803024,
0.010576304979622364,
0.08340424299240112,
0.05190001428127289,
-0.0033158448059111834,
0.03359478339552879,
-0.11392346024513245,
-0.018371889367699623,
-0.0025143229868263006,
-0.1120673269033432,
-0.013156741857528687,
0.043184131383895874,
0.06196508929133415,
0.029656430706381798,
-0.04070382937788963,
-0.042618073523044586,
-0.03952549770474434,
-0.007970230653882027,
0.018177000805735588,
-0.07072175294160843,
-0.015879884362220764,
-0.0012456451077014208,
-0.07065178453922272,
-0.1010010614991188,
-0.0830322876572609,
-0.08660051226615906,
0.05762242153286934,
0.008682179264724255,
-0.01710912212729454,
0.007785634603351355,
0.04844998940825462,
0.04948190972208977,
0.04945717006921768,
-0.026684075593948364,
-0.11233672499656677,
0.004080663435161114,
0.03922499716281891,
-0.03265265375375748,
-0.055048517882823944,
-0.01735672540962696,
0.03217979148030281,
-0.006402085535228252,
-0.016067886725068092,
0.052421003580093384,
-0.030988795682787895,
-0.007829885929822922,
0.044220998883247375,
0.019895775243639946,
0.0016910971608012915,
0.022388944402337074,
0.058412328362464905,
-0.021113187074661255,
0.005273493006825447,
0.023296140134334564,
0.04394395276904106,
0.03527393564581871,
-0.023030422627925873,
0.10518301278352737,
-0.003256760537624359,
-0.0016734484815970063,
-0.03939871862530708,
-0.01361917331814766,
-0.13438920676708221,
-0.05013813078403473,
-0.0790148675441742,
0.025539923459291458,
-0.06000781059265137,
-0.018418388441205025,
0.037444163113832474,
-0.05732323229312897,
0.028602125123143196,
-0.03994660824537277,
0.08079830557107925,
0.010199056006968021,
-6.741518417505879e-34,
-0.05412958562374115,
0.021784139797091484,
0.06819016486406326,
0.0669705867767334,
0.03703116998076439,
-0.05045641213655472,
0.005782807245850563,
-0.004345918074250221,
0.001866747159510851,
-0.04571964591741562,
-0.04044374078512192,
-0.044478241354227066,
0.09950832277536392,
-0.03334256261587143,
0.1090511828660965,
0.0038585125003010035,
-0.019605645909905434,
0.006256473250687122,
-0.016580862924456596,
-0.002588320057839155,
0.014027919620275497,
0.10823605209589005,
-0.07382295280694962,
0.022899333387613297,
-0.03169560059905052,
-0.002664353000000119,
0.009786753915250301,
-0.05903133749961853,
-0.01458826009184122,
-0.016419073566794395,
-0.0026842409279197454,
0.03706281632184982,
-0.09433046728372574,
0.003811619011685252,
-0.020185552537441254,
0.05587880685925484,
0.05554221197962761,
-0.09204373508691788,
-0.027760455384850502,
0.04103364050388336,
0.044109147042036057,
-0.010039026848971844,
-0.08222554624080658,
-0.05331021547317505,
0.0075536444783210754,
0.0175777617841959,
-0.11046848446130753,
-0.05285051092505455,
-0.04736117646098137,
-0.04417232424020767,
-0.010091218166053295,
0.03791411593556404,
-0.008914854377508163,
-0.022618383169174194,
-0.06367149204015732,
-0.11888732761144638,
-0.025370724499225616,
-0.06631318479776382,
0.03495018184185028,
0.004919631872326136,
-0.00661783991381526,
-0.024088654667139053,
0.06255819648504257,
0.004142759367823601,
-0.01937923952937126,
-0.03152032941579819,
0.01195605006068945,
0.0379750020802021,
-0.03648620843887329,
-0.005861622281372547,
0.06811341643333435,
-0.04510799050331116,
-0.007002750877290964,
0.0018304013647139072,
-0.010178045369684696,
0.032986830919981,
0.015056942589581013,
-0.01370844617486,
-0.005285997875034809,
-0.03415624797344208,
0.03898059204220772,
0.023263700306415558,
0.0940409004688263,
0.06294134259223938,
0.00628863088786602,
0.1252709925174713,
0.06646785140037537,
0.12955906987190247,
0.018749594688415527,
0.06980834901332855,
-0.002465182216838002,
-0.09624464809894562,
-0.02861989289522171,
0.0476863868534565,
-0.02934171073138714,
-3.702167106212073e-8,
-0.06753445416688919,
0.02975093014538288,
-0.057922959327697754,
0.11197948455810547,
-0.03336843475699425,
0.02579905092716217,
-0.018826600164175034,
0.05075186491012573,
-0.002039511688053608,
0.05320379137992859,
0.05829653888940811,
0.02155212312936783,
-0.011663427576422691,
-0.022369422018527985,
0.007541326805949211,
0.09277937561273575,
-0.09079908579587936,
0.046334464102983475,
0.015941431745886803,
-0.04474858567118645,
0.05462697893381119,
0.05502690002322197,
0.022020122036337852,
0.0764569565653801,
-0.016199322417378426,
-0.049553874880075455,
-0.07134708762168884,
0.10745212435722351,
-0.005395422223955393,
-0.0011704904027283192,
-0.006103431340306997,
-0.010306219570338726,
-0.036239322274923325,
-0.07516627758741379,
0.0494864247739315,
0.024783512577414513,
-0.07359009981155396,
-0.03271329030394554,
-0.005131513345986605,
0.018929356709122658,
0.008411595597863197,
0.06314563751220703,
-0.059540387243032455,
0.053243737667798996,
0.08943503350019455,
-0.011453862302005291,
-0.041430383920669556,
-0.12207422405481339,
-0.013768834993243217,
-0.06350196897983551,
0.016420798376202583,
-0.07443640381097794,
0.0009510907693766057,
0.009780324064195156,
0.0695093646645546,
0.013742124661803246,
-0.012922501191496849,
0.05202306807041168,
0.10678470879793167,
-0.02034449391067028,
0.03345021232962608,
0.01472017727792263,
-0.02241106890141964,
0.026176542043685913
] | https://github.com/suquark/leaderbot | 3 | 0 | 0 | 0 |
ONLINE EPSILON NET & PIERCING SET FOR GEOMETRIC CONCEPTS | https://openreview.net/forum?id=nNiWRRj6r9 | [
"Sujoy Bhore",
"Devdan Dey",
"Satyam Singh"
] | Poster | VC-dimension (Vapnik & Chervonenkis (1971)) and $\varepsilon$-nets (Haussler & Welzl (1987)) are key concepts in Statistical Learning Theory. Intuitively, VC-dimension is a measure of the size of a class of sets. The famous $\varepsilon$-net theorem, a fundamental result in Discrete Geometry, asserts that if the VC-dimension of a set system is bounded, then a small sample exists that intersects all sufficiently large sets.
In online learning scenarios where data arrives sequentially, the VC-dimension helps to bound the complexity of the set system, and $\varepsilon$-nets ensure the selection of a small representative set. This sampling framework is crucial in various domains, including spatial data analysis, motion planning in dynamic environments, optimization of sensor networks, and feature extraction in computer vision, among others. Motivated by these applications, we study the online $\varepsilon$-net problem for geometric concepts with bounded VC-dimension. While the offline version of this problem has been extensively studied, surprisingly, there are no known theoretical results for the online version to date. We present the first deterministic online algorithm with an optimal competitive ratio for intervals in $\mathbb{R}$. Next, we give a randomized online algorithm with a near-optimal competitive ratio for axis-aligned boxes in $\mathbb{R}^d$, for $d\le 3$. Furthermore, we introduce a novel technique to analyze similar-sized objects of constant description complexity in $\mathbb{R}^d$, which may be of independent interest.
Next, we focus on the continuous version of this problem (called online piercing set), where ranges of the set system are geometric concepts in $\mathbb{R}^d$ arriving in an online manner, but the universe is the entire ambient space, and the objective is to choose a small sample that intersects all the ranges. Although online piercing set is a very well-studied problem in the literature, to our surprise, very few works have addressed generic geometric concepts without any assumption about the sizes. We advance this field by proposing asymptotically optimal competitive deterministic algorithms for boxes and ellipsoids in $\mathbb{R}^d$, for any $d\in\mathbb{N}$. | Theoretical machine learning, VC-dimension, Geometric sampling | null | 13,983 | 2410.07059 | [
0.01166477520018816,
-0.0758422240614891,
0.03966336324810982,
-0.0081912437453866,
0.08764712512493134,
0.02597629837691784,
0.06462221592664719,
0.01774171181023121,
0.0785980373620987,
-0.021075429394841194,
-0.05021291598677635,
-0.005227779503911734,
0.03512934595346451,
0.027918968349695206,
-0.046693213284015656,
-0.009196929633617401,
0.07244221866130829,
0.04371422156691551,
-0.06842834502458572,
0.00305911130271852,
0.004369631875306368,
0.00012612619320861995,
0.042526889592409134,
-0.00960495974868536,
0.037822529673576355,
-0.04930116981267929,
0.058342721313238144,
0.02498292736709118,
0.008648043498396873,
0.016815287992358208,
-0.034595344215631485,
0.01676148921251297,
-0.006773282773792744,
-0.07369129359722137,
0.01898142695426941,
0.0371556356549263,
-0.0723997950553894,
0.0746435895562172,
-0.04922814667224884,
0.04292552173137665,
0.016370560973882675,
0.07485764473676682,
0.05564060062170029,
0.0634109228849411,
0.03723103553056717,
0.07409745454788208,
-0.024220019578933716,
-0.0721917599439621,
-0.0007542778621427715,
-0.09328371286392212,
-0.08364783227443695,
-0.021465066820383072,
-0.07306920737028122,
-0.001999531639739871,
-0.01522944774478674,
-0.011470827274024487,
-0.013849475421011448,
0.0006648304988630116,
0.0334017314016819,
0.0034842563327401876,
0.07288123667240143,
-0.12775973975658417,
0.023014431819319725,
-0.03518342226743698,
-0.042651936411857605,
0.030586494132876396,
0.03914885222911835,
0.0455959253013134,
0.0034433361142873764,
0.05021169036626816,
-0.0768192708492279,
0.06709219515323639,
-0.11360234022140503,
0.056803539395332336,
0.0411851704120636,
-0.030053358525037766,
0.03834255412220955,
0.002797664375975728,
0.0010690432973206043,
-0.011278620921075344,
-0.008833005093038082,
0.022910453379154205,
-0.009816641919314861,
0.010146032087504864,
0.015580734238028526,
0.0515749417245388,
-0.017078667879104614,
0.06476183235645294,
0.10966812819242477,
-0.029949311167001724,
-0.0905701071023941,
0.006416928488761187,
0.020340172573924065,
-0.02756728045642376,
0.004336598329246044,
-0.0737546905875206,
-0.016840776428580284,
-0.08525168895721436,
0.09709802269935608,
0.0713796466588974,
-0.03394627943634987,
-0.002735921647399664,
0.05366018787026405,
0.03361718729138374,
0.047692108899354935,
0.050271496176719666,
0.03477652743458748,
0.004700531251728535,
0.04295549541711807,
-0.12822213768959045,
0.061681050807237625,
-0.06011636555194855,
0.024610599502921104,
0.03581566363573074,
0.002833039965480566,
-0.008851852267980576,
0.013362781144678593,
-0.0022424247581511736,
0.010929695330560207,
0.04092199355363846,
-0.02576873078942299,
-0.04781040921807289,
0.002114588161930442,
-0.017090892419219017,
0.04041062667965889,
-0.014614581130445004,
0.00772354518994689,
1.044014389327853e-33,
0.018684392794966698,
0.022685179486870766,
0.0801854357123375,
-0.019050654023885727,
0.08701376616954803,
0.05092897638678551,
0.0833897590637207,
0.012332111597061157,
0.04406959190964699,
0.0926617681980133,
-0.06344807147979736,
-0.03862753137946129,
-0.009247818030416965,
0.09535788744688034,
0.14079055190086365,
-0.011244216002523899,
0.03359205275774002,
0.031512301415205,
-0.03129785880446434,
-0.06183562055230141,
0.06661508232355118,
-0.05889385566115379,
-0.003919277805835009,
-0.04226858913898468,
-0.004246499855071306,
-0.06469471007585526,
-0.04003124311566353,
0.06025012582540512,
-0.01655847765505314,
-0.026530789211392403,
-0.05433196574449539,
0.027771316468715668,
0.03072582185268402,
0.07244144380092621,
0.056909654289484024,
-0.00028535386081784964,
-0.06333043426275253,
-0.013580728322267532,
0.011793229728937149,
-0.07548032701015472,
0.04432191327214241,
0.03712020441889763,
0.025606222450733185,
0.014378241263329983,
-0.08204623311758041,
-0.07869177311658859,
-0.038472093641757965,
0.09213103353977203,
-0.03862797096371651,
-0.036466680467128754,
0.004057674203068018,
0.03491203486919403,
-0.056966111063957214,
-0.06615625321865082,
-0.012050897814333439,
0.018253818154335022,
-0.04481218382716179,
0.013683470897376537,
-0.0111624114215374,
0.07469100505113602,
-0.047604236751794815,
-0.07214651256799698,
-0.023795736953616142,
0.022110873833298683,
0.040494654327631,
0.042293526232242584,
0.003239138051867485,
-0.02542886883020401,
0.11695964634418488,
-0.060135941952466965,
0.00264056958258152,
0.03378242999315262,
-0.07263520359992981,
-0.12019721418619156,
0.05169365555047989,
0.0025841668248176575,
-0.003550052409991622,
-0.025201736018061638,
-0.013526396825909615,
0.09221412241458893,
-0.0288032665848732,
-0.026127079501748085,
0.00499736750498414,
-0.14487798511981964,
-0.06570513546466827,
-0.06735314428806305,
0.08050774782896042,
-0.03276974707841873,
-0.050470128655433655,
-0.029183614999055862,
-0.07983842492103577,
-0.022376304492354393,
-0.031040342524647713,
0.018506383523344994,
-0.040951844304800034,
-9.405910902646465e-34,
-0.03429160639643669,
-0.018515685573220253,
0.02507171779870987,
0.02599802426993847,
-0.043826159089803696,
-0.007731081917881966,
-0.03417510539293289,
0.01619601994752884,
-0.07085592299699783,
-0.1115657240152359,
-0.058581288903951645,
-0.010382868349552155,
-0.0049974448047578335,
-0.05612260475754738,
-0.0052648684941232204,
0.0773235559463501,
-0.01396979670971632,
-0.03839666768908501,
0.010044705122709274,
-0.018508827313780785,
0.001092861988581717,
-0.008787163533270359,
-0.05938219651579857,
0.017209092155098915,
-0.016670431941747665,
0.017056360840797424,
0.0024359351955354214,
0.024894123896956444,
-0.14444762468338013,
0.05483439192175865,
-0.041287560015916824,
-0.13272725045681,
-0.14061343669891357,
0.04483537748456001,
-0.047796912491321564,
-0.03507750481367111,
0.0960741639137268,
0.008513307198882103,
-0.03721727058291435,
0.07301109284162521,
-0.022641729563474655,
-0.08054081350564957,
-0.025432512164115906,
0.009973163716495037,
-0.015704091638326645,
-0.01257951557636261,
0.00764033617451787,
0.06076136231422424,
-0.021296488121151924,
0.0064510321244597435,
-0.030372828245162964,
0.06909578293561935,
-0.019632253795862198,
0.061822764575481415,
0.042399559170007706,
0.02401682734489441,
-0.030047938227653503,
0.019447937607765198,
0.11689476668834686,
-0.01978990249335766,
-0.02127791941165924,
-0.08191279321908951,
-0.0690779983997345,
0.13958773016929626,
0.042674481868743896,
0.03366616740822792,
-0.04624299705028534,
-0.02661687508225441,
-0.024334046989679337,
0.032471902668476105,
0.013991624116897583,
-0.03473581001162529,
-0.01993163861334324,
-0.03897714987397194,
-0.14450494945049286,
-0.03840584680438042,
0.08955178409814835,
0.02626418136060238,
-0.0213974229991436,
0.022762250155210495,
0.017931364476680756,
0.025881005451083183,
0.05748653784394264,
0.04157942906022072,
0.05867593735456467,
0.06922957301139832,
0.05545908212661743,
0.013759002089500427,
0.04253741353750229,
0.0366656519472599,
-0.013581358827650547,
-0.0026514215860515833,
0.004732300993055105,
0.018588580191135406,
-0.035983338952064514,
-4.146429688489661e-8,
-0.08333317190408707,
0.048056554049253464,
-0.0049744113348424435,
-0.04281727224588394,
0.09719369560480118,
-0.013927350752055645,
0.042049746960401535,
0.09275337308645248,
-0.08182605355978012,
0.029220089316368103,
0.027637582272291183,
0.01617208682000637,
-0.04586060345172882,
0.03370237722992897,
0.021706998348236084,
-0.013133328408002853,
-0.029446635395288467,
0.006620385218411684,
0.020160872489213943,
0.052951667457818985,
-0.01670696772634983,
-0.0020752439741045237,
0.030973542481660843,
0.012166326865553856,
0.04873006418347359,
-0.05591797083616257,
-0.03930639103055,
0.0637463927268982,
0.06180788204073906,
0.02361181564629078,
0.028625626116991043,
-0.0034832952078431845,
0.14011229574680328,
-0.019401632249355316,
0.09116671979427338,
0.01881592534482479,
-0.09323932975530624,
0.007969548925757408,
-0.05040784180164337,
-0.07703344523906708,
-0.010575396940112114,
0.049707259982824326,
0.010614436119794846,
-0.004486418329179287,
0.10383792966604233,
0.03773018345236778,
0.02424010820686817,
-0.047882627695798874,
0.030316820368170738,
-0.0050394777208566666,
-0.02323034405708313,
0.019072463735938072,
-0.030111581087112427,
0.06976515054702759,
0.04335019364953041,
0.042557407170534134,
0.020118189975619316,
-0.05964263528585434,
0.01158615667372942,
0.04048244282603264,
-0.047941092401742935,
0.02616989240050316,
-0.08284200727939606,
-0.03188896179199219
] | 0 | 0 | 0 | 0 |
|
SimulPL: Aligning Human Preferences in Simultaneous Machine Translation | https://openreview.net/forum?id=XBF63bHDZw | [
"Donglei Yu",
"Yang Zhao",
"Jie Zhu",
"Yangyifan Xu",
"Yu Zhou",
"Chengqing Zong"
] | Poster | Simultaneous Machine Translation (SiMT) generates translations while receiving streaming source inputs. This requires the SiMT model to learn a read/write policy, deciding when to translate and when to wait for more source input. Numerous linguistic studies indicate that audiences in SiMT scenarios have distinct preferences, such as accurate translations, simpler syntax, and no unnecessary latency. Aligning SiMT models with these human preferences is crucial to improve their performances. However, this issue still remains unexplored. Additionally, preference optimization for SiMT task is also challenging. Existing methods focus solely on optimizing the generated responses, ignoring human preferences related to latency and the optimization of read/write policy during the preference optimization phase. To address these challenges, we propose Simultaneous Preference Learning (SimulPL), a preference learning framework tailored for the SiMT task. In the SimulPL framework, we categorize SiMT human preferences into five aspects: **translation quality preference**, **monotonicity preference**, **key point preference**, **simplicity preference**, and **latency preference**. By leveraging the first four preferences, we construct human preference prompts to efficiently guide GPT-4/4o in generating preference data for the SiMT task. In the preference optimization phase, SimulPL integrates **latency preference** into the optimization objective and enables SiMT models to improve the read/write policy, thereby aligning with human preferences more effectively. Experimental results indicate that SimulPL exhibits better alignment with human preferences across all latency levels in Zh$\rightarrow$En, De$\rightarrow$En and En$\rightarrow$Zh SiMT tasks. Our data and code will be available at https://github.com/EurekaForNLP/SimulPL. | simultaneous machine translation, simultaneous preference optimization, human preferences | null | 13,982 | 2502.00634 | [
0.0012093124678358436,
-0.07053827494382858,
-0.018033631145954132,
-0.0021536711137741804,
-0.017011990770697594,
0.012583891861140728,
0.028261182829737663,
0.034645382314920425,
0.0887381061911583,
-0.013798994943499565,
-0.023891594260931015,
-0.0629531517624855,
0.00004453867950360291,
-0.019681554287672043,
0.059880055487155914,
-0.06478643417358398,
0.095355324447155,
0.015629922971129417,
-0.08211864531040192,
-0.043553341180086136,
0.021693220362067223,
-0.0176826324313879,
0.0339488610625267,
-0.04851068556308746,
0.03021123819053173,
-0.016812104731798172,
-0.061544910073280334,
-0.014173733070492744,
0.019262351095676422,
0.042912084609270096,
-0.03824771195650101,
-0.007146232761442661,
-0.01728847809135914,
-0.01193282287567854,
-0.0615866482257843,
0.02730645425617695,
-0.08656713366508484,
-0.0844438225030899,
-0.03832082450389862,
-0.029780961573123932,
-0.04005806893110275,
-0.006566093303263187,
-0.019628295674920082,
0.021100016310811043,
0.06606040894985199,
0.06192883104085922,
0.02836027182638645,
-0.009756704792380333,
0.025009384378790855,
0.042880360037088394,
-0.09220243245363235,
-0.020870108157396317,
0.015614544041454792,
0.0049817850813269615,
0.0628482922911644,
0.06093234941363335,
-0.018623415380716324,
0.08245538920164108,
0.029838846996426582,
-0.03788217529654503,
-0.11998803913593292,
-0.08661356568336487,
-0.03579353541135788,
0.004476325120776892,
-0.004975811578333378,
-0.019776523113250732,
0.01336789969354868,
0.08072658628225327,
-0.02596593089401722,
0.05555931106209755,
-0.030683964490890503,
0.08117841184139252,
-0.0158231258392334,
0.03836028650403023,
-0.022952264174818993,
0.03379365801811218,
-0.031190862879157066,
-0.06702709197998047,
0.030404480174183846,
-0.018129494041204453,
-0.01747814379632473,
-0.0029792357236146927,
-0.0014390982687473297,
0.03280987590551376,
0.03837951272726059,
-0.05160230025649071,
0.032260339707136154,
0.0959256961941719,
0.01937423273921013,
0.04050512611865997,
-0.06833935528993607,
0.027355695143342018,
0.09193485975265503,
0.02064475603401661,
-0.08907568454742432,
0.0853324681520462,
-0.02633504942059517,
0.004151734057813883,
-0.05764245614409447,
0.040323056280612946,
0.04372691363096237,
0.05380275100469589,
-0.003988966345787048,
-0.017001643776893616,
-0.023623814806342125,
-0.07283033430576324,
0.002559800399467349,
0.0819811001420021,
0.037112876772880554,
-0.060352105647325516,
-0.02336292900145054,
0.022915344685316086,
-0.01114440243691206,
-0.054120004177093506,
0.005682218354195356,
0.02322901040315628,
0.028464732691645622,
-0.032354019582271576,
0.05052207037806511,
0.059876468032598495,
-0.05126265808939934,
-0.026764601469039917,
0.008574194274842739,
0.00004126553176320158,
0.046234130859375,
-0.0704912394285202,
0.01120968908071518,
-2.5380548522535184e-34,
-0.006935340352356434,
0.03898889943957329,
0.04479461908340454,
-0.02564937248826027,
0.04679425060749054,
-0.01449937280267477,
0.007147531490772963,
-0.008194581605494022,
-0.09421460330486298,
-0.09213326126337051,
-0.0013395705027505755,
-0.0738503560423851,
-0.06369507312774658,
0.13495664298534393,
-0.07027970254421234,
-0.04980950057506561,
0.038799818605184555,
0.09519382566213608,
-0.04577452689409256,
-0.001506261876784265,
0.09288500994443893,
0.049388621002435684,
0.006573685444891453,
-0.01598878763616085,
0.048906195908784866,
-0.0035144987050443888,
0.021035298705101013,
-0.06544908136129379,
0.021069705486297607,
-0.008207864128053188,
-0.10213249921798706,
-0.040336232632398605,
0.017478177323937416,
-0.04786132276058197,
0.00665142061188817,
-0.06577112525701523,
0.009236404672265053,
-0.03411936014890671,
-0.057796601206064224,
-0.015764599665999413,
-0.01570788025856018,
0.00215452304109931,
0.02877742424607277,
-0.012981897220015526,
-0.056206539273262024,
-0.009595315903425217,
-0.009055186994373798,
0.03728195279836655,
0.012287717312574387,
-0.04508369415998459,
0.004067795351147652,
0.07022213190793991,
-0.021835384890437126,
0.0022166245616972446,
0.03387463092803955,
0.03677516058087349,
0.04328622668981552,
0.02207668498158455,
0.0031041339971125126,
0.009885075502097607,
-0.02846416085958481,
0.02664208970963955,
0.02149832434952259,
0.07290413975715637,
0.08676104247570038,
0.03256509080529213,
-0.023320065811276436,
-0.03321768343448639,
0.07558121532201767,
-0.04742962494492531,
-0.06337501108646393,
-0.005040401592850685,
-0.013383808545768261,
-0.0026534763164818287,
0.04265547916293144,
0.08040966838598251,
0.0016613199841231108,
-0.06833984702825546,
0.09341608732938766,
0.010538456030189991,
-0.030658412724733353,
-0.018469180911779404,
-0.004495314322412014,
-0.08742980659008026,
-0.028302649036049843,
-0.0790463536977768,
0.04208759218454361,
-0.056748006492853165,
0.013554931618273258,
0.0560314804315567,
-0.04176832363009453,
0.055594947189092636,
-0.02207677811384201,
0.016409745439887047,
0.028486521914601326,
2.6335890230192727e-34,
0.00919055100530386,
-0.0628833994269371,
0.0010958631755784154,
0.12086402624845505,
-0.04031548276543617,
-0.023719025775790215,
0.02050621435046196,
-0.03733613342046738,
0.04482792690396309,
-0.05186716094613075,
-0.11352606862783432,
-0.0823795348405838,
0.05142765864729881,
-0.054066337645053864,
0.035554733127355576,
0.030063502490520477,
0.034486155956983566,
-0.017909139394760132,
0.02445576898753643,
0.06445811688899994,
-0.05144534632563591,
0.03306182846426964,
-0.07905668020248413,
0.084281325340271,
-0.005694314371794462,
0.006437766365706921,
-0.0576251856982708,
0.02751137688755989,
-0.02832076884806156,
-0.03649630770087242,
-0.0479983352124691,
-0.0211432296782732,
-0.05363107845187187,
-0.0703597366809845,
0.021279772743582726,
0.05216655507683754,
0.039451755583286285,
0.05553317815065384,
-0.08222779631614685,
0.20044273138046265,
0.08506395667791367,
-0.044061873108148575,
-0.037607572972774506,
0.012712784111499786,
-0.027259275317192078,
0.058118999004364014,
-0.08415035903453827,
-0.08210081607103348,
-0.047416411340236664,
-0.027244364842772484,
0.043590061366558075,
0.025415726006031036,
-0.061571214348077774,
0.004428731743246317,
0.05882647633552551,
-0.12923327088356018,
0.0065280660055577755,
-0.059505071491003036,
0.093027763068676,
-0.05056526139378548,
0.04216643422842026,
-0.09195297956466675,
0.02631925791501999,
-0.051344577223062515,
0.02419484779238701,
-0.0663008987903595,
0.08031830936670303,
-0.020411930978298187,
0.005074752494692802,
-0.02219727635383606,
0.06192877143621445,
-0.12694905698299408,
0.022703008726239204,
-0.02802014723420143,
0.02391943521797657,
-0.01893772929906845,
0.028835047036409378,
-0.02743932418525219,
0.025840409100055695,
0.02079715207219124,
-0.014633667655289173,
0.1249324381351471,
0.028434712439775467,
0.06023627892136574,
0.009886735118925571,
0.1256657838821411,
0.052906956523656845,
0.11612100899219513,
0.05326450988650322,
0.011621331796050072,
0.012338140979409218,
0.06705435365438461,
-0.011247129179537296,
0.05234713479876518,
-0.12653706967830658,
-4.36785292379227e-8,
-0.10779734700918198,
-0.017172150313854218,
-0.019776200875639915,
0.0634617805480957,
-0.015939263626933098,
-0.07562897354364395,
-0.034526072442531586,
0.05936519801616669,
0.02623739466071129,
0.0048309373669326305,
0.013198411092162132,
0.0032333186827600002,
-0.004661858081817627,
-0.06938166171312332,
-0.010674409568309784,
0.054888948798179626,
0.0633002296090126,
0.0028644020203500986,
0.027614373713731766,
0.026848411187529564,
0.03594983369112015,
0.0801193118095398,
0.02723320946097374,
-0.10036715120077133,
0.01848924160003662,
0.02660558745265007,
0.0047187721356749535,
0.044482871890068054,
0.08420022577047348,
-0.009211620315909386,
-0.009242652915418148,
-0.031366582959890366,
-0.02343449555337429,
-0.06062674894928932,
0.011883242055773735,
0.00545417470857501,
0.020735695958137512,
-0.0019510302226990461,
0.050554562360048294,
-0.014396030455827713,
0.11384334415197372,
0.001718908199109137,
-0.03694814071059227,
0.04061080887913704,
0.027998879551887512,
-0.03896301984786987,
0.00873994454741478,
-0.06999927759170532,
0.004118956159800291,
-0.09230681508779526,
0.030202604830265045,
-0.0879889726638794,
0.04497219994664192,
0.06546247005462646,
0.05097571387887001,
-0.015675503760576248,
0.01606578938663006,
-0.03368893265724182,
0.05382201448082924,
0.0640687346458435,
0.0730544924736023,
0.05839347466826439,
-0.1167823076248169,
-0.0031180179212242365
] | 0 | 0 | 0 | 0 |
|
Neural Interactive Proofs | https://openreview.net/forum?id=R2834dhBlo | [
"Lewis Hammond",
"Sam Adam-Day"
] | Poster | We consider the problem of how a trusted, but computationally bounded agent (a 'verifier') can learn to interact with one or more powerful but untrusted agents ('provers') in order to solve a given task. More specifically, we study the case in which agents are represented using neural networks and refer to solutions of this problem as neural interactive proofs. First we introduce a unifying framework based on prover-verifier games (Anil et al., 2021), which generalises previously proposed interaction protocols. We then describe several new protocols for generating neural interactive proofs, and provide a theoretical comparison of both new and existing approaches. Finally, we support this theory with experiments in two domains: a toy graph isomorphism problem that illustrates the key ideas, and a code validation task using large language models. In so doing, we aim to create a foundation for future work on neural interactive proofs and their application in building safer AI systems. | interactive proofs, game theory, neural networks, safety, multi-agent reinforcement learning | We study how a trusted, weak model can learn to interact with one or more stronger but untrusted models in order to solve a given task. | 13,981 | 2412.08897 | [
-0.12331586331129074,
-0.013705730438232422,
0.0019060141639783978,
-0.019777653738856316,
0.048008278012275696,
0.010603954084217548,
0.059414248913526535,
0.01946588046848774,
0.010366405360400677,
0.010246955789625645,
-0.048818789422512054,
-0.07716729491949081,
0.07724231481552124,
0.06557555496692657,
-0.04295304790139198,
0.02523643523454666,
0.04261937364935875,
0.0532865896821022,
-0.038046035915613174,
-0.04612516984343529,
-0.010115332901477814,
-0.08804110437631607,
-0.02025221846997738,
-0.029426369816064835,
-0.026623593643307686,
-0.04389467090368271,
-0.008392776362597942,
-0.026771588250994682,
0.008112659677863121,
0.018189936876296997,
0.07081739604473114,
-0.03321214020252228,
-0.02673356421291828,
0.027220232412219048,
-0.04267149791121483,
0.10277952998876572,
-0.06169053539633751,
-0.03041743114590645,
0.008750353939831257,
-0.06453469395637512,
-0.0491347461938858,
-0.0512600876390934,
0.010690586641430855,
-0.03130582347512245,
0.0485832504928112,
0.04148593917489052,
-0.01394714880734682,
0.046088770031929016,
-0.04461384937167168,
0.010238186456263065,
-0.15796148777008057,
-0.0508730486035347,
-0.016717813909053802,
-0.023651892319321632,
-0.000271405529929325,
-0.05954352766275406,
0.060512181371450424,
0.013484104536473751,
-0.04092104360461235,
-0.007850741036236286,
0.003648159559816122,
-0.009925919584929943,
-0.0010731387883424759,
0.02434052899479866,
0.0014622168382629752,
0.09023856371641159,
0.020156975835561752,
0.04289161413908005,
-0.03715142607688904,
0.023956386372447014,
0.044168271124362946,
0.05634380877017975,
-0.07552674412727356,
0.000870840041898191,
-0.036665018647909164,
0.027860518544912338,
-0.03247498720884323,
-0.02418562024831772,
0.003048286074772477,
-0.06591380387544632,
-0.01752576418220997,
-0.034926194697618484,
-0.009272796101868153,
0.05632956326007843,
0.037403374910354614,
-0.0063682314939796925,
-0.016917241737246513,
0.01306875515729189,
0.07420226186513901,
0.029199479147791862,
-0.03216392546892166,
-0.000711352564394474,
0.07647760957479477,
0.044241275638341904,
0.06939728558063507,
0.014432390220463276,
-0.017019934952259064,
-0.06733312457799911,
-0.021067079156637192,
0.07933025062084198,
-0.09441199153661728,
0.043162453919649124,
0.05059698969125748,
-0.13277727365493774,
0.08133260905742645,
0.09697490930557251,
0.04763639718294144,
-0.05009501054883003,
0.056537777185440063,
-0.053157638758420944,
-0.014054234139621258,
0.03908120095729828,
0.03660905361175537,
0.08593788743019104,
0.045372478663921356,
-0.0009216639446094632,
0.017520783469080925,
0.023915626108646393,
0.022789515554904938,
0.028818277642130852,
0.06602702289819717,
0.038961492478847504,
0.014180329628288746,
0.03715948387980461,
0.06276074051856995,
-0.0392737090587616,
0.011025233194231987,
-1.371058131726226e-33,
0.03075217828154564,
0.02134452946484089,
0.017188092693686485,
0.05671390891075134,
0.07564069330692291,
0.0026362077333033085,
0.057215366512537,
0.009663622826337814,
-0.04439469426870346,
-0.0152586093172431,
-0.05289286747574806,
0.04920932650566101,
-0.011667036451399326,
0.04852129518985748,
0.022615928202867508,
0.007345898542553186,
0.015435083769261837,
0.006702616345137358,
-0.03794095292687416,
-0.10096529126167297,
0.07382310181856155,
-0.05030931904911995,
0.017796948552131653,
-0.046666745096445084,
0.026998592540621758,
0.03002273663878441,
0.010392018593847752,
-0.04199846088886261,
0.08558446168899536,
-0.025348477065563202,
-0.07850643247365952,
-0.023599643260240555,
0.010906657204031944,
0.08224879950284958,
0.0208772923797369,
0.04205397143959999,
-0.013959343545138836,
-0.06100279837846756,
0.04701041057705879,
0.045847088098526,
0.04527166113257408,
-0.053767960518598557,
0.01447244081646204,
0.05270107835531235,
0.03459657356142998,
-0.024766556918621063,
-0.07488938421010971,
0.009982186369597912,
-0.050923313945531845,
-0.0859011560678482,
-0.01558851357549429,
0.06341376900672913,
-0.013351493515074253,
-0.10209894180297852,
0.028718046844005585,
-0.040810901671648026,
-0.012093141674995422,
0.05728812515735626,
-0.029634255915880203,
0.08787824958562851,
-0.0027906985487788916,
0.07067212462425232,
-0.0006851132493466139,
0.012825756333768368,
-0.11191792041063309,
0.010604625567793846,
-0.07032274454832077,
0.0034168267156928778,
0.12707370519638062,
0.02113211527466774,
-0.043752819299697876,
0.07451321929693222,
-0.031039061024785042,
0.0025828287471085787,
-0.003995455801486969,
-0.029319224879145622,
-0.014878110028803349,
-0.1131291463971138,
-0.020113926380872726,
-0.009869823232293129,
0.01761624775826931,
-0.05033930018544197,
-0.017405418679118156,
0.05635708570480347,
-0.015702174976468086,
-0.07942411303520203,
-0.03572072461247444,
0.014727002941071987,
-0.009870871901512146,
-0.056674323976039886,
-0.05112075433135033,
-0.08865471184253693,
0.023077145218849182,
0.010579518042504787,
0.029622871428728104,
-2.439924721600118e-34,
-0.04387859255075455,
-0.022797780111432076,
-0.06544553488492966,
0.10121674090623856,
-0.08831410855054855,
-0.0850827693939209,
-0.055665310472249985,
-0.0027137594297528267,
-0.00106039980892092,
0.01239934004843235,
-0.06675831973552704,
0.01219469215720892,
0.09454779326915741,
-0.05496865510940552,
0.02034858614206314,
-0.09775794297456741,
-0.03089933842420578,
0.031100956723093987,
0.029373561963438988,
0.0008386581321246922,
0.07405336201190948,
0.0915304496884346,
-0.06261306256055832,
0.03760321065783501,
0.030345415696501732,
-0.006376318633556366,
0.013961409218609333,
0.01385193970054388,
0.030239352956414223,
0.10118257254362106,
0.007000159937888384,
-0.0540212020277977,
-0.02892952412366867,
-0.06933589279651642,
0.005503765773028135,
0.02785896696150303,
0.1272965967655182,
-0.022572437301278114,
-0.03432983160018921,
0.028502082452178,
0.07131980359554291,
-0.12932060658931732,
-0.11233842372894287,
-0.01860046572983265,
-0.03923068940639496,
0.029228657484054565,
-0.04785808548331261,
0.004903513006865978,
0.005376772955060005,
-0.027402391657233238,
0.02206405997276306,
-0.05330898240208626,
-0.022064756602048874,
-0.10110867023468018,
-0.07266170531511307,
-0.04184747487306595,
-0.005624782294034958,
0.03094295598566532,
0.04109187051653862,
0.014033421874046326,
-0.012942967005074024,
-0.031209327280521393,
0.00023773135035298765,
0.04155558720231056,
-0.01576991192996502,
0.002193216700106859,
-0.07354482263326645,
0.054901372641325,
0.03751980885863304,
-0.01518908143043518,
0.088422991335392,
-0.04062562435865402,
-0.05875762552022934,
0.04798613861203194,
0.06380363553762436,
0.039015281945466995,
-0.020510699599981308,
-0.03302814066410065,
0.013065098784863949,
-0.0007222775020636618,
0.0500730499625206,
-0.025342252105474472,
0.11454679071903229,
0.012308123521506786,
0.049845948815345764,
0.05484543368220329,
-0.002899894258007407,
0.04055216908454895,
-0.007550063543021679,
0.04458136111497879,
0.027407964691519737,
0.04381950572133064,
-0.03592969477176666,
0.023494020104408264,
-0.08520112186670303,
-3.783576474347683e-8,
-0.04690692946314812,
0.04894673451781273,
0.08998964726924896,
0.010195635259151459,
0.04519159346818924,
-0.02185250073671341,
0.01727159321308136,
-0.05408357456326485,
-0.05027221143245697,
0.014176647178828716,
0.0782567635178566,
0.022140122950077057,
-0.018168505281209946,
-0.10953487455844879,
0.020044293254613876,
0.05757415294647217,
-0.03623778745532036,
-0.057968489825725555,
-0.0027114793192595243,
0.0036853416822850704,
0.034857869148254395,
0.012716135010123253,
-0.032945722341537476,
0.02171574905514717,
-0.013602862134575844,
-0.046990152448415756,
-0.05119682848453522,
0.02630455233156681,
-0.0043222797103226185,
0.1367119550704956,
-0.015744954347610474,
-0.05458705127239227,
0.10219060629606247,
0.002474085194990039,
0.07218123227357864,
0.14647597074508667,
0.006552842911332846,
-0.02492811158299446,
-0.02542012557387352,
-0.06136751174926758,
-0.052130647003650665,
0.023220868781208992,
-0.0034145554527640343,
-0.06064094230532646,
0.018949678167700768,
-0.019677376374602318,
0.03643965348601341,
-0.1478162556886673,
0.02007146365940571,
-0.04080065339803696,
-0.02509842813014984,
-0.06230293959379196,
-0.040174394845962524,
0.05236981436610222,
0.01875632256269455,
0.03582485765218735,
-0.060475852340459824,
-0.03934631124138832,
0.04947381094098091,
0.10830066353082657,
0.022485144436359406,
0.13461817800998688,
0.06836193799972534,
-0.0053991167806088924
] | 0 | 0 | 0 | 0 |
|
Oracle efficient truncated statistics | https://openreview.net/forum?id=ZS7UEI3vG5 | [
"Konstantinos Karatapanis",
"Vasilis Kontonis",
"Christos Tzamos"
] | Poster | We study the problem of learning from truncated samples: instead of observing
samples from some underlying population $p^\ast$, we observe only the examples that fall in some survival set $S \subset \mathbb{R}^d$ whose probability mass (measured with respect to $p^\ast$) is at least $\alpha$. Assuming membership oracle access to the truncation set $S$, prior works obtained algorithms for the case where $p^\ast$ is Gaussian or more generally an exponential family with strongly convex likelihood --- albeit with a super-polynomial
dependency on the (inverse) survival mass $1/\alpha$
both in terms of runtime and in number of oracle calls to the set $S$. In this work we design a new learning method with runtime and query complexity polynomial in $1/\alpha$.
Our result significantly improves over the prior works
by focusing on efficiently solving the underlying optimization problem using a general
purpose optimization algorithm with minimal assumptions. | truncated statistics, exponential family, statistical learning | null | 13,970 | null | [
-0.07238094508647919,
-0.059179749339818954,
-0.0008119094418361783,
0.004057813435792923,
0.08675213158130646,
-0.06791325658559799,
0.04436790943145752,
0.05260126292705536,
-0.020209209993481636,
0.03518619388341904,
-0.016237419098615646,
0.08830048888921738,
-0.02232467569410801,
-0.030054578557610512,
-0.015317309647798538,
-0.0027564712800085545,
0.03390840068459511,
0.053688064217567444,
-0.009712819010019302,
0.04113125056028366,
-0.06335394829511642,
0.04247796908020973,
0.01551993377506733,
0.017561350017786026,
0.06271184235811234,
-0.06254886090755463,
0.0465279296040535,
-0.05867929011583328,
0.039778709411621094,
0.054031696170568466,
0.012533515691757202,
-0.015586493536829948,
0.032910946756601334,
-0.03882163017988205,
-0.0178305022418499,
0.058581672608852386,
-0.053167399019002914,
0.03842780366539955,
0.01065076608210802,
0.055840253829956055,
0.019046850502490997,
0.04472040757536888,
-0.014124265871942043,
0.002781092654913664,
0.005286393687129021,
0.026513295248150826,
-0.029827991500496864,
-0.043656229972839355,
-0.01664458028972149,
-0.026974687352776527,
-0.02640756405889988,
0.043538544327020645,
-0.056435924023389816,
0.007654896005988121,
-0.05436112359166145,
-0.08115332573652267,
-0.00637823948636651,
-0.0546528659760952,
-0.055037721991539,
0.010768537409603596,
-0.02515326626598835,
-0.08262713253498077,
-0.030623216181993484,
-0.02473866380751133,
-0.0031140611972659826,
0.043440643697977066,
0.05265571549534798,
0.0071222782135009766,
0.05463103950023651,
0.08807916194200516,
-0.07468027621507645,
0.08977299928665161,
-0.17962734401226044,
0.1253444254398346,
-0.08762151002883911,
0.04082631319761276,
0.05518657714128494,
-0.030741563066840172,
-0.013680360279977322,
0.003972863778471947,
0.07326953858137131,
-0.015736063942313194,
-0.04435522481799126,
-0.01259850338101387,
0.09782837331295013,
0.005612778477370739,
0.01660439744591713,
0.026635991409420967,
0.10034108906984329,
-0.006865647155791521,
-0.0190530177205801,
-0.011829059571027756,
-0.036967430263757706,
-0.02258976548910141,
-0.04322204738855362,
-0.01770063117146492,
0.06650250405073166,
-0.020771639421582222,
0.0719938799738884,
0.09494002163410187,
-0.07369997352361679,
0.10258187353610992,
0.03250429779291153,
-0.03865329176187515,
-0.06471656262874603,
-0.06838248670101166,
-0.053744908422231674,
0.01644747518002987,
0.05531599000096321,
-0.09068230539560318,
0.05971868708729744,
-0.1022145226597786,
0.0400598980486393,
0.009111195802688599,
0.004855217412114143,
0.007032277062535286,
-0.0023891145829111338,
0.018233513459563255,
-0.058538515120744705,
0.03134874999523163,
-0.03417208418250084,
0.035530708730220795,
0.023927772417664528,
0.09327995777130127,
0.035703036934137344,
0.0632038563489914,
-0.007147024851292372,
3.1555179738404726e-33,
-0.010013124905526638,
-0.014815766364336014,
0.039100416004657745,
-0.07989930361509323,
-0.00001710135256871581,
0.09816663712263107,
0.013107589446008205,
-0.02386520616710186,
-0.021751880645751953,
0.04791015014052391,
0.01694697141647339,
-0.0808967873454094,
0.00900606345385313,
-0.05466020479798317,
0.05920293182134628,
0.030257824808359146,
-0.008523240685462952,
0.07295956462621689,
-0.07053963840007782,
-0.07448466867208481,
0.06994250416755676,
-0.04935063049197197,
-0.0015097929863259196,
-0.053154412657022476,
0.010312050580978394,
-0.061736997216939926,
0.010153867304325104,
0.0019003591733053327,
0.0016183608677238226,
-0.0006038232822902501,
-0.05555294454097748,
0.06018972024321556,
-0.013023373670876026,
0.021947413682937622,
0.01753317378461361,
-0.041345641016960144,
0.0025044605135917664,
-0.021046148613095284,
0.018688468262553215,
-0.007820824161171913,
-0.016515329480171204,
0.04056594893336296,
0.09788880497217178,
-0.003181666601449251,
-0.11036650091409683,
-0.14417560398578644,
0.03768427297472954,
0.07094430178403854,
-0.015014313161373138,
-0.0846831351518631,
0.009712518192827702,
-0.05961097031831741,
-0.07721264660358429,
-0.028066543862223625,
-0.014954851008951664,
0.03585556894540787,
-0.0022210287861526012,
0.030414508655667305,
0.03256639838218689,
0.08772171288728714,
0.044240497052669525,
-0.04622829705476761,
-0.004996027331799269,
0.01598571613430977,
-0.0021105410996824503,
0.07426975667476654,
-0.009371228516101837,
-0.026158025488257408,
0.09532329440116882,
-0.05744066834449768,
0.1141100600361824,
-0.04230183735489845,
0.051533497869968414,
-0.06579509377479553,
0.0432933084666729,
0.0052225361578166485,
0.09732957929372787,
0.017889220267534256,
-0.034113138914108276,
-0.007704746909439564,
-0.0025923922657966614,
0.00017733323329593986,
0.0014087505405768752,
-0.028565485030412674,
0.03789703920483589,
-0.07146621495485306,
0.02357906848192215,
-0.11632490158081055,
-0.07840108871459961,
-0.08457868546247482,
-0.08851896226406097,
-0.05807437002658844,
-0.07598506659269333,
0.01901691034436226,
-0.009569458663463593,
-1.9463032818679016e-33,
-0.06846164911985397,
-0.05750547721982002,
0.00004081616862094961,
0.012596352957189083,
0.015317551791667938,
-0.06866688281297684,
-0.06262166053056717,
-0.021238861605525017,
-0.05201289430260658,
-0.09372944384813309,
-0.039144136011600494,
0.011620716191828251,
-0.030439961701631546,
0.025466425344347954,
-0.021075036376714706,
0.012496789917349815,
0.05773003026843071,
0.06720300763845444,
-0.00286394776776433,
0.002000166103243828,
0.03185524791479111,
0.029517389833927155,
-0.05345042794942856,
0.06806252151727676,
0.05021461471915245,
0.04640236124396324,
-0.023819096386432648,
0.1282338798046112,
-0.03433519974350929,
-0.005135994404554367,
-0.08281850814819336,
-0.07214438915252686,
-0.10083901137113571,
-0.040411315858364105,
-0.08074738830327988,
0.08201988786458969,
0.05152055621147156,
0.0935726910829544,
-0.07239256054162979,
0.049464236944913864,
0.029563261196017265,
0.02387542836368084,
-0.021978158503770828,
-0.008333452977240086,
0.09901827573776245,
-0.04920843616127968,
0.03650461509823799,
0.034671321511268616,
0.07165645807981491,
0.04410713538527489,
0.004717451985925436,
-0.00499887065961957,
-0.006313278339803219,
0.1486513763666153,
-0.017881568521261215,
-0.032496802508831024,
-0.06512310355901718,
0.0129391485825181,
-0.020004725083708763,
-0.09209991246461868,
-0.05187943950295448,
0.010527394711971283,
-0.036122892051935196,
0.10524317622184753,
0.017486119642853737,
-0.013627837412059307,
0.02902023307979107,
-0.003423623973503709,
-0.05262618139386177,
-0.0012280321680009365,
-0.02094944566488266,
-0.03487687557935715,
0.033782508224248886,
0.0355457179248333,
0.030616849660873413,
0.06078002601861954,
0.0173182375729084,
-0.002356605837121606,
0.03556244447827339,
0.00858037918806076,
-0.024493707343935966,
-0.020182499662041664,
0.025931136682629585,
-0.0036580990999937057,
0.055947866290807724,
0.050177887082099915,
0.006756751332432032,
-0.023395074531435966,
0.05084146186709404,
-0.03650043159723282,
-0.01571360044181347,
0.005900830961763859,
0.058925941586494446,
-0.041780006140470505,
-0.05875714123249054,
-4.314591706133797e-8,
0.05479046329855919,
0.017859313637018204,
-0.02509595826268196,
-0.019461525604128838,
0.20804138481616974,
0.006356570869684219,
-0.003155624261125922,
0.07870155572891235,
-0.07265931367874146,
-0.09673552960157394,
-0.0003016431874129921,
-0.03534705191850662,
0.015973972156643867,
0.005630210507661104,
-0.0065969135612249374,
0.043776676058769226,
-0.022703517228364944,
0.037809740751981735,
-0.07147960364818573,
-0.009140209294855595,
-0.052465636283159256,
0.008123540319502354,
0.0097576342523098,
-0.025431429967284203,
0.01623903214931488,
-0.012161302380263805,
0.006321399938315153,
0.005260651931166649,
0.09633439779281616,
0.03488903492689133,
-0.08736769109964371,
-0.009889988228678703,
0.036081526428461075,
0.006946153473109007,
0.057900991290807724,
0.058313582092523575,
-0.02483566850423813,
-0.03230741247534752,
-0.06404329091310501,
0.0006628718110732734,
0.022395139560103416,
0.0020429398864507675,
0.03833986073732376,
0.04830338433384895,
0.08651789277791977,
0.04715818911790848,
-0.00963228102773428,
0.023119021207094193,
0.04654168337583542,
0.03712470456957817,
0.07149163633584976,
-0.05482491850852966,
0.013904500752687454,
0.007616381626576185,
0.05208497494459152,
-0.00880624633282423,
0.012451589107513428,
-0.025327710434794426,
-0.02805519476532936,
-0.0022268928587436676,
0.021371610462665558,
-0.041144419461488724,
0.004303287714719772,
0.0659954845905304
] | 0 | 0 | 0 | 0 |
|
Drop-Upcycling: Training Sparse Mixture of Experts with Partial Re-initialization | https://openreview.net/forum?id=gx1wHnf5Vp | [
"Taishi Nakamura",
"Takuya Akiba",
"Kazuki Fujii",
"Yusuke Oda",
"Rio Yokota",
"Jun Suzuki"
] | Poster | The Mixture of Experts (MoE) architecture reduces the training and inference cost significantly compared to a dense model of equivalent capacity. Upcycling is an approach that initializes and trains an MoE model using a pre-trained dense model. While upcycling leads to initial performance gains, the training progresses slower than when trained from scratch, leading to suboptimal performance in the long term. We propose Drop-Upcycling - a method that effectively addresses this problem. Drop-Upcycling combines two seemingly contradictory approaches: utilizing the knowledge of pre-trained dense models while statistically re-initializing some parts of the weights. This approach strategically promotes expert specialization, significantly enhancing the MoE model's efficiency in knowledge acquisition.
Extensive large-scale experiments demonstrate that Drop-Upcycling significantly outperforms previous MoE construction methods in the long term, specifically when training on hundreds of billions of tokens or more.
As a result, our MoE model with 5.9B active parameters achieves comparable performance to a 13B dense model in the same model family, while requiring approximately 1/4 of the training FLOPs.
All experimental resources, including source code, training data, model checkpoints and logs, are publicly available to promote reproducibility and future research on MoE. | mixture of experts, large language models, continual pre-training | null | 13,966 | null | [
-0.05108951777219772,
-0.042665719985961914,
0.04591430723667145,
0.11048727482557297,
-0.0196387842297554,
-0.05645209178328514,
-0.007800832856446505,
0.01246290560811758,
-0.060779597610235214,
-0.07727112621068954,
-0.0732007622718811,
-0.03914529085159302,
0.029739057645201683,
0.02370358817279339,
-0.04717816412448883,
0.022581128403544426,
0.11005640029907227,
0.045273881405591965,
-0.12727226316928864,
-0.016714802011847496,
-0.035294219851493835,
-0.04195321351289749,
0.006053128279745579,
0.016165075823664665,
0.060768820345401764,
-0.06080689653754234,
-0.011758510954678059,
-0.02414005808532238,
0.03206125646829605,
-0.06101369857788086,
0.003058208618313074,
0.05323505029082298,
0.06858395785093307,
0.02191663160920143,
-0.02966434508562088,
0.0585666224360466,
-0.05806302651762962,
0.09232307970523834,
0.04955456405878067,
-0.007205107714980841,
0.010675838217139244,
0.014623364433646202,
0.013213691301643848,
0.08385241031646729,
0.051160771399736404,
0.04736560955643654,
0.059787001460790634,
0.013044623658061028,
0.056793104857206345,
-0.02397509664297104,
-0.06495431065559387,
-0.012089605443179607,
0.009342719800770283,
0.02565135434269905,
-0.018940292298793793,
0.062127675861120224,
0.010488098487257957,
0.023703686892986298,
-0.019921207800507545,
-0.008903191424906254,
-0.10710719972848892,
-0.057881757616996765,
-0.06065192446112633,
-0.0011639837175607681,
0.018263807520270348,
-0.005374143365770578,
0.011307116597890854,
0.04144681990146637,
0.01100848987698555,
0.014085152186453342,
0.06663230061531067,
0.0942244678735733,
-0.07590604573488235,
0.02633323334157467,
0.02795453742146492,
0.040483370423316956,
0.059936683624982834,
0.0310239028185606,
0.03331303596496582,
-0.0408145934343338,
0.013283449225127697,
-0.01090441644191742,
-0.04357312619686127,
-0.09133824706077576,
0.06844120472669601,
0.007675837725400925,
0.0025717762764543295,
0.01708008348941803,
0.032380979508161545,
-0.03143709525465965,
0.00836903601884842,
-0.02917594090104103,
0.027343211695551872,
0.007561757229268551,
0.03212609514594078,
0.11346086114645004,
0.052848994731903076,
-0.08611539006233215,
-0.031270090490579605,
0.15034185349941254,
-0.021957475692033768,
0.01714949682354927,
0.07258006930351257,
0.03588611260056496,
-0.060111988335847855,
-0.04970039799809456,
0.025235172361135483,
0.035678964108228683,
-0.003823758801445365,
-0.029585031792521477,
0.04419242963194847,
0.014689477160573006,
-0.00039977990672923625,
-0.017929472029209137,
-0.032215751707553864,
0.06404714286327362,
-0.06338515132665634,
-0.014515389688313007,
-0.03103860840201378,
0.08793163299560547,
-0.07567060738801956,
0.05266766622662544,
0.003927809186279774,
0.001935093547217548,
-0.0017184976022690535,
0.019558122381567955,
-0.14261889457702637,
4.280207458820221e-33,
0.05846405029296875,
0.023037699982523918,
0.048043884336948395,
-0.0012393133947625756,
-0.028304653242230415,
-0.06106291338801384,
0.041320230811834335,
0.04249880835413933,
-0.03329404816031456,
-0.01334734819829464,
-0.019620848819613457,
0.007430653553456068,
-0.018530787900090218,
0.0398329459130764,
0.04699695482850075,
-0.02533872239291668,
-0.03715825825929642,
0.08396946638822556,
-0.047561146318912506,
-0.01439544279128313,
0.04040813818573952,
-0.007951557636260986,
-0.05411052703857422,
-0.001077265595085919,
-0.05900624766945839,
-0.009784312918782234,
0.05701342597603798,
-0.054788388311862946,
-0.07918576896190643,
0.04888148978352547,
-0.12352442741394043,
-0.05257261171936989,
-0.036726001650094986,
0.049326542764902115,
0.03158276900649071,
-0.028024639934301376,
0.021699680015444756,
-0.04483191296458244,
0.0012188967084512115,
-0.06290847063064575,
-0.05003702640533447,
0.00513159204274416,
-0.02163107506930828,
-0.007709418423473835,
-0.10084403306245804,
-0.06587222963571548,
0.00011913787602679804,
-0.014935733750462532,
-0.06758780032396317,
0.012089570052921772,
0.051298804581165314,
-0.016210710629820824,
0.003811413422226906,
-0.02843894436955452,
-0.051527585834264755,
0.03948276489973068,
0.0669630765914917,
0.029021887108683586,
0.07347959280014038,
0.018392756581306458,
-0.006794946733862162,
0.08582029491662979,
-0.03679867833852768,
0.11411496996879578,
0.01696496270596981,
0.0012978854356333613,
-0.0033199975732713938,
0.08839212357997894,
0.054940082132816315,
-0.03879343718290329,
-0.037030864506959915,
-0.018513360992074013,
0.001996577950194478,
-0.06177787855267525,
0.09742674231529236,
-0.0053234645165503025,
0.028700290247797966,
-0.08853822201490402,
0.01284728106111288,
0.006054393015801907,
0.03135102987289429,
0.05953767150640488,
-0.03308643400669098,
-0.10004299879074097,
-0.027341436594724655,
0.012811916880309582,
0.05007357895374298,
-0.06569933891296387,
0.0030239641200751066,
-0.033938828855752945,
-0.031755901873111725,
-0.0462704673409462,
-0.03053002990782261,
-0.07732261717319489,
0.0068311369977891445,
-2.9434944251246412e-33,
-0.018179399892687798,
0.04152371734380722,
-0.016292212530970573,
0.07619450986385345,
0.008132313378155231,
-0.0683586522936821,
0.024490125477313995,
-0.0061773937195539474,
-0.06621726602315903,
-0.09803051501512527,
0.009736503474414349,
-0.03676428273320198,
0.013585723005235195,
-0.02164161205291748,
0.03046608716249466,
0.005466260481625795,
-0.11323371529579163,
0.02415436878800392,
0.06354545801877975,
0.04290341958403587,
0.03420305997133255,
-0.06579593569040298,
-0.11132453382015228,
0.04748136177659035,
-0.06237590312957764,
0.0034237729851156473,
-0.0960618406534195,
0.15014776587486267,
0.04838771000504494,
0.056617289781570435,
-0.06841105967760086,
0.013679580762982368,
-0.040412839502096176,
-0.01292456779628992,
-0.06343137472867966,
0.09889979660511017,
0.005214867182075977,
0.036406297236680984,
-0.01767675019800663,
0.04708615690469742,
-0.00005567644984694198,
-0.031391691416502,
-0.08304863423109055,
0.01904364675283432,
-0.03599414601922035,
-0.000152961554704234,
0.00029280237504281104,
-0.01662607118487358,
0.051736246794462204,
-0.008943573571741581,
-0.010336709208786488,
0.046555351465940475,
-0.04395099729299545,
0.020332684740424156,
-0.06420108675956726,
-0.037939999252557755,
0.036943480372428894,
0.00269591948017478,
-0.01932365447282791,
0.008658244274556637,
-0.013332713395357132,
-0.03593859076499939,
-0.00041107265860773623,
-0.04694497585296631,
-0.0169927179813385,
-0.032802291214466095,
0.014038262888789177,
0.03682322055101395,
-0.08994874358177185,
0.03996969386935234,
-0.06794475018978119,
-0.005944757256656885,
0.03791941702365875,
-0.016849039122462273,
0.007219571154564619,
0.016947587952017784,
0.014093823730945587,
-0.036943357437849045,
0.013111691921949387,
-0.011625656858086586,
-0.07042581588029861,
-0.0029800066258758307,
-0.040567804127931595,
0.054783064872026443,
0.09520850330591202,
0.0786413848400116,
-0.006633241660892963,
0.042300350964069366,
0.00765095092356205,
0.02720462717115879,
-0.03455503657460213,
0.005104922689497471,
0.07278502732515335,
0.05922967568039894,
-0.10829149186611176,
-4.6258282537792184e-8,
0.00127974443603307,
0.051950130611658096,
0.0064486232586205006,
0.0832839235663414,
0.13191619515419006,
-0.029039014130830765,
-0.04461533948779106,
0.11821269243955612,
0.0033065874595195055,
-0.08452729135751724,
0.05291879549622536,
-0.029363511130213737,
-0.03423638641834259,
-0.0017050511669367552,
0.08091291040182114,
0.05703600496053696,
0.04330752044916153,
-0.0024980728048831224,
-0.053672295063734055,
-0.032560981810092926,
0.10416562110185623,
-0.0071099442429840565,
0.05821859464049339,
-0.0654972717165947,
0.0813012644648552,
-0.10037966817617416,
-0.014802982099354267,
0.11623407155275345,
0.06779846549034119,
0.07050083577632904,
-0.03039402700960636,
-0.017956893891096115,
0.0010543386451900005,
-0.06656456738710403,
0.12213491648435593,
0.0869278535246849,
0.02936760149896145,
-0.03571000695228577,
-0.02056662179529667,
-0.04760867729783058,
-0.02967202104628086,
0.004937338177114725,
-0.05171133950352669,
0.025007613003253937,
0.03083675540983677,
-0.00998820923268795,
-0.09589585661888123,
-0.029296161606907845,
0.06175379455089569,
-0.03823632746934891,
-0.029433339834213257,
0.010169930756092072,
-0.03678388148546219,
0.0935426652431488,
0.07467048615217209,
0.0716334655880928,
0.015112316235899925,
-0.02865142561495304,
-0.03194386884570122,
-0.0037361904978752136,
-0.00717555545270443,
-0.0534093976020813,
-0.02123119682073593,
0.06880929321050644
] | 0 | 0 | 0 | 0 |
|
Black-Box Detection of Language Model Watermarks | https://openreview.net/forum?id=E4LAVLXAHW | [
"Thibaud Gloaguen",
"Nikola Jovanović",
"Robin Staab",
"Martin Vechev"
] | Poster | Watermarking has emerged as a promising way to detect LLM-generated text, by augmenting LLM generations with later detectable signals. Recent work has proposed multiple families of watermarking schemes, several of which focus on preserving the LLM distribution. This distribution-preservation property is motivated by the fact that it is a tractable proxy for retaining LLM capabilities, as well as the inherently implied undetectability of the watermark by downstream users. Yet, despite much discourse around undetectability, no prior work has investigated the practical detectability of any of the current watermarking schemes in a realistic black-box setting. In this work we tackle this for the first time, developing rigorous statistical tests to detect the presence, and estimate parameters, of all three popular watermarking scheme families, using only a limited number of black-box queries. We experimentally confirm the effectiveness of our methods on a range of schemes and a diverse set of open-source models. Further, we validate the feasibility of our tests on real-world APIs. Our findings indicate that current watermarking schemes are more detectable than previously believed. | llm, watermarking | null | 13,958 | 2405.20777 | [
-0.09321487694978714,
-0.04461606591939926,
0.01982736773788929,
0.03837455436587334,
0.12631262838840485,
0.006052641198039055,
-0.052731141448020935,
-0.08584925532341003,
0.05494708940386772,
-0.08448407053947449,
-0.052314676344394684,
-0.005248229950666428,
0.13072505593299866,
0.004520892631262541,
-0.026958800852298737,
-0.028276123106479645,
0.006080181337893009,
0.04820536449551582,
-0.05455605313181877,
-0.09818754345178604,
0.03526068478822708,
-0.03292332589626312,
0.008638049475848675,
-0.010258719325065613,
0.05999770388007164,
0.00005004222475690767,
-0.006743248552083969,
-0.01700594276189804,
-0.014810832217335701,
0.03197525069117546,
0.043095823377370834,
-0.01049518771469593,
-0.029149161651730537,
0.04596956819295883,
0.034687500447034836,
-0.01767127402126789,
-0.019223814830183983,
-0.04373583942651749,
0.06579726189374924,
-0.055823955684900284,
-0.03597501665353775,
-0.07669225335121155,
-0.06949500739574432,
0.04104578122496605,
0.005160249304026365,
-0.009601428173482418,
0.052232492715120316,
0.07295772433280945,
-0.14234280586242676,
0.06774913519620895,
0.025188986212015152,
-0.05356376990675926,
-0.05339594930410385,
0.025125859305262566,
-0.03482086583971977,
-0.09180117398500443,
-0.012866172008216381,
0.00859170500189066,
-0.09818920493125916,
0.019526563584804535,
-0.044836223125457764,
-0.05461893603205681,
-0.015842851251363754,
0.01567794196307659,
0.010424260050058365,
0.060427676886320114,
0.024007048457860947,
0.04887743294239044,
0.0367082878947258,
0.0007810689276084304,
-0.03832058981060982,
0.03504516929388046,
-0.01999097503721714,
0.08111892640590668,
-0.0347912460565567,
0.07382306456565857,
-0.01847171038389206,
0.015309007838368416,
-0.01763298176229,
-0.060092855244874954,
0.032609131187200546,
-0.08828001469373703,
0.05230279639363289,
0.09208086133003235,
0.06677918136119843,
0.05609262362122536,
0.0347520187497139,
-0.013384109362959862,
0.029438085854053497,
0.013758341781795025,
-0.05582885071635246,
-0.06613720953464508,
0.12551501393318176,
0.020844977349042892,
-0.00889839231967926,
0.038255784660577774,
-0.02442360110580921,
-0.01594817452132702,
0.02716096304357052,
0.06474065035581589,
-0.07050777971744537,
-0.011474478989839554,
0.03764353692531586,
-0.07510191202163696,
0.10824176669120789,
0.06837623566389084,
0.018445678055286407,
-0.0023964906577020884,
0.06066630408167839,
-0.034228820353746414,
0.001573732472024858,
0.030112318694591522,
-0.008647415786981583,
0.0031158647034317255,
0.01554798148572445,
-0.023365560919046402,
-0.024438980966806412,
-0.05602554231882095,
0.004508441314101219,
-0.0016388145741075277,
0.06616802513599396,
0.04228484630584717,
-0.003687917487695813,
-0.019000230357050896,
-0.0179898701608181,
0.03792847320437431,
0.027409931644797325,
1.630056179575343e-33,
0.07312431186437607,
0.020002001896500587,
-0.05732116475701332,
-0.0163744930177927,
0.05946207791566849,
0.02001033164560795,
0.0156706515699625,
0.0357479602098465,
-0.055482808500528336,
-0.07520280033349991,
0.04646958038210869,
0.00042870864854194224,
-0.03304244950413704,
0.055484358221292496,
0.06360723078250885,
0.03028402477502823,
-0.012232509441673756,
-0.029250340536236763,
-0.017693273723125458,
-0.002854147693142295,
0.023046594113111496,
-0.008789610117673874,
-0.0009869368514046073,
-0.10252632200717926,
-0.011406456120312214,
0.07546907663345337,
0.06570350378751755,
-0.04324113950133324,
0.00007349713268922642,
0.017755787819623947,
-0.08993294090032578,
-0.04677501693367958,
0.06026553362607956,
-0.025340132415294647,
0.027705350890755653,
0.01806851476430893,
-0.0070336670614778996,
-0.04601972550153732,
0.024465564638376236,
0.009473375044763088,
0.018409905955195427,
0.01156825851649046,
-0.056435469537973404,
-0.055239323526620865,
-0.09074867516756058,
0.0030077395495027304,
-0.0739714503288269,
-0.0037834299728274345,
0.002317843958735466,
-0.07959798723459244,
0.07898739725351334,
0.07854995876550674,
-0.04505694657564163,
-0.034705426543951035,
-0.041315142065286636,
-0.06600449234247208,
-0.029201462864875793,
-0.0471637099981308,
0.030340541154146194,
0.04209241643548012,
-0.007854660041630268,
-0.0024742998648434877,
0.05300629884004593,
0.033488012850284576,
-0.002353126183152199,
-0.004470056854188442,
0.021249549463391304,
0.05251897871494293,
-0.0018335956847295165,
-0.011098616756498814,
0.041519083082675934,
-0.025531647726893425,
-0.031191090121865273,
0.04701334238052368,
-0.07959490269422531,
-0.083229199051857,
0.06817340105772018,
0.026379156857728958,
0.007784099783748388,
-0.03508617356419563,
0.06942863017320633,
-0.09936944395303726,
0.10240653157234192,
-0.04432647302746773,
-0.06696480512619019,
-0.08439574390649796,
0.017286818474531174,
-0.08009529113769531,
-0.07241019606590271,
-0.010026409290730953,
0.0178378913551569,
-0.06393051892518997,
-0.0032105089630931616,
-0.036147650331258774,
-0.004525276832282543,
-3.288597158665002e-33,
-0.08129819482564926,
-0.004541012458503246,
-0.06793772429227829,
0.09814413636922836,
-0.05515342950820923,
-0.0773007869720459,
0.044347815215587616,
0.097056083381176,
-0.0030547184869647026,
-0.017656931653618813,
-0.07532037794589996,
0.011461008340120316,
-0.007366212084889412,
0.011954283341765404,
-0.02487456426024437,
-0.022754734382033348,
-0.0009742074180394411,
0.04127601906657219,
-0.01716335490345955,
0.010884526185691357,
0.022251421585679054,
0.011921056546270847,
-0.06159369647502899,
0.055029820650815964,
0.019150281324982643,
0.018451401963829994,
-0.006536092609167099,
-0.009015055373311043,
-0.059803612530231476,
-0.010482629761099815,
-0.0032540929969400167,
0.07627324014902115,
-0.011195608414709568,
-0.06346825510263443,
-0.028488852083683014,
-0.0004048037517350167,
0.031545791774988174,
0.09098584204912186,
-0.023219695314764977,
0.06536131352186203,
0.09103923290967941,
-0.03925846889615059,
-0.025292402133345604,
-0.013519445434212685,
-0.014204028062522411,
0.009138299152255058,
-0.09553801268339157,
0.07816073298454285,
0.07137919962406158,
-0.033565085381269455,
0.025179943069815636,
-0.030970633029937744,
0.03166648745536804,
-0.005372481420636177,
-0.027381453663110733,
-0.010446536354720592,
-0.025293346494436264,
-0.006737768184393644,
-0.0011790302814915776,
0.02693737857043743,
-0.04174251854419708,
0.001300809090025723,
-0.025491300970315933,
-0.05909140035510063,
0.04266999661922455,
-0.05536019802093506,
0.010136757045984268,
0.06155537813901901,
-0.06399703770875931,
-0.024717247113585472,
0.11741716414690018,
-0.09752782434225082,
-0.07298287004232407,
0.05044857785105705,
0.05788739025592804,
0.00931664276868105,
-0.004161089193075895,
0.02380136027932167,
-0.05212745815515518,
0.051307883113622665,
0.04931836575269699,
0.01902468502521515,
0.0014450709568336606,
0.05473608896136284,
0.12466110289096832,
-0.016571758314967155,
-0.02935081161558628,
-0.0013186490396037698,
-0.04213341698050499,
0.06181245669722557,
-0.013430616818368435,
0.09127229452133179,
-0.10724107176065445,
0.06572705507278442,
0.04131884127855301,
-4.114975737934401e-8,
-0.05210740864276886,
0.0008341195061802864,
-0.06778904050588608,
0.02252018079161644,
-0.04163366183638573,
0.04602054879069328,
-0.033393923193216324,
0.06339084357023239,
0.0442538820207119,
-0.13698461651802063,
0.03129882365465164,
-0.07230562716722488,
-0.12724661827087402,
-0.08261633664369583,
-0.06777551025152206,
0.028814980760216713,
-0.042638786137104034,
0.02685750648379326,
-0.06142960488796234,
-0.02383682131767273,
-0.0594843327999115,
0.10450350493192673,
0.03331145644187927,
-0.04329947382211685,
-0.02984805405139923,
0.025210866704583168,
0.08663727343082428,
0.09690957516431808,
0.03931314870715141,
0.07631377875804901,
-0.0009454443352296948,
0.004254891071468592,
0.018723154440522194,
-0.051133040338754654,
0.04175572842359543,
0.10432063043117523,
-0.02914089523255825,
-0.0070412480272352695,
0.007888098247349262,
0.06616847217082977,
0.056803543120622635,
-0.03372590243816376,
-0.035370856523513794,
-0.028557801619172096,
0.05370667949318886,
-0.04714788496494293,
0.023588484153151512,
-0.07960391789674759,
0.05151819437742233,
-0.048547204583883286,
0.03097544237971306,
-0.0863007977604866,
0.01156467106193304,
0.06177777796983719,
0.027996845543384552,
-0.04194284975528717,
-0.013829996809363365,
0.048996638506650925,
0.05431564524769783,
0.012301867827773094,
0.13160516321659088,
-0.01482307817786932,
0.06352769583463669,
0.0623607337474823
] | 0 | 0 | 0 | 0 |
|
ProAdvPrompter: A Two-Stage Journey to Effective Adversarial Prompting for LLMs | https://openreview.net/forum?id=tpHqsyZ3YX | [
"Hao Di",
"Tong He",
"Haishan Ye",
"Yinghui Huang",
"Xiangyu Chang",
"Guang Dai",
"Ivor Tsang"
] | Poster | As large language models (LLMs) are increasingly being integrated into various real-world applications, the identification of their vulnerabilities to jailbreaking attacks becomes an essential component of ensuring the safety and reliability of LLMs.
Previous studies have developed LLM assistants, known as the adversarial prompter, to automatically generate suffixes that manipulate target LLMs into generating harmful and undesirable outputs.
However, these approaches often suffer from low performance or generate semantically meaningless prompts, which can be easily identified by perplexity-based defenses.
In this paper, we introduce a novel two-stage method, $\texttt{ProAdvPrompter}$, that significantly improves the performance of adversarial prompters.
In $\texttt{ProAdvPrompter}$, the first stage (Exploration) utilizes the loss information to guide the adversarial prompter in generating suffixes that are more likely to elicit harmful responses.
Then the second stage (Exploitation) iteratively fine-tunes the prompter using high-quality generated adversarial suffixes to further boost performance.
Additionally, we incorporate the prompt template to aid in the Exploration stage and propose a filtering mechanism to accelerate the training process in the Exploitation stage.
We evaluate $\texttt{ProAdvPrompter}$ against the well-aligned LLMs (i.e., Llama2-Chat-7B and Llama3-chat-8B), achieving attack success rates of 99.68% and 97.12% respectively after 10 trials on the AdvBench dataset, thereby enhancing performance by $\sim 2$ times compared to previous works.
Moreover, $\texttt{ProAdvPrompter}$ reduces training time by 20% on Llama3-Instruct-8B, generates more generalized adversarial suffixes, and demonstrates resilience against the perplexity defense.
An ablation study further evaluates the effects of key components in $\texttt{ProAdvPrompter}$ (the prompt template and the filtering mechanism). | jailbreaking attacks; large language model | null | 13,954 | null | [
-0.0655977874994278,
0.006110694259405136,
0.020438922569155693,
-0.0010409349342808127,
0.07952897995710373,
0.03858701139688492,
0.024257060140371323,
-0.020083477720618248,
0.02229168266057968,
-0.05476012080907822,
-0.039281055331230164,
-0.09934522956609726,
0.11343167722225189,
-0.02371976338326931,
-0.01804785430431366,
0.05584834888577461,
0.0887892097234726,
0.08703334629535675,
-0.05649520084261894,
0.021494189277291298,
0.013957135379314423,
0.04521578922867775,
0.026320315897464752,
-0.0157551858574152,
-0.021466098725795746,
-0.03949350491166115,
-0.019606465473771095,
-0.016280459240078926,
0.013530398719012737,
-0.0006281940150074661,
0.06392276287078857,
-0.0004094833566341549,
-0.0253678597509861,
0.019310135394334793,
-0.03408108279109001,
0.036882348358631134,
-0.03557247295975685,
0.0565594807267189,
0.0013759383000433445,
-0.12078448385000229,
-0.02613792009651661,
-0.029860984534025192,
-0.037821825593709946,
0.008474328555166721,
0.004637556150555611,
-0.0815284252166748,
-0.006187218241393566,
0.0376288965344429,
-0.008377407677471638,
-0.031518302857875824,
-0.011720268055796623,
-0.037794794887304306,
0.02810034714639187,
0.04138946533203125,
-0.0633842945098877,
-0.054752450436353683,
0.011516040191054344,
0.08014308661222458,
-0.03656955063343048,
0.06406810134649277,
-0.0588562935590744,
-0.040136680006980896,
-0.049155205488204956,
-0.0024081829469650984,
-0.0062682838179171085,
-0.04276310279965401,
0.011715576983988285,
0.08104594796895981,
-0.03675156831741333,
0.07933107018470764,
0.023455753922462463,
0.024349650368094444,
-0.022678660228848457,
0.031220130622386932,
0.03270532190799713,
0.08607666939496994,
-0.08039906620979309,
-0.05696948617696762,
0.033782247453927994,
-0.08109210431575775,
0.01449901144951582,
-0.021111924201250076,
0.04100523889064789,
0.043761663138866425,
0.054752200841903687,
0.06066874787211418,
-0.008948564529418945,
0.0011327555403113365,
0.12615706026554108,
0.051843371242284775,
-0.04596647992730141,
-0.08217789232730865,
0.04596998915076256,
0.05071290209889412,
0.06545206904411316,
0.03605445474386215,
-0.0555444061756134,
-0.1299436390399933,
-0.07315627485513687,
0.07338795810937881,
-0.01647738739848137,
0.032414838671684265,
0.028128324076533318,
-0.10804783552885056,
0.04553183540701866,
-0.004314292687922716,
0.03984038904309273,
-0.005818549077957869,
0.05324043706059456,
-0.0444558747112751,
0.01982256956398487,
0.045390136539936066,
0.06746295094490051,
-0.036144353449344635,
0.045286327600479126,
-0.010027122683823109,
-0.021918537095189095,
0.02076098881661892,
0.053890347480773926,
-0.02733851596713066,
0.0312042199075222,
0.03999089077115059,
0.03499717637896538,
0.04449383541941643,
0.022807525470852852,
-0.07549645006656647,
-0.06377579271793365,
1.9089526835544926e-33,
0.0362115353345871,
0.0721086710691452,
-0.06756774336099625,
0.08506458252668381,
0.04474521800875664,
-0.0021095394622534513,
0.0000788364777690731,
0.07836122065782547,
0.024568205699324608,
-0.00024008758191484958,
-0.04660669341683388,
-0.03945980593562126,
0.007805341854691505,
0.04508199170231819,
0.06948830932378769,
0.046188998967409134,
-0.0019873594865202904,
0.00915029551833868,
-0.06792184710502625,
-0.01654870994389057,
0.03824484348297119,
0.032803792506456375,
0.05106575787067413,
-0.014072664082050323,
0.04740730673074722,
0.011660931631922722,
0.04038533940911293,
-0.012507234700024128,
0.044892821460962296,
0.04173751920461655,
-0.11312424391508102,
-0.0454641617834568,
0.010650965385138988,
0.02439851313829422,
0.023331651464104652,
-0.0028623752295970917,
-0.03207476809620857,
-0.055114567279815674,
-0.016852224245667458,
0.011108225211501122,
-0.03346910700201988,
0.009705426171422005,
-0.01187033299356699,
0.015021600760519505,
-0.01497901976108551,
-0.054565053433179855,
-0.0780264362692833,
0.04903647303581238,
-0.08205074071884155,
-0.04511798918247223,
0.014323722571134567,
0.052051644772291183,
0.007719888351857662,
0.011200152337551117,
-0.11259534955024719,
-0.016390886157751083,
-0.00791128259152174,
0.011534356512129307,
-0.030783452093601227,
0.04965934157371521,
-0.020437544211745262,
0.027103567495942116,
-0.023550724610686302,
0.04435201361775398,
-0.0275198332965374,
-0.08881513774394989,
-0.026912443339824677,
0.03520527482032776,
-0.004059510305523872,
0.025247706100344658,
-0.07251632213592529,
0.004322377033531666,
0.010669084265828133,
0.0025348637718707323,
-0.018645677715539932,
-0.08665461838245392,
0.043643366545438766,
-0.013084985315799713,
0.0990421399474144,
-0.027678515762090683,
0.01717814989387989,
-0.02836371213197708,
0.01157455239444971,
0.021052438765764236,
-0.05690880864858627,
-0.044148143380880356,
0.004480733070522547,
-0.07889121025800705,
-0.016161436215043068,
0.018719566985964775,
-0.042225901037454605,
0.007144182454794645,
-0.004260493908077478,
0.06004234030842781,
-0.03231974318623543,
-3.729852758708734e-33,
-0.06354755163192749,
-0.005392904859036207,
-0.09572488069534302,
0.054662253707647324,
-0.07648713141679764,
-0.06932980567216873,
0.00005702843554900028,
0.047881994396448135,
-0.03594239056110382,
-0.05223019793629646,
-0.11042183637619019,
0.01309207919985056,
0.05249253287911415,
-0.02724643237888813,
0.05210024490952492,
-0.05709444358944893,
0.02699246257543564,
0.06251256912946701,
-0.035120680928230286,
0.03979644551873207,
0.016309240832924843,
0.13271312415599823,
-0.10896461457014084,
0.004746947903186083,
-0.018356073647737503,
-0.06010453775525093,
0.011598714627325535,
0.0979347974061966,
0.03552940487861633,
0.07070516049861908,
-0.002366960747167468,
0.0628267154097557,
-0.013337061740458012,
0.000400614197133109,
-0.09483522921800613,
0.08174602687358856,
0.10635793954133987,
-0.009197833947837353,
-0.02107907459139824,
0.06602220982313156,
0.08280980587005615,
-0.0507778599858284,
-0.06908390671014786,
-0.0033874022774398327,
-0.035591766238212585,
-0.019956668838858604,
-0.031154006719589233,
-0.00026416470063850284,
0.06586100161075592,
-0.052962251007556915,
0.029501650482416153,
-0.017254238948225975,
-0.01810227893292904,
0.029820697382092476,
-0.05804517865180969,
-0.14052753150463104,
-0.017448700964450836,
-0.0552348867058754,
0.048034194856882095,
-0.005656187888234854,
-0.06256663799285889,
-0.0131752360612154,
0.044697802513837814,
-0.05605049803853035,
0.010959762148559093,
-0.0204561036080122,
-0.0401635617017746,
0.04270206764340401,
0.03085818514227867,
-0.03472625091671944,
0.07920195907354355,
0.002628308953717351,
-0.050873253494501114,
0.021696066483855247,
0.012352386489510536,
-0.0029154771473258734,
-0.05470855161547661,
-0.12305448204278946,
-0.015149455517530441,
-0.04438650980591774,
0.0380668081343174,
-0.018926365301012993,
0.016979798674583435,
0.00437944894656539,
0.018874527886509895,
0.018203454092144966,
0.04771636053919792,
0.10897582024335861,
-0.017340438440442085,
0.06749206781387329,
0.057537224143743515,
0.06107589229941368,
-0.032554104924201965,
0.10220415890216827,
-0.018949931487441063,
-4.674333098364514e-8,
-0.03485238179564476,
0.008084540255367756,
-0.009926285594701767,
-0.007761585526168346,
0.06602425128221512,
0.013240295462310314,
-0.10963989794254303,
0.03729124739766121,
0.049617547541856766,
-0.09385528415441513,
0.027575982734560966,
0.01604900136590004,
-0.041823647916316986,
-0.09450998157262802,
-0.016369355842471123,
0.06508632004261017,
-0.03950973600149155,
0.03828348219394684,
-0.06195473298430443,
-0.020469721406698227,
-0.02047879807651043,
0.050827931612730026,
-0.06468477845191956,
-0.08971023559570312,
0.01935439370572567,
-0.06876900047063828,
-0.03391832485795021,
0.03210297226905823,
0.00785005185753107,
0.06320197135210037,
-0.025331327691674232,
-0.02110852487385273,
-0.04245680198073387,
0.0012082878965884447,
-0.008361414074897766,
0.21999981999397278,
-0.0012198967160657048,
-0.060145024210214615,
0.040555793792009354,
0.0009233465534634888,
0.025579998269677162,
0.0023181550204753876,
0.006847284734249115,
-0.0196292195469141,
-0.030644109472632408,
-0.01216131541877985,
0.006265360862016678,
-0.1734546273946762,
0.09720741212368011,
-0.03521550074219704,
0.01537492498755455,
-0.04567874222993851,
0.008452873677015305,
0.06818944215774536,
0.005855754483491182,
0.01702013984322548,
-0.04359060898423195,
-0.04866094887256622,
0.01979794166982174,
0.09845252335071564,
0.074893057346344,
0.029048996046185493,
0.04555930942296982,
-0.03194586932659149
] | 0 | 0 | 0 | 0 |
|
Ward: Provable RAG Dataset Inference via LLM Watermarks | https://openreview.net/forum?id=kVrwHLAb20 | [
"Nikola Jovanović",
"Robin Staab",
"Maximilian Baader",
"Martin Vechev"
] | Poster | RAG enables LLMs to easily incorporate external data, raising concerns for data owners regarding unauthorized usage of their content. The challenge of detecting such unauthorized usage remains underexplored, with datasets and methods from adjacent fields being ill-suited for its study. We take several steps to bridge this gap. First, we formalize this problem as (black-box) RAG Dataset Inference (RAG-DI). We then introduce a novel dataset designed for realistic benchmarking of RAG-DI methods, alongside a set of baselines. Finally, we propose Ward, a method for RAG-DI based on LLM watermarks that equips data owners with rigorous statistical guarantees regarding their dataset's misuse in RAG corpora. Ward consistently outperforms all baselines, achieving higher accuracy, superior query efficiency and robustness. Our work provides a foundation for future studies of RAG-DI and highlights LLM watermarks as a promising approach to this problem. | llm, watermarks, dataset inference, rag | We formalize RAG Dataset Inference, introduce a suitable dataset and baselines, and propose Ward, a rigorous method based on LLM watermarks. | 13,947 | 2410.03537 | [
-0.10166645795106888,
-0.0015490468358621001,
0.008840436115860939,
0.0026479263324290514,
0.04581863433122635,
-0.021305585280060768,
0.003833295311778784,
-0.052159395068883896,
-0.010709540918469429,
-0.033594466745853424,
-0.0024080269504338503,
-0.03188306465744972,
0.12144479900598526,
-0.04783047363162041,
-0.0794958621263504,
0.061405155807733536,
0.07200324535369873,
0.10088671743869781,
-0.061612483114004135,
-0.07348024845123291,
-0.08788122236728668,
0.008342154324054718,
0.017307139933109283,
-0.023747697472572327,
0.030409110710024834,
-0.030285701155662537,
-0.01316634751856327,
0.05545273795723915,
0.01924365758895874,
-0.01633523963391781,
-0.01597432605922222,
0.028245391324162483,
-0.016062811017036438,
0.07287602126598358,
0.0006486402708105743,
-0.024145115166902542,
-0.03411043435335159,
0.032974518835544586,
0.017507266253232956,
-0.029258370399475098,
0.007926280610263348,
-0.04321523755788803,
-0.028597556054592133,
0.05100960657000542,
-0.0066909827291965485,
0.022738691419363022,
0.023563919588923454,
0.045791108161211014,
-0.04430427774786949,
0.012258713133633137,
-0.0356026217341423,
-0.0005562760634347796,
0.0004426667874213308,
0.06544017046689987,
-0.07511147111654282,
-0.08026006072759628,
0.016194656491279602,
0.00850035808980465,
-0.050786521285772324,
0.004076386336237192,
-0.07516568154096603,
-0.10795874893665314,
-0.02952839806675911,
0.015865566208958626,
-0.0024032306391745806,
0.02810855582356453,
-0.017654521390795708,
0.10103097558021545,
0.008227312937378883,
0.034782472997903824,
-0.027306072413921356,
0.08349563181400299,
-0.08282015472650528,
0.12000726163387299,
-0.006737398449331522,
0.10669556260108948,
-0.056703854352235794,
-0.03266572952270508,
0.011463087052106857,
-0.11315096169710159,
-0.011364111676812172,
-0.04601861909031868,
0.03407828137278557,
0.08759033679962158,
0.0697173923254013,
0.014955602586269379,
0.033555883914232254,
-0.0004376200377009809,
0.00139889947604388,
0.017920976504683495,
-0.026335949078202248,
-0.00907793641090393,
0.10615357011556625,
-0.023784780874848366,
0.0055074989795684814,
0.04339030757546425,
0.012456257827579975,
-0.002332398435100913,
0.019970405846834183,
0.10487154126167297,
-0.08621679991483688,
0.06572958081960678,
0.012103301472961903,
-0.08423515409231186,
0.0533536933362484,
-0.02779756486415863,
0.06158704683184624,
-0.01431476790457964,
0.04284120351076126,
-0.043147217482328415,
0.05970852077007294,
0.010421205312013626,
-0.00039290424319915473,
-0.03809010982513428,
0.06051022186875343,
-0.06352220475673676,
0.011368626728653908,
-0.019212879240512848,
-0.05965445935726166,
0.0009023564634844661,
0.05698343366384506,
0.08198540657758713,
0.0254882350564003,
-0.015715787187218666,
-0.04771123826503754,
0.04245799034833908,
-0.004075638018548489,
1.62078667193514e-33,
0.10294657945632935,
0.013767794705927372,
-0.04514285922050476,
-0.035741791129112244,
-0.017492614686489105,
-0.021501004695892334,
-0.0637192502617836,
0.036045923829078674,
-0.030183594673871994,
-0.006802921649068594,
0.03083062171936035,
-0.0002479981048963964,
0.002099712612107396,
0.022001229226589203,
0.07681428641080856,
0.10986579954624176,
0.028326619416475296,
0.020123200491070747,
-0.03881441056728363,
0.02026359736919403,
0.09261041134595871,
-0.03538191691040993,
0.007837217301130295,
-0.03644510358572006,
0.003944354597479105,
-0.0018850266933441162,
0.040094755589962006,
0.00602002115920186,
0.020940950140357018,
0.02941814810037613,
-0.02681906893849373,
-0.03975759074091911,
0.045077383518218994,
0.07096303999423981,
0.05796315148472786,
-0.005340539384633303,
-0.005280136130750179,
-0.052570901811122894,
0.016046229749917984,
0.045852191746234894,
-0.020664747804403305,
-0.0031047523953020573,
-0.017986010760068893,
-0.03531918674707413,
-0.08575267344713211,
-0.0651061087846756,
-0.023995930328965187,
-0.010056732222437859,
0.043162956833839417,
0.021935364231467247,
0.12794211506843567,
0.01632021553814411,
-0.01550076249986887,
-0.029315335676074028,
-0.09849430620670319,
-0.017582079395651817,
-0.03814183920621872,
-0.052288737148046494,
0.055593784898519516,
0.11650077998638153,
-0.018758786842226982,
0.026122132316231728,
0.019830193370580673,
0.031416501849889755,
-0.04693799838423729,
-0.024060703814029694,
0.009933184832334518,
0.03773416578769684,
0.009107999503612518,
0.008088882081210613,
0.03617453947663307,
-0.036523811519145966,
-0.04293111339211464,
0.0007743090391159058,
-0.050808392465114594,
-0.047934748232364655,
0.08794865757226944,
0.03383573517203331,
0.006393411662429571,
-0.015290760435163975,
0.11357024312019348,
-0.04886607825756073,
0.02657219022512436,
-0.02897784672677517,
-0.13107705116271973,
-0.06619620323181152,
0.02717762254178524,
-0.06542100757360458,
-0.07436130195856094,
-0.03244027867913246,
0.010550057515501976,
-0.005885221995413303,
-0.05084553360939026,
-0.03206712007522583,
-0.010945012792944908,
-2.8202038021645995e-33,
-0.01583811640739441,
-0.0030241315253078938,
0.015401029027998447,
0.03889511898159981,
-0.04554859176278114,
-0.025453926995396614,
-0.014282102696597576,
0.04234728217124939,
0.029845453798770905,
-0.005566318519413471,
-0.025181245058774948,
-0.03273971378803253,
-0.005297604948282242,
-0.07248368114233017,
0.029047371819615364,
0.040116775780916214,
-0.03109901025891304,
0.038641493767499924,
-0.10923652350902557,
0.0208596121519804,
0.011923125013709068,
0.07398253679275513,
-0.06162193417549133,
0.03156467527151108,
-0.036995839327573776,
-0.0053476858884096146,
0.005676710978150368,
0.005372996442019939,
-0.031857460737228394,
-0.04533086717128754,
-0.05444563552737236,
-0.005009213928133249,
-0.04815437272191048,
-0.10365623980760574,
-0.06668658554553986,
-0.01549252588301897,
0.05530146509408951,
0.08667895942926407,
-0.08621682226657867,
0.01840594783425331,
0.034248363226652145,
0.020692013204097748,
-0.09887626767158508,
-0.0030807103030383587,
-0.05925551801919937,
0.0020897965878248215,
-0.11403415352106094,
0.023104287683963776,
0.06019870936870575,
0.0021704554092139006,
-0.004126028623431921,
-0.020179610699415207,
0.07437475770711899,
0.005069989711046219,
0.011282108724117279,
-0.06298942118883133,
-0.0007096542976796627,
-0.020322928205132484,
-0.014305232092738152,
0.06504480540752411,
-0.01629670150578022,
0.04659014195203781,
-0.07593288272619247,
0.01985389180481434,
0.03613227233290672,
-0.0859505757689476,
-0.019053425639867783,
-0.04804661124944687,
-0.12454095482826233,
-0.03269506245851517,
0.0313890241086483,
-0.0709073394536972,
-0.026418786495923996,
-0.042732905596494675,
0.029696058481931686,
0.04175947979092598,
0.011523853987455368,
-0.0436103530228138,
-0.04224748909473419,
0.06922008097171783,
0.057495277374982834,
-0.06734873354434967,
0.07510663568973541,
0.04744438827037811,
0.03762798756361008,
0.04165804013609886,
-0.04698709771037102,
-0.013090584427118301,
0.02855573780834675,
0.0655902847647667,
-0.0511523075401783,
-0.03197876363992691,
-0.1260174959897995,
0.0921950563788414,
0.001672003767453134,
-3.809124748954673e-8,
-0.010627026669681072,
0.014853743836283684,
-0.015646660700440407,
0.08167469501495361,
0.013368614949285984,
0.042353007942438126,
-0.07430628687143326,
0.1888718158006668,
-0.02738266810774803,
-0.012681465595960617,
0.1019347757101059,
-0.031527210026979446,
-0.07586245983839035,
-0.09659859538078308,
-0.015094744972884655,
0.028672276064753532,
0.01415900606662035,
-0.006717080250382423,
-0.08107157051563263,
-0.001830714987590909,
-0.07668399810791016,
0.08048232644796371,
0.04723970964550972,
-0.028337934985756874,
0.024164171889424324,
-0.03094630129635334,
-0.008927673101425171,
0.11882663518190384,
0.06504862755537033,
0.062245938926935196,
0.001029479899443686,
-0.0227960292249918,
0.03497382998466492,
-0.09593132883310318,
-0.008144143968820572,
0.09833498299121857,
-0.05850900709629059,
-0.007927974686026573,
-0.01371005829423666,
0.037889737635850906,
0.022191068157553673,
0.006566612049937248,
-0.026702139526605606,
0.03291289135813713,
0.035666324198246,
-0.053298093378543854,
0.0018998708110302687,
-0.027333777397871017,
0.09471030533313751,
-0.047711338847875595,
0.02848263829946518,
-0.10312584787607193,
0.05247505009174347,
0.08662272244691849,
0.013082215562462807,
0.05252945050597191,
0.014244193211197853,
-0.021092884242534637,
0.03690627962350845,
0.04176631197333336,
0.09598737955093384,
-0.05800154432654381,
0.005828655790537596,
0.07423923909664154
] | 0 | 0 | 0 | 0 |
|
SCOPE: A Self-supervised Framework for Improving Faithfulness in Conditional Text Generation | https://openreview.net/forum?id=dTkqaCKLPp | [
"Song Duong",
"Florian Le Bronnec",
"Alexandre Allauzen",
"Vincent Guigue",
"Alberto Lumbreras",
"Laure Soulier",
"Patrick Gallinari"
] | Poster | Large Language Models (LLMs), when used for conditional text generation, often produce hallucinations, i.e., information that is unfaithful or not grounded in the input context. This issue arises in typical conditional text generation tasks, such as text summarization and data-to-text generation, where the goal is to produce fluent text based on contextual input. When fine-tuned on specific domains, LLMs struggle to provide faithful answers to a given context, often adding information or generating errors. One underlying cause of this issue is that LLMs rely on statistical patterns learned from their training data. This reliance can interfere with the model's ability to stay faithful to a provided context, leading to the generation of ungrounded information. We build upon this observation and introduce a novel self-supervised method for generating a training set of unfaithful samples. We then refine the model using a training process that encourages the generation of grounded outputs over unfaithful ones, drawing on preference-based training. Our approach leads to significantly more grounded text generation, outperforming existing self-supervised techniques in faithfulness, as evaluated through automatic metrics, LLM-based assessments, and human evaluations. | faithfulness, hallucination, conditional text generation, natural language processing, large language models | We propose a self-supervised method for faithfulness enhancement for conditional text generation. | 13,935 | 2502.13674 | [
-0.051056552678346634,
-0.05049978569149971,
0.010658320039510727,
0.0482599176466465,
0.011756833642721176,
0.00982419028878212,
-0.017712756991386414,
-0.028439681977033615,
0.07010138779878616,
-0.0573715940117836,
-0.02269951067864895,
-0.07438890635967255,
0.09120427072048187,
0.01304544135928154,
-0.027668630704283714,
0.013026837259531021,
0.15973669290542603,
0.0034842039458453655,
-0.12669135630130768,
-0.07281488180160522,
0.07004370540380478,
0.06713611632585526,
0.07423797994852066,
0.007920059375464916,
0.006172229070216417,
-0.030497727915644646,
0.013689953833818436,
-0.01258756686002016,
0.029199151322245598,
-0.018868334591388702,
0.024186450988054276,
0.07179851084947586,
-0.04601417854428291,
0.025467174127697945,
-0.022447660565376282,
0.01332484744489193,
-0.0806165263056755,
0.09952033311128616,
0.028173396363854408,
-0.06892526894807816,
-0.03343942016363144,
0.012191827408969402,
0.0709758773446083,
0.033035021275281906,
0.04877771437168121,
-0.02628825604915619,
-0.0591740645468235,
0.016257239505648613,
-0.03604842349886894,
0.01483707781881094,
-0.07857533544301987,
0.012967066839337349,
-0.018658649176359177,
0.011132057756185532,
-0.02877984754741192,
0.0309530608355999,
0.08086584508419037,
0.05738670751452446,
-0.03309875726699829,
-0.05780423805117607,
-0.026249751448631287,
-0.0990670919418335,
-0.07303114235401154,
-0.08140412718057632,
-0.024885477498173714,
-0.002633471041917801,
0.0059451456181705,
0.1033129096031189,
-0.057764168828725815,
0.0704113095998764,
0.010388943366706371,
0.02487337589263916,
-0.038230616599321365,
0.0771075040102005,
0.012813636101782322,
0.10000499337911606,
-0.04088318720459938,
-0.08156290650367737,
0.03766963630914688,
-0.07919353246688843,
-0.013194460421800613,
0.013969533145427704,
0.10956303775310516,
0.06270501762628555,
0.03199063241481781,
0.021840054541826248,
0.07868918031454086,
0.06380327045917511,
0.010573231615126133,
0.041976191103458405,
-0.11502708494663239,
-0.06009753420948982,
0.06135150045156479,
-0.042730048298835754,
-0.010094898752868176,
0.04784223437309265,
-0.016550039872527122,
-0.0851430743932724,
0.002673275535926223,
0.06885506957769394,
0.0526006855070591,
0.10175972431898117,
-0.0003407007607165724,
-0.07038626074790955,
0.02485612966120243,
-0.002501837443560362,
0.03119148313999176,
0.07059711217880249,
0.023605242371559143,
-0.06415973603725433,
0.05035945773124695,
0.07179392874240875,
-0.06224450096487999,
-0.06025911867618561,
0.04909002408385277,
-0.02535463497042656,
0.059688251465559006,
-0.02423054724931717,
-0.013356279581785202,
0.0680316686630249,
0.009841276332736015,
0.019865140318870544,
0.062498703598976135,
0.06012820079922676,
-0.04592932015657425,
-0.11953502893447876,
-0.0032367997337132692,
1.957342459519069e-33,
0.04185019060969353,
0.033169835805892944,
0.022283608093857765,
0.04329978674650192,
0.04651201143860817,
0.020049061626195908,
-0.0405891127884388,
0.011344613507390022,
0.04373276233673096,
-0.04578239470720291,
0.007397817447781563,
-0.003252946073189378,
-0.02633640356361866,
-0.011140077374875546,
0.08080602437257767,
0.019437799230217934,
-0.042641088366508484,
0.027283232659101486,
-0.008090105839073658,
0.020726457238197327,
0.06836892664432526,
0.04623812809586525,
0.03220156952738762,
-0.0544656440615654,
0.04669450595974922,
-0.009531215764582157,
0.050895269960165024,
-0.0010924929520115256,
-0.06193457171320915,
0.00634507741779089,
-0.05521587282419205,
-0.040098950266838074,
0.08317719399929047,
0.013464936055243015,
0.019355688244104385,
-0.010921075940132141,
-0.0428655743598938,
-0.05868984013795853,
0.03960182145237923,
-0.008838699199259281,
-0.028041958808898926,
0.011023059487342834,
0.04696265608072281,
0.009280148893594742,
-0.0424548014998436,
-0.07059218734502792,
-0.0409129299223423,
0.033731091767549515,
-0.0206836499273777,
-0.06308160722255707,
0.06263435631990433,
0.06922217458486557,
0.03857704624533653,
-0.043307822197675705,
-0.053201597183942795,
0.006001812405884266,
0.026670455932617188,
0.005033239722251892,
0.0619838684797287,
0.010637087747454643,
0.08177988976240158,
0.04492349177598953,
-0.05854097753763199,
0.06044834852218628,
0.014397792518138885,
-0.04730696976184845,
-0.00005526853783521801,
-0.013050680048763752,
0.07664520293474197,
-0.09628196060657501,
-0.030792996287345886,
-0.010490184649825096,
-0.07551881670951843,
-0.04611742123961449,
0.04243910312652588,
0.012450844049453735,
0.028881141915917397,
-0.010525139048695564,
0.02697131410241127,
-0.051351822912693024,
0.09033342450857162,
-0.0035944695118814707,
0.016302401199936867,
-0.09907018393278122,
-0.057556912302970886,
-0.03292325511574745,
0.03606049716472626,
-0.033302560448646545,
-0.037412408739328384,
0.0083121657371521,
0.011023834347724915,
-0.05480413883924484,
0.006959923077374697,
0.0072842808440327644,
0.020448561757802963,
-1.2853002168684744e-33,
-0.01760483905673027,
0.0273244958370924,
-0.03169994056224823,
0.05478658154606819,
0.013551853597164154,
-0.0007545796106569469,
0.022712349891662598,
0.047730952501297,
-0.009547943249344826,
-0.07916173338890076,
-0.08542690426111221,
0.00024142606707755476,
0.10260716080665588,
0.018765244632959366,
-0.0317838229238987,
-0.008409369736909866,
0.04206015542149544,
0.03140296787023544,
-0.039876073598861694,
0.0686543881893158,
-0.039213817566633224,
0.08756428211927414,
-0.08939259499311447,
0.04975300654768944,
-0.018910789862275124,
0.033985476940870285,
-0.03006051480770111,
0.07908415794372559,
-0.007482821121811867,
0.04385053366422653,
0.008979796431958675,
0.05835683271288872,
-0.017804164439439774,
-0.08310214430093765,
-0.08938486874103546,
-0.024764660745859146,
0.04719330742955208,
-0.0009729614248499274,
-0.09769055992364883,
0.07981158047914505,
0.016381196677684784,
0.019515283405780792,
-0.09809324890375137,
-0.002500953385606408,
-0.09259390085935593,
-0.08544924855232239,
-0.0796656608581543,
0.0013946491526439786,
0.05641015246510506,
-0.028269514441490173,
-0.030234582722187042,
-0.008365289308130741,
-0.06663814932107925,
-0.0060751354321837425,
-0.06446166336536407,
-0.11936113983392715,
-0.03849329054355621,
-0.02269742079079151,
0.03807734325528145,
0.009221944957971573,
-0.08213882148265839,
0.011683596298098564,
0.06897144019603729,
-0.07870440930128098,
0.04501733183860779,
-0.10435720533132553,
-0.033484406769275665,
-0.014831825159490108,
0.004347620997577906,
0.03001640923321247,
0.09518314898014069,
-0.007068670354783535,
-0.014248210936784744,
0.020378021523356438,
0.022441210225224495,
-0.014822470024228096,
0.031702958047389984,
-0.0740036740899086,
-0.04173289239406586,
-0.06319087743759155,
0.05833876132965088,
-0.005796039942651987,
0.027326906099915504,
0.09283847361803055,
0.06103956699371338,
-0.0034784292802214622,
0.03697388991713524,
0.08943343162536621,
0.04226855933666229,
0.0827600434422493,
-0.026689741760492325,
-0.0034649898298084736,
-0.08438938111066818,
0.04291433095932007,
-0.02818482741713524,
-4.106125572889141e-8,
-0.03491700813174248,
-0.021906375885009766,
-0.03474678471684456,
0.13394379615783691,
0.0008060064865276217,
-0.0030810823664069176,
0.03433288261294365,
-0.0016876086592674255,
0.010254774242639542,
-0.06320555508136749,
-0.019010653719305992,
0.008313657715916634,
-0.05640583112835884,
-0.07578259706497192,
-0.031668636947870255,
0.04452839866280556,
0.01676095463335514,
0.041320689022541046,
0.01213338878005743,
-0.01938987523317337,
0.05722222104668617,
0.003196832723915577,
-0.03387277200818062,
-0.057114873081445694,
0.0656091645359993,
0.005762539338320494,
-0.019885029643774033,
0.023911718279123306,
0.03147600218653679,
0.02695809118449688,
0.03914232552051544,
-0.04145781695842743,
-0.047932352870702744,
-0.015211998485028744,
-0.05913160368800163,
0.1274951845407486,
0.024512486532330513,
-0.032179977744817734,
0.026400581002235413,
-0.07996571063995361,
0.055477071553468704,
0.07429858297109604,
-0.04546155035495758,
0.04455043002963066,
0.03855929523706436,
-0.04206368327140808,
0.02910188026726246,
-0.10842978209257126,
0.024427803233265877,
-0.08281053602695465,
-0.05048862844705582,
-0.10906830430030823,
0.009390628896653652,
0.005497183185070753,
0.06019677594304085,
0.02139492705464363,
-0.0017884025583043694,
0.039940983057022095,
0.021587874740362167,
0.07656517624855042,
0.025752194225788116,
0.02539014257490635,
0.006594092585146427,
-0.03189992159605026
] | 0 | 0 | 0 | 0 |
|
Clique Number Estimation via Differentiable Functions of Adjacency Matrix Permutations | https://openreview.net/forum?id=DFSb67ksVr | [
"Indradyumna Roy",
"Eeshaan Jain",
"Soumen Chakrabarti",
"Abir De"
] | Poster | Estimating the clique number in a graph is central to various applications, e.g., community detection, graph retrieval, etc.
Existing estimators often rely on non-differentiable combinatorial components. Here, we propose a full differentiable estimator for clique number estimation, which can be trained from distant supervision of clique numbers, rather than demonstrating actual cliques.
Our key insight is a formulation of the maximum clique problem (MCP) as a maximization of the size of fully dense square submatrix, within a suitably row-column-permuted adjacency matrix.
We design a differentiable mechanism to search for permutations that lead to the discovery of such dense blocks.
However, the optimal permutation is not unique, which leads to the learning of spurious permutations. To tackle this problem, we view the MCP problem as a sequence of subgraph matching tasks, each detecting progressively larger cliques in a nested manner. This allows effective navigation through suitable node permutations.
These steps result in MxNet, an end-to-end differentiable model, which learns to predict clique number without explicit clique demonstrations, with the added benefit of interpretability. Experiments on eight datasets show the superior accuracy of our approach. | Graph neural network, distant supervision | We propose a differentiable model for clique number estimation, learning from distant supervision by searching for dense submatrices in permuted adjacency matrices. | 13,932 | null | [
-0.007750284858047962,
-0.09869354218244553,
-0.022677896544337273,
-0.02886972203850746,
0.10222214460372925,
-0.019429216161370277,
0.002833140082657337,
-0.09173252433538437,
0.027371656149625778,
0.014535426162183285,
0.03368612378835678,
0.01711444929242134,
0.059782858937978745,
0.0815502256155014,
-0.1170596331357956,
0.10151676088571548,
0.10640566796064377,
0.050630684942007065,
0.04383383318781853,
0.017279962077736855,
-0.030625078827142715,
0.005177798215299845,
-0.04872430860996246,
-0.022857103496789932,
-0.04053312912583351,
0.02573603205382824,
0.002344513777643442,
-0.026929140090942383,
-0.008942472748458385,
-0.018653374165296555,
0.02725236862897873,
0.03140878677368164,
0.04543290287256241,
-0.018187010660767555,
-0.057279255241155624,
0.03033091500401497,
-0.06328673660755157,
0.029163548722863197,
0.09820397198200226,
0.058768074959516525,
0.06337691843509674,
0.0491330660879612,
0.041246697306632996,
0.03767133504152298,
0.03675141558051109,
0.01648663356900215,
0.010502736084163189,
-0.02562624029815197,
-0.03664318099617958,
0.045410193502902985,
-0.026645034551620483,
0.0016983440145850182,
-0.04469505697488785,
-0.026031814515590668,
0.019600195810198784,
-0.15517795085906982,
-0.020648663863539696,
-0.05377129092812538,
0.03794290870428085,
0.008989045396447182,
0.05599523335695267,
-0.083704873919487,
0.02895689755678177,
-0.0727207288146019,
-0.016504410654306412,
-0.044447705149650574,
0.02254479192197323,
0.051374346017837524,
-0.0015712211607024074,
0.05444449558854103,
0.06486384570598602,
0.11020981520414352,
-0.08483462035655975,
0.08112312108278275,
0.06057550385594368,
0.07972503453493118,
0.011134018190205097,
-0.005277170334011316,
0.00449568685144186,
-0.09523282200098038,
-0.06516731530427933,
0.005542922765016556,
0.05648360773921013,
-0.026431584730744362,
0.06695644557476044,
-0.01370865385979414,
-0.005987487267702818,
0.036714546382427216,
0.008516928181052208,
0.00411613704636693,
-0.12903475761413574,
0.08595173805952072,
0.008657959289848804,
-0.03717513754963875,
-0.029879331588745117,
0.05226818472146988,
0.04114505648612976,
0.010004348121583462,
0.04609638452529907,
0.0995510146021843,
-0.05476304888725281,
0.10451535880565643,
0.04743650555610657,
-0.007075674366205931,
-0.047073762863874435,
0.039060767740011215,
0.018012646585702896,
0.022585760802030563,
0.013715320266783237,
-0.05755855515599251,
-0.01778525300323963,
0.10603216290473938,
0.05033799260854721,
0.03622916340827942,
0.05964426323771477,
-0.11739128082990646,
0.05292583256959915,
0.020121710374951363,
0.08029139786958694,
0.07793620973825455,
0.03301410749554634,
0.021362442523241043,
0.03925122693181038,
-0.018536534160375595,
-0.011660861782729626,
0.09670550376176834,
-0.0764671042561531,
2.771737251725235e-33,
0.07356914132833481,
0.0579436719417572,
0.07574785500764847,
0.06595861166715622,
0.07958520203828812,
-0.051543183624744415,
-0.026469429954886436,
0.011867916211485863,
0.04308675602078438,
-0.014611084945499897,
0.009310493245720863,
0.040883343666791916,
0.030968530103564262,
0.013039388693869114,
0.01029361691325903,
-0.0370488166809082,
0.043756768107414246,
0.014586553908884525,
-0.036261461675167084,
-0.08908163011074066,
0.047968264669179916,
-0.06526833772659302,
0.055668849498033524,
0.027215290814638138,
0.04904448613524437,
-0.0032264613546431065,
0.06460762023925781,
-0.02391684614121914,
0.09502043575048447,
-0.03199423477053642,
-0.0509379617869854,
0.0739746242761612,
0.028590919449925423,
0.06482475250959396,
0.025743836537003517,
-0.044093772768974304,
0.04198872670531273,
-0.027433447539806366,
0.03190970420837402,
-0.04816722869873047,
-0.006902605760842562,
-0.024398040026426315,
0.040253955870866776,
-0.05193676799535751,
-0.01928100176155567,
0.0145560372620821,
-0.0022647539153695107,
0.04765715077519417,
-0.0831986591219902,
-0.0006611755234189332,
0.03036155179142952,
-0.06783042103052139,
-0.0390731543302536,
0.08256318420171738,
-0.013212722726166248,
-0.06303176283836365,
-0.01111710723489523,
0.027025343850255013,
0.054623398929834366,
0.040201637893915176,
0.03257407993078232,
0.027620747685432434,
0.028153549879789352,
0.04711655154824257,
0.006062401924282312,
0.01548397820442915,
-0.057581428438425064,
-0.008758273907005787,
0.12282446026802063,
-0.07719890773296356,
0.08243332803249359,
0.0066875857301056385,
-0.04291931912302971,
-0.050201404839754105,
-0.027070775628089905,
-0.014049176126718521,
-0.03371448442339897,
-0.1098543033003807,
-0.031193522736430168,
0.004379159305244684,
-0.029489323496818542,
-0.006159917917102575,
0.0027821180410683155,
-0.08255699276924133,
-0.04323457181453705,
0.009352130815386772,
-0.010831928811967373,
-0.022897731512784958,
0.012098513543605804,
-0.02722572162747383,
-0.00598133122548461,
-0.037376414984464645,
-0.013358855620026588,
-0.010189144872128963,
0.04285339638590813,
-1.0859595175140463e-33,
-0.016499411314725876,
0.031050564721226692,
0.061926886439323425,
-0.00706896185874939,
0.040370237082242966,
-0.12421755492687225,
-0.004597825929522514,
-0.045932039618492126,
-0.02853884920477867,
-0.040648069232702255,
-0.07918014377355576,
-0.039939120411872864,
0.0451347790658474,
-0.013459447771310806,
-0.018975770100951195,
0.036547593772411346,
0.0178469717502594,
0.04699532687664032,
-0.05482962727546692,
-0.029775450006127357,
-0.08051765710115433,
0.07133365422487259,
-0.04660452902317047,
0.10593031346797943,
0.024942053481936455,
0.05417623743414879,
0.0318034328520298,
0.032315608114004135,
0.030200796201825142,
0.0015231353463605046,
-0.0642428919672966,
-0.03505600243806839,
-0.06938281655311584,
-0.010036244988441467,
-0.0005643116310238838,
0.06485538184642792,
-0.023821938782930374,
0.09460388123989105,
-0.0181853249669075,
0.0514274425804615,
-0.03065820410847664,
-0.031718723475933075,
-0.07726631313562393,
0.008283902890980244,
0.007843188010156155,
0.046182047575712204,
0.019279170781373978,
0.02535325288772583,
0.006484456360340118,
-0.0753985345363617,
-0.03243603557348251,
0.040630120784044266,
-0.013414405286312103,
-0.055264681577682495,
-0.03115081414580345,
-0.02411431446671486,
-0.04103671386837959,
0.10737849771976471,
0.08722776174545288,
-0.010675286874175072,
-0.05811816081404686,
0.04319143667817116,
-0.05548172444105148,
0.04327165707945824,
-0.005866955500096083,
-0.06332750618457794,
-0.0722515881061554,
-0.020118417218327522,
-0.046848002821207047,
0.051698002964258194,
-0.036137040704488754,
-0.05389348044991493,
-0.06976346671581268,
-0.021788733080029488,
0.05924247205257416,
0.025913704186677933,
0.0042663998901844025,
0.023221923038363457,
-0.0028415555134415627,
0.05940200388431549,
-0.06018197908997536,
-0.01584400236606598,
0.010508367791771889,
-0.024275533854961395,
-0.0010551945306360722,
-0.01886507123708725,
0.023497814312577248,
-0.014417462982237339,
0.02663044072687626,
-0.00027141228201799095,
-0.036932241171598434,
-0.0524071641266346,
-0.04405844956636429,
-0.027440788224339485,
0.05945666879415512,
-3.956225214096776e-8,
-0.12974321842193604,
0.007454381790012121,
0.030465485528111458,
-0.057530052959918976,
0.05382222309708595,
0.009716475382447243,
0.04950447008013725,
0.10350071638822556,
0.011784772388637066,
0.06739506870508194,
0.05129215121269226,
-0.0031195119954645634,
0.022850289940834045,
0.026628054678440094,
-0.03689217194914818,
0.04570314660668373,
0.05671118199825287,
-0.049223288893699646,
0.0244063138961792,
-0.05847598984837532,
-0.05894561856985092,
0.0002844241389539093,
-0.008494921959936619,
0.05956738069653511,
-0.03599386289715767,
-0.13809670507907867,
-0.09673990309238434,
-0.012310859747231007,
0.022895503789186478,
0.027828894555568695,
-0.0286214891821146,
-0.012837860733270645,
0.07029066979885101,
-0.033279530704021454,
0.04084179550409317,
0.11625687777996063,
-0.03305581584572792,
-0.025991274043917656,
0.007768615148961544,
-0.06351165473461151,
0.007011703215539455,
0.004268486052751541,
-0.03494951128959656,
-0.0392020121216774,
-0.026686856523156166,
-0.0034367807675153017,
-0.017322970554232597,
-0.011583348736166954,
-0.031414903700351715,
-0.12392859905958176,
-0.005328393541276455,
-0.10176108777523041,
0.03523409366607666,
-0.031093157827854156,
0.06259043514728546,
0.0023359402548521757,
-0.04658055678009987,
-0.0627746507525444,
0.14440086483955383,
0.0010809272062033415,
-0.0623180978000164,
0.009068471379578114,
-0.03683537617325783,
-0.07339345663785934
] | 0 | 0 | 0 | 0 |
|
Reliable and Diverse Evaluation of LLM Medical Knowledge Mastery | https://openreview.net/forum?id=TXfzH933qV | [
"Yuxuan Zhou",
"Xien Liu",
"Chen Ning",
"Xiao Zhang",
"Ji Wu"
] | Poster | Mastering medical knowledge is crucial for medical-specific LLMs. However, despite the existence of medical benchmarks like MedQA, a unified framework that fully leverages existing knowledge bases to evaluate LLMs' mastery of medical knowledge is still lacking. We propose PretexEval, a novel framework that dynamically generates reliable and diverse test samples to evaluate LLMs for any given medical knowledge base. We notice that test samples produced directly from knowledge bases by templates or LLMs may introduce factual errors and also lack diversity. To address these issues, our framework employs predicate equivalence transformations to produce a series of variants for any given medical knowledge point. Finally, these produced predicate variants are converted into textual language, resulting in a series of reliable and diverse test samples. Here, we use our proposed framework to systematically investigate the mastery of medical factual knowledge of 12 well-known LLMs, based on two knowledge bases that are crucial for clinical diagnosis and treatment. The evaluation results illustrate that current LLMs still exhibit significant deficiencies in fully mastering medical knowledge, despite achieving considerable success on some famous public benchmarks. These new findings provide valuable insights for developing medical-specific LLMs, highlighting that current LLMs urgently need to strengthen their comprehensive and in-depth mastery of medical knowledge before being applied to real-world medical scenarios. | LLM Evaluation, Medical Evaluation, Large Language Model | We propose a reliable and diverse evaluation method, aiming to probe the medical knowledge mastery of LLMs. | 13,909 | 2409.14302 | [
0.02601810172200203,
0.06738622486591339,
0.03314797207713127,
-0.008308898657560349,
-0.010839561931788921,
-0.005436368752270937,
0.018117433413863182,
0.06345564872026443,
-0.01575603522360325,
0.023822616785764694,
-0.05381259322166443,
-0.02008175291121006,
0.05098992586135864,
0.09982355684041977,
0.009776024147868156,
0.040545132011175156,
-0.024399403482675552,
0.03632180020213127,
-0.042494576424360275,
0.040463246405124664,
0.018502984195947647,
0.08415460586547852,
0.012286858633160591,
0.06184154748916626,
0.020246613770723343,
-0.004050849471241236,
-0.04891829192638397,
-0.04119072109460831,
0.02809198386967182,
0.027422158047556877,
0.08390086144208908,
0.06021439656615257,
0.04113879054784775,
0.05758882313966751,
-0.011674403212964535,
0.046808600425720215,
0.0006696897908113897,
0.0400431826710701,
-0.028387999162077904,
0.021680613979697227,
0.02179097570478916,
-0.044508013874292374,
0.046674784272909164,
0.0891885906457901,
0.03737879544496536,
0.06249189004302025,
-0.0462796725332737,
0.057832080870866776,
-0.037379927933216095,
0.07917535305023193,
-0.10997777432203293,
-0.013563920743763447,
-0.03438479080796242,
-0.009828465059399605,
0.04979874938726425,
0.09699974209070206,
-0.09198840707540512,
-0.028730731457471848,
-0.1298738718032837,
-0.10707778483629227,
-0.05683937668800354,
-0.12652990221977234,
-0.01986902765929699,
0.015959426760673523,
0.013688893988728523,
-0.00856752973049879,
0.0063209994696080685,
0.011122098192572594,
0.004143241792917252,
0.046768344938755035,
-0.08141618967056274,
0.047305382788181305,
0.0067513482645154,
0.17477530241012573,
-0.03700987622141838,
0.06008170545101166,
0.055316075682640076,
-0.020110763609409332,
-0.004579901695251465,
-0.033774491399526596,
-0.01925545744597912,
0.03926528990268707,
-0.028990618884563446,
0.008144715800881386,
0.06696401536464691,
-0.013650501146912575,
0.035180240869522095,
0.017045387998223305,
-0.0547243170440197,
-0.004698783624917269,
0.03731699287891388,
-0.07447246462106705,
0.04748232290148735,
0.01722083054482937,
0.09150407463312149,
0.02731589414179325,
-0.01766141690313816,
-0.06689181923866272,
0.04487357661128044,
0.037666093558073044,
0.0391169972717762,
0.09927497804164886,
-0.018244856968522072,
-0.03588831424713135,
-0.05859902501106262,
-0.0647573322057724,
0.010416479781270027,
-0.02456304244697094,
0.048411302268505096,
-0.0742705836892128,
0.044678136706352234,
0.05214255675673485,
0.055677734315395355,
-0.06271281838417053,
0.01544089987874031,
-0.007433136459439993,
-0.013647593557834625,
-0.04738224670290947,
0.03201426938176155,
0.00038403045618906617,
-0.02114233188331127,
-0.007528939284384251,
0.03700627014040947,
0.009695550426840782,
0.03753277286887169,
-0.00783436931669712,
-0.032016605138778687,
2.251866426208432e-33,
0.05918702855706215,
0.061127353459596634,
0.07818279415369034,
0.09850107878446579,
-0.011801481246948242,
0.0359431728720665,
0.031710028648376465,
0.052100636065006256,
-0.08457504212856293,
0.03282887488603592,
-0.032182298600673676,
0.02968045137822628,
-0.04677213728427887,
-0.00023584043083246797,
0.0074488199315965176,
0.04789664223790169,
-0.03804739937186241,
-0.0075107174925506115,
-0.026750581339001656,
0.03201247379183769,
0.07079026848077774,
-0.0103231742978096,
-0.07836457341909409,
-0.07005667686462402,
-0.07653772830963135,
0.03523242846131325,
0.03589503467082977,
0.007412112317979336,
0.04115018621087074,
-0.018286796286702156,
-0.12441599369049072,
-0.04204173758625984,
0.04795679822564125,
0.03502677008509636,
0.000271439814241603,
0.000026662459276849404,
-0.0468723438680172,
-0.0803605243563652,
0.026852300390601158,
-0.01248464360833168,
-0.02441476471722126,
-0.005669416859745979,
0.05070611089468002,
-0.015686122700572014,
0.020853610709309578,
-0.046544112265110016,
-0.07705910503864288,
-0.043481484055519104,
0.014006655663251877,
-0.07549644261598587,
0.05452602356672287,
-0.018358584493398666,
-0.01720534637570381,
-0.0378960520029068,
0.0038991060573607683,
0.05468756705522537,
-0.09836691617965698,
0.07199839502573013,
-0.012116018682718277,
0.09443353861570358,
0.010471105575561523,
0.04575246199965477,
-0.04761483892798424,
0.0604124516248703,
-0.02229505218565464,
0.016443144530057907,
-0.044778626412153244,
-0.04616349935531616,
0.05735649913549423,
0.03936007246375084,
-0.04316865652799606,
-0.013695412315428257,
-0.05700976029038429,
-0.03819538652896881,
-0.000890331226401031,
0.0017392990412190557,
-0.006930212490260601,
-0.08336152136325836,
0.01976722665131092,
-0.04589221999049187,
0.02418055571615696,
0.030425161123275757,
-0.035291869193315506,
0.00792590994387865,
0.017710871994495392,
-0.10653815418481827,
0.016001949086785316,
-0.046750761568546295,
-0.08176494389772415,
-0.037819307297468185,
-0.00365231791511178,
0.014485691674053669,
-0.05366136506199837,
0.01652889885008335,
0.0025285156443715096,
-3.879905714617153e-33,
-0.030380209907889366,
0.0070396605879068375,
-0.01758517511188984,
0.09690148383378983,
0.017729008570313454,
-0.04658243805170059,
0.010466446168720722,
0.002396959811449051,
-0.03218498453497887,
-0.06676696240901947,
0.056440725922584534,
-0.08681276440620422,
-0.030956001952290535,
-0.05468049272894859,
-0.08317118138074875,
0.06047763302922249,
-0.07984839379787445,
-0.031141193583607674,
-0.030637210234999657,
0.05899905040860176,
-0.03472491726279259,
0.12974990904331207,
-0.09070375561714172,
0.06620237976312637,
-0.06470108032226562,
0.060004591941833496,
0.041206542402505875,
0.04409205541014671,
-0.011759239248931408,
0.001697394298389554,
-0.036664679646492004,
-0.06057218089699745,
-0.05147576704621315,
0.005732533056288958,
-0.033583641052246094,
-0.02095877379179001,
0.06474307179450989,
-0.03739767149090767,
0.013203511945903301,
0.014849121682345867,
0.0429101325571537,
-0.0020903367549180984,
-0.04151790961623192,
-0.09266044199466705,
-0.017747994512319565,
-0.05594867095351219,
-0.07207057625055313,
0.005723171401768923,
0.1423330307006836,
0.003256050404161215,
-0.006000314373522997,
-0.061242882162332535,
-0.04764268919825554,
0.02542266994714737,
-0.028725575655698776,
-0.1226438358426094,
-0.08670207113027573,
-0.04400399327278137,
-0.010321330279111862,
-0.054682958871126175,
-0.0176005307585001,
0.031432297080755234,
-0.04850263148546219,
0.05523395910859108,
0.030915582552552223,
-0.029953446239233017,
-0.030767880380153656,
0.07242844253778458,
-0.05519828572869301,
0.019440891221165657,
-0.11747768521308899,
-0.0035508249420672655,
-0.06126194819808006,
-0.0640961155295372,
0.08708219230175018,
-0.02351241186261177,
-0.01765877939760685,
-0.07517686486244202,
-0.04216798022389412,
-0.06766533106565475,
0.04032270982861519,
-0.09959720075130463,
0.013677977956831455,
0.10074221342802048,
-0.01239610556513071,
0.08821701258420944,
0.08760780096054077,
0.049607813358306885,
0.015670986846089363,
0.07920350879430771,
-0.00027458861586637795,
0.03611502796411514,
-0.05652456358075142,
0.003549138782545924,
-0.029880937188863754,
-4.438990686139732e-8,
-0.0855073630809784,
0.009111432358622551,
-0.035683076828718185,
0.029550250619649887,
-0.019236665219068527,
-0.0019025098299607635,
-0.10250210762023926,
0.06008294224739075,
-0.04815277084708214,
-0.019912462681531906,
0.009233788587152958,
0.039807841181755066,
-0.07597548514604568,
-0.06526164710521698,
0.016699256375432014,
0.03245614469051361,
-0.06764721870422363,
0.04135744646191597,
-0.02182563953101635,
-0.040374185889959335,
0.028995992615818977,
0.027575112879276276,
0.005592388100922108,
-0.03930285573005676,
0.05768216773867607,
-0.03437434509396553,
-0.023160455748438835,
0.11143045127391815,
0.01256793923676014,
0.028825053945183754,
-0.008067173883318901,
-0.04020090028643608,
0.011212127283215523,
-0.018688270822167397,
0.05458479002118111,
0.0204586461186409,
0.027072018012404442,
-0.0638781264424324,
0.03634703904390335,
-0.03495824709534645,
-0.02260914072394371,
0.005678561981767416,
-0.00921024102717638,
0.08061505109071732,
-0.04013999551534653,
-0.12568573653697968,
-0.01643069088459015,
0.018723031505942345,
0.0022319017443805933,
-0.059737544506788254,
0.03985513001680374,
0.01933998614549637,
0.03965957835316658,
-0.019006500020623207,
0.037866901606321335,
0.0753474235534668,
-0.01319904811680317,
0.02806016430258751,
-0.007591072469949722,
-0.06017711013555527,
0.15298409759998322,
0.033733658492565155,
0.028345925733447075,
0.042849741876125336
] | 0 | 0 | 0 | 0 |
|
SWE-Search: Enhancing Software Agents with Monte Carlo Tree Search and Iterative Refinement | https://openreview.net/forum?id=G7sIFXugTX | [
"Antonis Antoniades",
"Albert Örwall",
"Kexun Zhang",
"Yuxi Xie",
"Anirudh Goyal",
"William Yang Wang"
] | Poster | Software engineers operating in complex and dynamic environments must continuously adapt to evolving requirements, learn iteratively from experience, and reconsider their approaches based on new insights. However, current large language model (LLM)-based software agents often follow linear, sequential processes that prevent backtracking and exploration of alternative solutions, limiting their ability to rethink their strategies when initial approaches prove ineffective. To address these challenges, we propose SWE-Search, a multi-agent framework that integrates Monte Carlo Tree Search (MCTS) with a self-improvement mechanism to enhance software agents' performance on repository-level software tasks. SWE-Search extends traditional MCTS by incorporating a hybrid value function that leverages LLMs for both numerical value estimation and qualitative evaluation. This enables self-feedback loops where agents iteratively refine their strategies based on both quantitative numerical evaluations and qualitative natural language assessments of pursued trajectories. The framework includes a SWE-Agent for adaptive exploration, a Value Agent for iterative feedback, and a Discriminator Agent that facilitates multi-agent debate for collaborative decision-making. Applied to the SWE-bench benchmark, our approach demonstrates a 23% relative improvement in performance across five models compared to standard open-source agents without MCTS. Our analysis reveals how performance scales with increased inference-time compute through deeper search, providing a pathway to improve software agents without requiring larger models or additional training data. This highlights the potential of self-evaluation driven search techniques in complex software engineering environments. | agents, LLM, SWE-agents, SWE-bench, search, planning, reasoning, self-improvement, open-ended | Introduce an inference-time Monte Carlo Tree Search method for Software Agents. | 13,886 | null | [
-0.023143157362937927,
-0.055276207625865936,
-0.011301964521408081,
0.02827429212629795,
0.03991659730672836,
-0.0563177689909935,
-0.013575132936239243,
0.04589453712105751,
0.06106317415833473,
-0.04503347724676132,
-0.10002964735031128,
-0.0691424235701561,
0.055514801293611526,
-0.0024919528514146805,
0.04957900941371918,
0.010564115829765797,
0.09267932921648026,
0.05141962692141533,
-0.014204345643520355,
-0.11062920838594437,
0.0322694405913353,
0.018546173349022865,
0.02499372698366642,
-0.016574189066886902,
-0.013968748040497303,
-0.01670382358133793,
-0.03358510509133339,
-0.06078384816646576,
0.013443919830024242,
0.07289232313632965,
0.04253217577934265,
0.057682186365127563,
-0.018346933647990227,
0.017727743834257126,
0.010969532653689384,
0.08303741365671158,
-0.12318098545074463,
0.03419918566942215,
0.009720348753035069,
0.039687614887952805,
-0.08160264045000076,
0.04517684504389763,
0.03082996793091297,
-0.06009308993816376,
0.07104980200529099,
-0.13951659202575684,
-0.04985031858086586,
-0.040494948625564575,
0.00921277329325676,
0.026684528216719627,
-0.08518028259277344,
-0.08356925845146179,
-0.05317242443561554,
-0.029322464019060135,
0.07920485734939575,
0.07561904191970825,
0.05374079570174217,
-0.011266353540122509,
0.030892975628376007,
-0.07465320825576782,
0.05712723359465599,
-0.0920351967215538,
0.0019330650102347136,
0.011128267273306847,
-0.04011109098792076,
-0.01795198768377304,
-0.027861174196004868,
0.046342216432094574,
0.039730004966259,
0.034041207283735275,
-0.047473907470703125,
0.07661808282136917,
0.03182241693139076,
0.00863584503531456,
-0.005097367335110903,
0.0005083585274405777,
-0.04076613858342171,
0.05060114338994026,
0.09801293909549713,
-0.03265748172998428,
-0.03351297602057457,
0.014338234439492226,
-0.019746003672480583,
0.04750974848866463,
0.04728928953409195,
0.00785751547664404,
0.07141641527414322,
0.08787884563207626,
0.11016062647104263,
0.07074329257011414,
0.0053314752876758575,
-0.02681131660938263,
0.01022302731871605,
0.013111181557178497,
0.00824856385588646,
0.04583335667848587,
0.016697941347956657,
-0.0769522413611412,
-0.0544598326086998,
0.035723455250263214,
0.006157503463327885,
0.04068383947014809,
0.034161195158958435,
-0.09975236654281616,
-0.07864909619092941,
0.005520532839000225,
0.08300701528787613,
0.009573877789080143,
-0.013141484931111336,
-0.0806063562631607,
-0.06703843921422958,
-0.022697295993566513,
0.08030000329017639,
0.013910713605582714,
0.01078894454985857,
0.030105672776699066,
0.08537932485342026,
0.006151373498141766,
-0.034520696848630905,
0.062253519892692566,
0.039709895849227905,
-0.008675816468894482,
0.038639117032289505,
-0.05586772412061691,
0.07751507312059402,
0.0709598958492279,
0.01685970090329647,
-6.778558428171282e-34,
0.07088436186313629,
0.0728723481297493,
0.041393838822841644,
0.06601966917514801,
0.03958410769701004,
-0.017427830025553703,
0.046312008053064346,
0.04088694229722023,
-0.05236578360199928,
-0.0418938547372818,
0.06387151032686234,
0.06359340995550156,
-0.1253928244113922,
0.002390902256593108,
0.07765759527683258,
-0.059224288910627365,
-0.09983561187982559,
0.05650683119893074,
-0.08452306687831879,
-0.0892934501171112,
0.01129713375121355,
-0.06629152595996857,
0.03933974355459213,
-0.07264778017997742,
0.07797874510288239,
0.035874538123607635,
0.0570589043200016,
-0.05301761254668236,
-0.0032762547489255667,
0.002442930592224002,
-0.06703991442918777,
0.03245585784316063,
-0.005558037664741278,
0.03290325030684471,
-0.022847700864076614,
-0.015782920643687248,
-0.057735707610845566,
-0.07260903716087341,
0.015299146994948387,
0.04548940435051918,
-0.04844100400805473,
-0.003534204326570034,
0.04819418117403984,
-0.009382299147546291,
0.013200478628277779,
-0.11990847438573837,
0.0000924780179047957,
-0.029761789366602898,
0.02764303795993328,
-0.06417670100927353,
0.026678167283535004,
0.046584516763687134,
0.024239055812358856,
-0.0042314836755394936,
-0.04993487149477005,
0.02207152545452118,
0.01726563647389412,
0.04536077752709389,
-0.025522347539663315,
0.04193129390478134,
0.006989776622503996,
0.007324910257011652,
0.0021880383137613535,
0.0693569928407669,
0.04816243797540665,
0.06425193697214127,
-0.01842191070318222,
0.11861401051282883,
0.08626129478216171,
0.03307781741023064,
-0.0037436403799802065,
-0.015293467789888382,
0.027342699468135834,
-0.0686277449131012,
0.02779839187860489,
-0.06443581730127335,
0.014285329729318619,
-0.04548433795571327,
0.03777990862727165,
-0.05518840253353119,
-0.0695677176117897,
-0.013668335042893887,
-0.04120504483580589,
-0.038182828575372696,
0.02462068945169449,
-0.05169357359409332,
-0.010314254090189934,
-0.01581084169447422,
0.008263153955340385,
-0.013198678381741047,
-0.07821586728096008,
0.004789708647876978,
-0.06202559173107147,
0.013480410911142826,
0.05392245948314667,
-5.06242254215965e-34,
-0.054133620113134384,
-0.12654298543930054,
-0.027017775923013687,
0.07462077587842941,
-0.014814034104347229,
-0.05397120490670204,
-0.09525477141141891,
-0.08688003569841385,
0.033547937870025635,
-0.09186829626560211,
-0.11475012451410294,
-0.033139120787382126,
0.005780517589300871,
-0.005803443491458893,
0.03793398290872574,
-0.01934926025569439,
-0.006597673054784536,
-0.006508775986731052,
0.03619663789868355,
0.037726741284132004,
0.019425105303525925,
0.08439625054597855,
-0.037785712629556656,
-0.025678006932139397,
0.006888779811561108,
-0.00884996447712183,
0.04074550420045853,
-0.012888767756521702,
0.0868581086397171,
-0.02333243563771248,
0.020048126578330994,
0.0015840594423934817,
-0.03656772896647453,
0.012473482638597488,
-0.023662569001317024,
0.05111899599432945,
0.06202622875571251,
-0.05372801050543785,
-0.018827494233846664,
0.08568592369556427,
0.097293421626091,
-0.1287638396024704,
-0.040923964232206345,
-0.03606904298067093,
0.009618023410439491,
0.02314719185233116,
-0.05319588631391525,
0.026457760483026505,
0.05211697146296501,
-0.033359095454216,
0.010329601354897022,
-0.025096429511904716,
-0.04203902557492256,
-0.0010900478810071945,
-0.06217077374458313,
-0.03438164293766022,
0.007712727878242731,
0.0027776777278631926,
0.0709451362490654,
-0.008195718750357628,
-0.07712975889444351,
0.012239263392984867,
0.10315275192260742,
-0.046021968126297,
-0.03621834143996239,
-0.017052913084626198,
-0.07423531264066696,
-0.0031790139619261026,
-0.020086683332920074,
-0.01974831335246563,
0.018667174503207207,
-0.034999918192625046,
-0.04042140394449234,
-0.015974214300513268,
0.05027385428547859,
-0.02238723635673523,
-0.04006147012114525,
-0.053845878690481186,
-0.03336571902036667,
-0.08967006206512451,
-0.010496976785361767,
0.11026427894830704,
0.01609320007264614,
-0.0023989358451217413,
-0.010374237783253193,
0.05886406451463699,
-0.03046703338623047,
0.1323232650756836,
0.021986596286296844,
-0.025224529206752777,
-0.08246142417192459,
-0.00029810002888552845,
-0.016155483201146126,
0.004999672528356314,
-0.11316826939582825,
-3.7807904362807676e-8,
-0.031231418251991272,
0.03346563130617142,
0.05623378977179527,
0.05821751430630684,
0.03692123293876648,
-0.026887966319918633,
-0.05681212618947029,
0.058560024946928024,
0.012325637973845005,
-0.00629882887005806,
0.0569535531103611,
-0.043058138340711594,
0.00482374569401145,
0.0024967463687062263,
0.0136790219694376,
0.07381163537502289,
-0.04713805764913559,
-0.024179939180612564,
-0.007612824440002441,
-0.050983112305402756,
0.03608507663011551,
0.03874373808503151,
0.017251497134566307,
0.04912879690527916,
0.019319364801049232,
-0.05290479212999344,
-0.016253434121608734,
0.0649838000535965,
0.07153438031673431,
-0.035080213099718094,
0.10192611068487167,
0.008907110430300236,
0.0030656836461275816,
0.04712632670998573,
0.03647507354617119,
0.01783563755452633,
0.05129094049334526,
0.017018044367432594,
0.0685541033744812,
-0.03850138187408447,
-0.029171478003263474,
0.09881847351789474,
-0.008352475240826607,
0.01236517634242773,
-0.013467101380228996,
-0.02663228288292885,
-0.10450205206871033,
-0.04811927303671837,
0.04218316823244095,
-0.07487348467111588,
-0.03071235492825508,
-0.0717562735080719,
-0.061998434364795685,
0.038177549839019775,
0.023901546373963356,
0.04356895014643669,
0.03514725714921951,
-0.011655070818960667,
0.07815880328416824,
0.017551733180880547,
-0.02001882530748844,
-0.0421246774494648,
-0.02301693521440029,
0.02840719185769558
] | 0 | 0 | 0 | 0 |
|
Language Models are Advanced Anonymizers | https://openreview.net/forum?id=82p8VHRsaK | [
"Robin Staab",
"Mark Vero",
"Mislav Balunovic",
"Martin Vechev"
] | Poster | Recent privacy research on large language models (LLMs) has shown that they achieve near-human-level performance at inferring personal data from online texts. With ever-increasing model capabilities, existing text anonymization methods are currently lacking behind regulatory requirements and adversarial threats. In this work, we take two steps to bridge this gap: First, we present a new setting for evaluating anonymization in the face of adversarial LLM inferences, allowing for a natural measurement of anonymization performance while remedying some of the shortcomings of previous metrics. Then, within this setting, we develop a novel LLM-based adversarial anonymization framework leveraging the strong inferential capabilities of LLMs to inform our anonymization procedure. We conduct a comprehensive experimental evaluation of adversarial anonymization across 13 LLMs on real-world and synthetic online texts, comparing it against multiple baselines and industry-grade anonymizers. Our evaluation shows that adversarial anonymization outperforms current commercial anonymizers both in terms of the resulting utility and privacy. We support our findings with a human study (n=50) highlighting a strong and consistent human preference for LLM-anonymized texts. | privacy, anonymization, large language models | We demonstrate how large language models can be employed in an adversarial framework to surpass state-of-the-art anonymization tools both in terms of privacy and utility. | 13,884 | 2402.13846 | [
-0.16379842162132263,
-0.06354240328073502,
-0.02385491132736206,
0.05401712656021118,
0.06125667318701744,
0.02077922224998474,
-0.002006814582273364,
-0.0628124326467514,
0.04231342300772667,
-0.04738250747323036,
0.004223017953336239,
-0.04373464360833168,
0.07764892280101776,
-0.010142500512301922,
-0.01675332337617874,
0.06341943144798279,
0.1050480529665947,
0.08458290249109268,
-0.11842722445726395,
-0.052336324006319046,
0.020472919568419456,
-0.003480373416095972,
0.08084263652563095,
-0.0063462830148637295,
0.04102446883916855,
-0.10435587912797928,
-0.01547737792134285,
-0.042067524045705795,
-0.0730304941534996,
0.0030712413135915995,
0.051457833498716354,
-0.00242236559279263,
0.0006611206335946918,
0.009977792389690876,
-0.010191530920565128,
-0.011282171122729778,
-0.014710141345858574,
-0.01007958222180605,
0.0388910174369812,
-0.06592752039432526,
0.016017882153391838,
-0.11842785030603409,
-0.008700785227119923,
-0.013581137172877789,
0.0017797292675822973,
-0.022143978625535965,
0.04008052870631218,
0.029450787231326103,
-0.0430762842297554,
0.0023907346185296774,
-0.026047494262456894,
-0.012401069514453411,
-0.0038246712647378445,
0.0006098771118558943,
-0.09403388947248459,
-0.06456008553504944,
0.02866293676197529,
0.03651810809969902,
-0.0351419560611248,
0.027237845584750175,
-0.010895196348428726,
-0.09621687233448029,
-0.03566361218690872,
0.04302406683564186,
-0.04648325592279434,
-0.0754663348197937,
0.019428478553891182,
0.06965312361717224,
-0.06485728919506073,
0.027455244213342667,
0.06273972988128662,
0.04509119316935539,
-0.021773502230644226,
0.06312929093837738,
0.03995388373732567,
0.05524945631623268,
-0.008525294251739979,
-0.02178872376680374,
0.0487324558198452,
-0.053956665098667145,
0.022614873945713043,
-0.00991094671189785,
0.08584828674793243,
0.0012524067424237728,
0.009794862940907478,
0.02960398979485035,
-0.021876135841012,
0.01437164656817913,
0.08400936424732208,
0.06872176378965378,
-0.057260774075984955,
-0.01607227511703968,
0.04655615612864494,
0.03398219496011734,
0.05800779536366463,
-0.021325254812836647,
-0.023160558193922043,
-0.02172728441655636,
-0.005408577620983124,
0.08644390106201172,
-0.04101461172103882,
0.05806545168161392,
-0.05344976857304573,
-0.03265761956572533,
0.007005896419286728,
0.06615634262561798,
0.008332326076924801,
-0.017078617587685585,
0.05140785500407219,
0.040976643562316895,
0.08373481035232544,
0.03179461508989334,
0.024820882827043533,
0.013197066262364388,
0.04900115355849266,
-0.047555435448884964,
0.0002550586068537086,
0.033554401248693466,
0.08376053720712662,
0.03876468539237976,
-0.03834659606218338,
0.05643216520547867,
0.029928771778941154,
-0.028251636773347855,
-0.022518141195178032,
0.011151333339512348,
-0.030117033049464226,
-9.808582921005371e-34,
0.06321434676647186,
0.07567642629146576,
-0.03354540467262268,
0.07951609045267105,
-0.018628880381584167,
0.03637056052684784,
-0.04451251029968262,
0.10212636739015579,
-0.06418014317750931,
0.013942791149020195,
-0.025006212294101715,
0.010524787940084934,
-0.008359415456652641,
0.07969564199447632,
0.00780304754152894,
0.08106075972318649,
0.03671727329492569,
-0.009614436887204647,
0.02162528596818447,
-0.005139682441949844,
0.1424013376235962,
0.012628970667719841,
-0.011194064281880856,
0.0031076560262590647,
0.018128378316760063,
-0.014186222106218338,
0.04096253588795662,
-0.06424067169427872,
0.00010059760097647086,
0.05354780703783035,
-0.13082797825336456,
0.007612820714712143,
0.0631268098950386,
0.01226543914526701,
0.1008642390370369,
-0.022556796669960022,
-0.050859905779361725,
-0.02699362114071846,
0.01240723580121994,
-0.0016693505458533764,
-0.05084194988012314,
0.05620900169014931,
-0.0002014431229326874,
-0.033829424530267715,
-0.0863313376903534,
0.02457541786134243,
-0.039200879633426666,
0.04526670277118683,
0.013861903920769691,
-0.03231821581721306,
0.052864931523799896,
0.06024879217147827,
-0.04963383823633194,
0.005217737052589655,
-0.07602445036172867,
0.007215264718979597,
-0.02760300599038601,
0.003091835416853428,
-0.0099206967279315,
0.031675055623054504,
-0.00854873564094305,
0.0757179781794548,
0.019443584606051445,
0.0022035574074834585,
-0.014549657702445984,
-0.04485878720879555,
-0.0008403819520026445,
0.013639944605529308,
0.010696375742554665,
-0.027905769646167755,
0.025877419859170914,
0.02412857860326767,
-0.02816806733608246,
0.0008663953049108386,
-0.023195866495370865,
-0.02824561856687069,
0.060512837022542953,
-0.0015430612256750464,
-0.004721104167401791,
0.005305042956024408,
0.08005588501691818,
-0.007825303822755814,
0.03169381991028786,
-0.05944133177399635,
-0.1272473782300949,
-0.028923405334353447,
0.011328158900141716,
0.005823644809424877,
0.04922928661108017,
0.02499173767864704,
0.0012003506999462843,
-0.04471266269683838,
-0.054401613771915436,
-0.002328765345737338,
0.004809450823813677,
-1.4719046186027354e-33,
-0.10499177128076553,
-0.0025788573548197746,
-0.06822507828474045,
0.020314449444413185,
-0.02209537662565708,
-0.055829886347055435,
0.026455281302332878,
0.08103047311306,
-0.000385731749702245,
-0.02025551348924637,
-0.09261871874332428,
0.01895829103887081,
0.08715757727622986,
-0.04502642899751663,
0.006697181146591902,
-0.004320627544075251,
0.013612290844321251,
0.03861786797642708,
-0.10847601294517517,
0.03413661569356918,
0.019603662192821503,
0.04164324328303337,
-0.10267405956983566,
0.02599172107875347,
0.03881681710481644,
-0.06184879317879677,
-0.030392944812774658,
0.026865661144256592,
0.04078184813261032,
0.02290099859237671,
-0.01744130067527294,
0.07317857444286346,
-0.04574265703558922,
-0.08386754989624023,
-0.09135908633470535,
0.0416192002594471,
0.07406703382730484,
0.03974572569131851,
-0.049237288534641266,
0.03285784274339676,
0.06283394992351532,
-0.0025679150130599737,
-0.09030697494745255,
-0.021797260269522667,
-0.03139315918087959,
-0.011305292136967182,
-0.08826044201850891,
-0.07307952642440796,
0.07239732891321182,
-0.03420157730579376,
0.020880689844489098,
0.006434069946408272,
0.018352612853050232,
0.03368552029132843,
-0.04156597703695297,
-0.13110564649105072,
-0.021557176485657692,
-0.08217586576938629,
-0.00031130321440286934,
0.027874642983078957,
-0.07583510875701904,
0.010898550041019917,
-0.02857598476111889,
0.047602660953998566,
0.0034817508421838284,
-0.046983104199171066,
-0.017471764236688614,
0.002390799578279257,
-0.0233895443379879,
-0.046652473509311676,
0.09154695272445679,
-0.08210285753011703,
-0.06699315458536148,
0.03404814749956131,
-0.029228920117020607,
-0.02641450986266136,
0.04374858736991882,
-0.048543039709329605,
-0.03284609317779541,
0.08166193217039108,
0.118421770632267,
-0.045370057225227356,
0.07333706319332123,
-0.07532362639904022,
0.11677376180887222,
-0.03924165666103363,
0.0680631548166275,
-0.002173907821998,
-0.014696690253913403,
0.08421552926301956,
0.06619047373533249,
0.017569581046700478,
-0.12230107933282852,
-0.02943478524684906,
0.014999359846115112,
-3.715216934097043e-8,
-0.03346133604645729,
-0.04099497199058533,
-0.006182588171213865,
0.10550931096076965,
-0.06305138766765594,
0.022938499227166176,
-0.051995135843753815,
0.04793070629239082,
-0.0196639746427536,
-0.05073285475373268,
0.049281276762485504,
-0.007670174352824688,
-0.03020687960088253,
-0.10767866671085358,
-0.01549459993839264,
0.017127757892012596,
-0.014360044151544571,
0.02017809823155403,
0.01311646867543459,
0.013432894833385944,
-0.021829089149832726,
0.042894188314676285,
-0.02900983765721321,
-0.09090155363082886,
0.03882507607340813,
-0.051839206367731094,
-0.03830663487315178,
0.0291892196983099,
0.06871858984231949,
0.08950173854827881,
-0.05341639742255211,
-0.07358895242214203,
-0.0569290816783905,
-0.02950623817741871,
-0.0754554271697998,
0.10042224824428558,
-0.04344382509589195,
-0.031954891979694366,
0.04223131760954857,
-0.01687164232134819,
0.033627647906541824,
0.046305928379297256,
-0.02702939696609974,
0.01309970486909151,
0.035631511360406876,
-0.009106410667300224,
-0.0416908822953701,
-0.1566348522901535,
0.08553216606378555,
-0.01859036087989807,
0.00502665713429451,
-0.03881639242172241,
0.07269313931465149,
0.09042622148990631,
0.025680378079414368,
-0.032513272017240524,
-0.002503560157492757,
0.006637174636125565,
0.02727395109832287,
0.060918547213077545,
0.0973861813545227,
-0.03829089179635048,
-0.04879182204604149,
-0.02960711345076561
] | https://github.com/eth-sri/llm-anonymization | 10 | 0 | 0 | 0 |
ADAM: An Embodied Causal Agent in Open-World Environments | https://openreview.net/forum?id=Ouu3HnIVBc | [
"Shu Yu",
"Chaochao Lu"
] | Poster | In open-world environments like Minecraft, existing agents face challenges in continuously learning structured knowledge, particularly causality. These challenges stem from the opacity inherent in black-box models and an excessive reliance on prior knowledge during training, which impair their interpretability and generalization capability. To this end, we introduce ADAM, An emboDied causal Agent in Minecraft, which can autonomously navigate the open world, perceive multimodal context, learn causal world knowledge, and tackle complex tasks through lifelong learning. ADAM is empowered by four key components: 1) an interaction module, enabling the agent to execute actions while recording the interaction processes; 2) a causal model module, tasked with constructing an ever-growing causal graph from scratch, which enhances interpretability and reduces reliance on prior knowledge; 3) a controller module, comprising a planner, an actor, and a memory pool, using the learned causal graph to accomplish tasks; 4) a perception module, powered by multimodal large language models, enabling ADAM to perceive like a human player. Extensive experiments show that ADAM constructs a nearly perfect causal graph from scratch, enabling efficient task decomposition and execution with strong interpretability. Notably, in the modified Minecraft game where no prior knowledge is available, ADAM excels with remarkable robustness and generalization capability. ADAM pioneers a novel paradigm that integrates causal methods and embodied agents synergistically. Our project page is at https://opencausalab.github.io/ADAM. | embodied agent, causality, large language model, interpretability, vision language navigation, cross-modal application, cross-modal information extraction, multimodality | null | 13,881 | 2410.22194 | [
0.04080472141504288,
-0.04553912207484245,
0.024279721081256866,
0.043105512857437134,
0.07941237837076187,
-0.00464848754927516,
0.016071021556854248,
-0.05555491894483566,
0.012740659527480602,
0.0035548261366784573,
-0.045584604144096375,
-0.04472701996564865,
0.009971964173018932,
0.0780671238899231,
0.017657019197940826,
0.012839210219681263,
0.09772159904241562,
0.07089763879776001,
-0.10136361420154572,
-0.07133794575929642,
0.01705275848507881,
0.011834685690701008,
0.03328031301498413,
-0.05296549201011658,
-0.058195363730192184,
0.02242952771484852,
0.01519085094332695,
-0.03843696415424347,
0.04728147014975548,
-0.04210725054144859,
0.09012207388877869,
-0.036159124225378036,
-0.06135217472910881,
-0.029771726578474045,
-0.05096976086497307,
0.13944579660892487,
-0.02417907677590847,
0.0687704011797905,
-0.028305165469646454,
-0.13463115692138672,
0.012857607565820217,
0.041026871651411057,
-0.009373658336699009,
-0.004183798562735319,
0.02740613743662834,
0.02702915668487549,
-0.03541555628180504,
-0.05852114409208298,
0.04394881799817085,
0.023477107286453247,
-0.14984580874443054,
-0.08674697577953339,
0.06341397017240524,
-0.03822873532772064,
0.050372831523418427,
0.029034312814474106,
0.007154494058340788,
-0.011877715587615967,
-0.025609342381358147,
-0.024278059601783752,
-0.03553376719355583,
-0.04613589122891426,
-0.004574471153318882,
-0.008163033984601498,
0.024268243461847305,
-0.018620209768414497,
0.009779294021427631,
-0.007104194723069668,
-0.029444200918078423,
0.015329809859395027,
0.04338693991303444,
0.05676817148923874,
0.02646961808204651,
-0.037332892417907715,
0.01915576122701168,
0.01142598781734705,
-0.0033849221654236317,
-0.029987139627337456,
0.008516070432960987,
-0.034272223711013794,
0.05494667589664459,
0.026111813262104988,
-0.0008603461901657283,
0.026052430272102356,
0.05757531896233559,
0.044302672147750854,
0.031685054302215576,
0.06352313607931137,
0.028654111549258232,
0.07763849943876266,
-0.0452057421207428,
-0.022254396229982376,
0.07416847348213196,
-0.014110537245869637,
0.08370757102966309,
0.020036697387695312,
0.02112494595348835,
-0.04884497821331024,
-0.04167122766375542,
0.08183427155017853,
0.015674417838454247,
0.023621922358870506,
-0.013928077183663845,
0.029664942994713783,
0.0004225146840326488,
0.0014754498843103647,
-0.002547891577705741,
0.05965043976902962,
-0.007368461228907108,
-0.054855264723300934,
-0.01080586388707161,
-0.04674477502703667,
0.01020590215921402,
0.014345518313348293,
0.05195920169353485,
-0.023778121918439865,
0.00870050024241209,
0.053272392600774765,
0.03582615405321121,
0.04966229945421219,
0.05780167132616043,
0.025360912084579468,
0.01282716915011406,
0.061718255281448364,
0.02924865297973156,
-0.06000782921910286,
0.015165850520133972,
-2.425740040452295e-33,
0.10437463968992233,
0.025942759588360786,
0.04738670587539673,
0.02929512783885002,
0.10203325748443604,
0.023718401789665222,
-0.043575916439294815,
0.03429362177848816,
-0.03764684498310089,
-0.02681376412510872,
-0.018186485394835472,
0.0063515277579426765,
-0.10242907702922821,
0.1425589621067047,
0.04289600998163223,
-0.0032948311418294907,
-0.0720105841755867,
0.007818263955414295,
0.02235809527337551,
-0.027327610179781914,
0.01893322356045246,
-0.0838065966963768,
-0.016477864235639572,
-0.0004568896838463843,
-0.023570269346237183,
0.039642397314310074,
0.06839864701032639,
-0.00695068622007966,
0.019845658913254738,
-0.014383516274392605,
-0.10742631554603577,
-0.0063437665812671185,
-0.05605483800172806,
0.01045428030192852,
0.005612429231405258,
-0.06500454992055893,
-0.07661819458007812,
-0.03964276239275932,
0.0029410081915557384,
-0.0459832064807415,
-0.00967132207006216,
-0.016794776543974876,
-0.028035609051585197,
-0.023496130481362343,
-0.10549060255289078,
-0.06153314933180809,
0.04462147504091263,
0.07741236686706543,
-0.08123486489057541,
0.0040612355805933475,
0.00003955196007154882,
0.03064594604074955,
0.03714634105563164,
-0.17061059176921844,
-0.054098233580589294,
-0.012552613392472267,
0.06551451981067657,
0.04868027940392494,
0.03552349656820297,
0.008148923516273499,
0.0261986143887043,
0.05226226523518562,
-0.011578117497265339,
0.12355858087539673,
0.030957210808992386,
0.02909642644226551,
-0.04296402633190155,
0.03289412707090378,
0.12540005147457123,
-0.024612780660390854,
-0.06529714912176132,
0.02770853601396084,
0.009359552524983883,
-0.01558495033532381,
-0.001705814036540687,
-0.044044289737939835,
-0.005889843218028545,
-0.12263605743646622,
0.020631372928619385,
0.03871927037835121,
-0.052274443209171295,
-0.0545845627784729,
-0.05204750970005989,
0.007345088291913271,
-0.040224380791187286,
-0.05821821466088295,
-0.02534177526831627,
-0.07212719321250916,
0.026001423597335815,
-0.02362067997455597,
0.0022400903981179,
-0.041716355830430984,
-0.03741738200187683,
-0.00006701495294691995,
-0.04862013831734657,
6.970245740946895e-34,
-0.0009526023059152067,
-0.023134708404541016,
0.003371796803548932,
-0.005313466768711805,
-0.0008999432320706546,
-0.07348144799470901,
-0.04485657811164856,
0.028533639386296272,
-0.01674923487007618,
-0.058481305837631226,
-0.03668344020843506,
0.02211209572851658,
0.07279828190803528,
0.02481897734105587,
0.0025929459370672703,
0.003397570224478841,
0.018062081187963486,
0.04864536225795746,
-0.006695596966892481,
-0.013523492962121964,
-0.010301053524017334,
0.09270773828029633,
-0.004001027904450893,
-0.027485359460115433,
-0.026254719123244286,
0.022306902334094048,
0.020779140293598175,
0.07277997583150864,
-0.0022023150231689215,
0.09840083122253418,
-0.011054236441850662,
-0.011101027019321918,
0.000664162274915725,
-0.040648169815540314,
-0.006580552086234093,
0.08512956649065018,
0.03326137363910675,
-0.033176131546497345,
-0.049754634499549866,
0.02945500984787941,
0.08141346275806427,
-0.053464751690626144,
-0.09943170100450516,
-0.01549925934523344,
-0.017956187948584557,
0.04417097941040993,
-0.11551651358604431,
0.07894894480705261,
-0.018861671909689903,
0.07198623567819595,
-0.04967951402068138,
0.05445335432887077,
-0.02312195673584938,
-0.13426345586776733,
-0.06484807282686234,
-0.09774702787399292,
0.06212693825364113,
0.003973644692450762,
0.1403227597475052,
-0.015203891322016716,
-0.03582022339105606,
-0.054857976734638214,
0.006966899149119854,
0.058095406740903854,
-0.0242436733096838,
0.021216794848442078,
-0.011093414388597012,
-0.015640370547771454,
0.022934867069125175,
-0.03248438984155655,
0.007217837031930685,
0.038175832480192184,
-0.08502719551324844,
-0.05485346540808678,
0.05616768077015877,
-0.003217533929273486,
-0.019830098375678062,
0.030606774613261223,
-0.05937166139483452,
-0.06128941848874092,
0.016363902017474174,
-0.008527095429599285,
0.07283642143011093,
0.007315970491617918,
0.03858710452914238,
0.05553410202264786,
-0.01190655492246151,
0.11864703893661499,
0.04387735575437546,
0.06695053726434708,
-0.01282292790710926,
0.08449263870716095,
-0.0927295908331871,
0.06018347293138504,
-0.09008060395717621,
-3.622945143888501e-8,
-0.14042308926582336,
-0.0020079179666936398,
0.06811472028493881,
-0.007282693404704332,
0.05689399689435959,
0.00595487654209137,
0.02210962027311325,
0.0047801402397453785,
0.0186351016163826,
0.015999028459191322,
0.07346384227275848,
0.0034618291538208723,
-0.0012365425936877728,
-0.0180742759257555,
0.003948454279452562,
0.06228216737508774,
-0.0209158007055521,
-0.009200545959174633,
-0.015830615535378456,
0.013782145455479622,
0.11510666459798813,
-0.031077586114406586,
-0.03432924300432205,
0.05023196339607239,
0.03633372113108635,
-0.08433374762535095,
-0.0850905254483223,
0.03739434480667114,
0.02410249039530754,
0.011209719814360142,
0.042061205953359604,
0.021326210349798203,
0.04005114361643791,
-0.06957654654979706,
-0.06273120641708374,
0.06292767822742462,
0.007181253284215927,
-0.048938266932964325,
-0.045402899384498596,
-0.05578910559415817,
-0.03019767999649048,
0.09894471615552902,
-0.02606561966240406,
-0.010205194354057312,
-0.0018166936933994293,
-0.012089123018085957,
-0.07185458391904831,
-0.11258581280708313,
-0.018376855179667473,
0.040464479476213455,
-0.056935373693704605,
-0.004327649250626564,
0.03082956187427044,
0.05315573140978813,
0.13510526716709137,
-0.06823709607124329,
0.03328299522399902,
0.005214765202254057,
0.001246198546141386,
0.08287395536899567,
0.06525655090808868,
0.04571222886443138,
-0.025925029069185257,
-0.03556394949555397
] | 0 | 0 | 0 | 0 |
|
Expected Return Symmetries | https://openreview.net/forum?id=wFg0shwoRe | [
"Darius Muglich",
"Johannes Forkel",
"Elise van der Pol",
"Jakob Nicolaus Foerster"
] | Poster | Symmetry is an important inductive bias that can improve model robustness and generalization across many deep learning domains. In multi-agent settings, a priori known symmetries have been shown to address a fundamental coordination failure mode known as mutually incompatible symmetry breaking; e.g. in a game where two independent agents can choose to move "left" or "right", and where a reward of +1 or -1 is received when the agents choose the same action or different actions, respectively. However, the efficient and automatic discovery of environment symmetries, in particular for decentralized partially observable Markov decision processes, remains an open problem. Furthermore, environmental symmetry breaking constitutes only one type of coordination failure, which motivates the search for a more accessible and broader symmetry class. In this paper, we introduce such a broader group of previously unexplored symmetries, which we call expected return symmetries, which contains environment symmetries as a subgroup. We show that agents trained to be compatible under the group of expected return symmetries achieve better zero-shot coordination results than those using environment symmetries. As an additional benefit, our method makes minimal a priori assumptions about the structure of their environment and does not require access to ground truth symmetries. | multi-agent reinforcement learning, zero-shot coordination | Discovering a symmetry class over policies that improves coordination between agents | 13,880 | 2502.01711 | [
-0.05155119672417641,
-0.05863641947507858,
0.07203437387943268,
-0.05141448229551315,
0.06509362906217575,
-0.04343339800834656,
-0.023957479745149612,
-0.07426121830940247,
0.005269088316708803,
0.025737399235367775,
-0.03820593282580376,
-0.04116051644086838,
-0.010301804170012474,
0.05321193113923073,
0.05777522921562195,
-0.015873156487941742,
0.045457419008016586,
0.05498271435499191,
-0.11659624427556992,
-0.0007711887592449784,
-0.06602486968040466,
-0.035654932260513306,
0.03855555132031441,
-0.012706421315670013,
-0.052892882376909256,
0.005846655461937189,
0.009503159672021866,
-0.0565190389752388,
-0.05622905492782593,
-0.06536947190761566,
0.04499344527721405,
-0.03134523332118988,
-0.04715321958065033,
-0.11242979019880295,
-0.03720230981707573,
0.09811710566282272,
-0.06682192534208298,
-0.010375063866376877,
0.011849437840282917,
-0.03909507766366005,
0.0029683494940400124,
0.055681560188531876,
-0.004598652943968773,
-0.034098852425813675,
-0.0019164092373102903,
0.05030526965856552,
0.055893637239933014,
-0.04866665229201317,
-0.003000420518219471,
-0.045636508613824844,
0.00022633800108451396,
0.06951317936182022,
-0.08160464465618134,
0.041652340441942215,
0.009584058076143265,
0.007337919902056456,
-0.00700357323512435,
-0.018316814675927162,
-0.027539962902665138,
-0.1202293336391449,
0.061435312032699585,
-0.0991050973534584,
-0.018728584051132202,
-0.029085638001561165,
-0.00974446814507246,
0.007187284994870424,
0.03420385718345642,
0.04764380306005478,
-0.004716671537607908,
-0.0355609655380249,
0.02927372418344021,
0.03484058752655983,
-0.05495397001504898,
-0.004761959426105022,
0.06958368420600891,
0.06381876766681671,
0.008679543621838093,
-0.03829579055309296,
0.016429629176855087,
-0.04453112930059433,
-0.030241316184401512,
0.009172149002552032,
-0.00814064685255289,
0.044321268796920776,
0.14314541220664978,
-0.071575827896595,
-0.02364867553114891,
0.02835819311439991,
0.1387328952550888,
-0.02199558913707733,
-0.043262120336294174,
0.012617269530892372,
0.006422976031899452,
-0.026507964357733727,
0.07220344990491867,
0.08158110827207565,
0.0037962356582283974,
-0.035251013934612274,
-0.04052053391933441,
0.13238680362701416,
0.0029198683332651854,
0.014877182431519032,
0.03431098163127899,
-0.04168324172496796,
0.09246689826250076,
-0.011586460284888744,
-0.022669164463877678,
-0.037368640303611755,
0.06953732669353485,
-0.09728183597326279,
-0.03636588901281357,
0.02250329963862896,
0.07990650832653046,
0.04319660738110542,
-0.030816160142421722,
0.008825773373246193,
0.0024715056642889977,
0.024745158851146698,
-0.05379391089081764,
0.020101282745599747,
0.03710488975048065,
-0.01593047007918358,
0.047090500593185425,
0.07566343992948532,
0.07104454934597015,
0.046396490186452866,
-0.03922136873006821,
2.9672884511661998e-33,
0.01603580079972744,
0.03749069944024086,
0.09204840660095215,
-0.03417922928929329,
0.126930370926857,
0.007754969410598278,
0.04333069920539856,
0.0011845616390928626,
-0.01100987195968628,
0.02651493065059185,
-0.042598918080329895,
0.017415791749954224,
-0.015659712255001068,
0.010263660922646523,
-0.03396853059530258,
0.0437820702791214,
-0.05152532085776329,
0.010199731215834618,
0.014417658559978008,
-0.07689874619245529,
0.044660650193691254,
-0.00349202542565763,
-0.029230354353785515,
-0.03851700574159622,
-0.03008541464805603,
0.013410793617367744,
0.05587621033191681,
0.023141173645853996,
0.016249310225248337,
-0.01442534476518631,
0.03636735677719116,
0.015094277448952198,
-0.02085035853087902,
0.12889045476913452,
0.02106981724500656,
-0.022998696193099022,
-0.09576268494129181,
0.03030470944941044,
0.02683379501104355,
-0.045961104333400726,
-0.027263445779681206,
0.03275773301720619,
-0.035804614424705505,
-0.0006700915982946754,
0.03269178420305252,
-0.09861066192388535,
0.0503048337996006,
-0.012690394185483456,
-0.0725954994559288,
-0.05418233200907707,
0.047399409115314484,
-0.03993711248040199,
0.03579513728618622,
-0.06853076070547104,
-0.04750247299671173,
0.03271155431866646,
0.06383371353149414,
0.07274836301803589,
-0.07651066035032272,
0.03905530273914337,
0.03414527699351311,
0.02785944938659668,
-0.03525584191083908,
0.022482020780444145,
-0.05236486345529556,
0.03446860611438751,
0.0010979644721373916,
0.0354374498128891,
0.09568790346384048,
-0.037101227790117264,
0.0477696917951107,
0.03669395670294762,
-0.0868111401796341,
-0.03707265481352806,
0.08772055804729462,
-0.011756567284464836,
0.08671951293945312,
-0.050803784281015396,
0.00014762725913897157,
-0.00007870687841204926,
-0.02969939075410366,
0.034549493342638016,
-0.0787450522184372,
0.037507522851228714,
-0.06376439332962036,
-0.007416012231260538,
-0.049186721444129944,
-0.18115316331386566,
0.018428310751914978,
0.02398211881518364,
-0.04858650639653206,
-0.03173224255442619,
-0.06077950447797775,
0.03760749101638794,
0.03184651583433151,
-3.5645951507144135e-33,
-0.024200839921832085,
-0.11367658525705338,
0.030911613255739212,
0.03457869216799736,
-0.02845006063580513,
-0.005133436061441898,
-0.0100697698071599,
0.010486895218491554,
0.006661437451839447,
-0.01535284984856844,
-0.03625839948654175,
0.05576613172888756,
-0.014202079735696316,
-0.035443518310785294,
0.018192438408732414,
-0.033015817403793335,
0.058663323521614075,
0.029805250465869904,
0.023721789941191673,
0.02123279683291912,
0.0516182966530323,
0.08280237764120102,
-0.10145517438650131,
0.07279606908559799,
0.09084045886993408,
0.08997455984354019,
-0.004509849939495325,
-0.008675405755639076,
0.11640623956918716,
0.04254605621099472,
-0.071166031062603,
-0.034824926406145096,
-0.04400058090686798,
0.03307504951953888,
-0.012837476097047329,
0.054870761930942535,
-0.0275376308709383,
0.056896574795246124,
-0.10892000049352646,
0.1433766633272171,
-0.016847744584083557,
-0.07438111305236816,
-0.033075250685214996,
0.08291676640510559,
0.026011763140559196,
-0.0747080147266388,
0.03030533343553543,
0.052600715309381485,
-0.013996955938637257,
0.0076025002636015415,
-0.022477561607956886,
-0.01619589328765869,
-0.07630491256713867,
0.009748312644660473,
-0.01213352382183075,
0.047804515808820724,
-0.006231005769222975,
-0.003290637396275997,
0.07810696959495544,
0.00774387689307332,
-0.06728722155094147,
-0.016739150509238243,
-0.0037979038897901773,
0.06985148042440414,
-0.04656846821308136,
0.03725147619843483,
-0.030858775600790977,
-0.00016548005805816501,
0.06740966439247131,
0.004037196282297373,
0.04602168872952461,
-0.025793706998229027,
-0.03037913702428341,
-0.11416579782962799,
0.10780248790979385,
0.08151089400053024,
-0.0294656865298748,
-0.032306283712387085,
-0.03433557599782944,
-0.024865401908755302,
-0.028176365420222282,
-0.007100589107722044,
-0.006947655696421862,
0.04344252124428749,
0.036581333726644516,
-0.027151871472597122,
0.08740504086017609,
0.03176339343190193,
0.00793792586773634,
-0.0394112728536129,
-0.014557118527591228,
0.03291965648531914,
0.03266735002398491,
0.10791090875864029,
-0.0728350356221199,
-5.085167131824164e-8,
-0.025817614048719406,
0.04145051911473274,
0.055716030299663544,
0.0006777436938136816,
0.011936287395656109,
0.012441499158740044,
-0.004999799188226461,
0.010859319940209389,
-0.07818887382745743,
0.01994326524436474,
0.03829874470829964,
-0.010092130862176418,
0.008556477725505829,
-0.039202868938446045,
-0.04010062664747238,
0.05202600359916687,
-0.01313733123242855,
-0.0042823008261621,
-0.07435716688632965,
0.0071880812756717205,
0.005383635871112347,
-0.02900380641222,
-0.020871294662356377,
-0.008946060203015804,
0.0027628212701529264,
-0.07866991311311722,
-0.09854217618703842,
-0.057009369134902954,
0.020818021148443222,
0.051590025424957275,
0.012961119413375854,
-0.06080038100481033,
0.08544205874204636,
-0.015949847176671028,
-0.008326496928930283,
0.09113423526287079,
-0.020064910873770714,
-0.027121512219309807,
0.01594332791864872,
-0.13848473131656647,
-0.022876646369695663,
0.1139901727437973,
0.002597726648673415,
0.019089212641119957,
0.05007389932870865,
0.024521086364984512,
0.010492299683392048,
-0.0617789700627327,
0.04961423575878143,
-0.027438731864094734,
-0.021836260333657265,
0.00046380763524211943,
-0.03388868272304535,
0.019459111616015434,
0.007249496877193451,
0.06551584601402283,
0.00966411828994751,
-0.09398109465837479,
0.05137837305665016,
0.05958866700530052,
0.00593637116253376,
-0.008136439137160778,
-0.09487772732973099,
-0.006920903921127319
] | 0 | 0 | 0 | 0 |
|
Beware of Calibration Data for Pruning Large Language Models | https://openreview.net/forum?id=x83w6yGIWb | [
"Yixin Ji",
"Yang Xiang",
"Juntao Li",
"Qingrong Xia",
"Ping Li",
"Xinyu Duan",
"Zhefeng Wang",
"Min Zhang"
] | Poster | As large language models (LLMs) are widely applied across various fields, model
compression has become increasingly crucial for reducing costs and improving
inference efficiency. Post-training pruning is a promising method that does not
require resource-intensive iterative training and only needs a small amount of
calibration data to assess the importance of parameters. Recent research has enhanced post-training pruning from different aspects but few of them systematically
explore the effects of calibration data, and it is unclear if there exist better calibration data construction strategies. We fill this blank and surprisingly observe that
calibration data is also crucial to post-training pruning, especially for high sparsity. Through controlled experiments on important influence factors of calibration
data, including the pruning settings, the amount of data, and its similarity with
pre-training data, we observe that a small size of data is adequate, and more similar data to its pre-training stage can yield better performance. As pre-training data
is usually inaccessible for advanced LLMs, we further provide a self-generating
calibration data synthesis strategy to construct feasible calibration data. Experimental results on recent strong open-source LLMs (e.g., DCLM, and LLaMA-3)
show that the proposed strategy can enhance the performance of strong pruning
methods (e.g., Wanda, DSnoT, OWL) by a large margin (up to 2.68%). | calibration data, post-training pruning, large language models | null | 13,874 | 2410.17711 | [
-0.014695784077048302,
-0.03532988578081131,
-0.044916294515132904,
0.006007678806781769,
0.037628207355737686,
0.04330733045935631,
-0.03254382684826851,
-0.0353672094643116,
0.030084002763032913,
-0.06575263291597366,
0.015352259390056133,
-0.04550294950604439,
0.05932916700839996,
0.03487304970622063,
0.0056715658865869045,
-0.06761126220226288,
0.019587215036153793,
0.09212469309568405,
-0.09197953343391418,
-0.12636961042881012,
-0.017158934846520424,
0.04551178216934204,
0.012260773219168186,
0.04633362963795662,
0.09802969545125961,
-0.02414461225271225,
-0.04789743945002556,
-0.040255408734083176,
-0.02285495027899742,
-0.03312969207763672,
0.06414494663476944,
-0.021748552098870277,
0.07952220737934113,
0.019292332231998444,
0.0005326823447830975,
0.0173606239259243,
0.0058108433149755,
-0.028237711638212204,
0.0018711986485868692,
-0.005995492916554213,
-0.0427989587187767,
0.0012173227733001113,
0.04712051898241043,
0.06459751725196838,
0.01947074942290783,
0.06213212013244629,
0.02639191225171089,
-0.031217362731695175,
-0.0410039946436882,
0.03014412708580494,
-0.057627588510513306,
0.021667219698429108,
-0.0074099707417190075,
0.052555352449417114,
-0.015192491933703423,
0.003948941361159086,
0.0010332567617297173,
0.055560626089572906,
-0.04500322416424751,
-0.029598549008369446,
-0.09123309701681137,
-0.06179273873567581,
-0.06579966843128204,
0.01309167966246605,
-0.026022354140877724,
-0.012705056928098202,
0.09160517901182175,
0.04049142450094223,
0.020411964505910873,
0.0684395283460617,
-0.007974151521921158,
0.08472461998462677,
0.014491799287497997,
0.12406115978956223,
0.02978476509451866,
0.07922356575727463,
0.0065197390504181385,
0.028842192143201828,
0.07509913295507431,
-0.06645180284976959,
-0.05034082382917404,
-0.00913463905453682,
0.024149728938937187,
0.02742965891957283,
0.023357568308711052,
-0.01408936083316803,
0.04677843302488327,
0.02135816402733326,
0.043194688856601715,
-0.013542449101805687,
-0.05663967877626419,
0.01886400394141674,
-0.02315453626215458,
0.03522466868162155,
-0.08532416820526123,
0.06959416717290878,
0.0031651565805077553,
-0.03178552910685539,
-0.00833046156913042,
0.08518635481595993,
-0.009014066308736801,
0.08129942417144775,
0.07347483932971954,
-0.004925695713609457,
0.0013734300155192614,
-0.06820196658372879,
0.03533010929822922,
0.05791420117020607,
-0.016747871413826942,
-0.012829907238483429,
0.0516941137611866,
0.007773654069751501,
-0.02392866089940071,
-0.0071637374348938465,
0.07948405295610428,
0.020816821604967117,
-0.0404084175825119,
-0.08238477259874344,
-0.004340856336057186,
0.038512226194143295,
-0.06609801203012466,
0.034984033554792404,
-0.013856260105967522,
0.028930624946951866,
0.022574611008167267,
-0.018434330821037292,
-0.04908769577741623,
4.590660659660172e-33,
0.09350442886352539,
0.13470926880836487,
0.03694118559360504,
0.01306262519210577,
0.01731392927467823,
-0.026106733828783035,
0.029475780203938484,
0.07612889260053635,
0.004346327390521765,
0.004317719954997301,
-0.05900182947516441,
-0.000925008556805551,
-0.03652898967266083,
0.03293699398636818,
0.06121315807104111,
0.009688890539109707,
-0.014409814029932022,
0.0756220668554306,
-0.09450806677341461,
-0.0026513258926570415,
0.07513925433158875,
0.0072958641685545444,
0.05026240274310112,
-0.09972626715898514,
0.01540308352559805,
0.05778825655579567,
0.07084660977125168,
0.0032372691202908754,
-0.14537692070007324,
0.01742623560130596,
-0.03347289189696312,
-0.03250342607498169,
0.007778913713991642,
0.00405610678717494,
0.050484757870435715,
-0.04157609865069389,
0.0013260197592899203,
-0.01147659495472908,
0.013878806494176388,
0.0437006950378418,
-0.0034977069590240717,
0.08323697745800018,
0.04905276000499725,
0.00036858164821751416,
-0.039076726883649826,
-0.06339535862207413,
0.023878159001469612,
-0.013089042156934738,
0.002717499155551195,
-0.0418703556060791,
0.07716554403305054,
-0.0029404363594949245,
-0.005311788525432348,
0.0008949306793510914,
-0.08132284879684448,
-0.01787884347140789,
0.0764961838722229,
-0.027183538302779198,
0.03315100446343422,
0.016047293320298195,
-0.047507502138614655,
-0.000319305487209931,
0.013846142217516899,
0.09009203314781189,
0.033694565296173096,
-0.03359243646264076,
-0.0454661101102829,
0.06045473366975784,
0.060776591300964355,
-0.0721948891878128,
-0.05915478989481926,
-0.06308498233556747,
-0.02834608405828476,
-0.05297749489545822,
0.034630727022886276,
-0.023790432140231133,
0.09330937266349792,
-0.10550987720489502,
0.013976807706058025,
-0.0035929610021412373,
0.0009966702200472355,
0.01980055496096611,
0.03882584348320961,
-0.09567300230264664,
-0.02713051810860634,
-0.08210758864879608,
0.05165529623627663,
-0.05759493261575699,
0.0130190160125494,
-0.021891729906201363,
-0.05179383605718613,
-0.03500663861632347,
-0.05883457511663437,
-0.06959326565265656,
0.03023107536137104,
-3.7199400352534405e-33,
-0.04015732556581497,
0.002080857753753662,
-0.045528605580329895,
0.11645301431417465,
-0.04616451635956764,
-0.04393117129802704,
0.07148241251707077,
0.11877832561731339,
-0.049561869353055954,
-0.07283376902341843,
-0.03936440870165825,
-0.08436205238103867,
0.09412720799446106,
-0.02392718195915222,
-0.012839029543101788,
0.022091902792453766,
-0.01881047710776329,
0.0022462974302470684,
0.039147429168224335,
0.02291995845735073,
-0.017412537708878517,
0.09448913484811783,
-0.07512198388576508,
-0.03316449001431465,
-0.00020889689039904624,
-0.01605619676411152,
-0.042305998504161835,
0.0031682788394391537,
0.005101946648210287,
0.0038213853258639574,
-0.022339453920722008,
-0.04033605009317398,
-0.028033118695020676,
-0.050387557595968246,
-0.1075991615653038,
-0.023680368438363075,
0.011378149501979351,
0.01290991809219122,
0.041642047464847565,
0.12879905104637146,
0.05005823075771332,
-0.006435323040932417,
-0.10075607895851135,
0.006318727973848581,
0.002156489295884967,
-0.052480727434158325,
-0.05551851913332939,
-0.06046512350440025,
0.045798156410455704,
0.058747436851263046,
0.005210917443037033,
0.07337533682584763,
-0.03020480088889599,
0.0016559067880734801,
-0.015299489721655846,
-0.10165412724018097,
-0.06665773689746857,
-0.022237394005060196,
0.011199813336133957,
-0.0177965946495533,
-0.05418771505355835,
-0.02027377113699913,
0.07195638120174408,
-0.04184498265385628,
0.048333898186683655,
-0.08666634559631348,
0.00003364462827448733,
0.03189381957054138,
0.04432268440723419,
-0.04953492432832718,
-0.0062094139866530895,
-0.040570396929979324,
-0.020854486152529716,
0.061761800199747086,
-0.03407064080238342,
-0.019174426794052124,
0.008765531703829765,
0.015487495809793472,
-0.016136063262820244,
-0.02528037130832672,
0.019433077424764633,
0.0170154832303524,
0.00811660848557949,
0.07154836505651474,
0.09262558072805405,
0.09914271533489227,
0.04338371381163597,
0.04567408561706543,
0.002839644905179739,
0.07929955422878265,
-0.055241819471120834,
0.07365497946739197,
0.06577255576848984,
0.10333086550235748,
-0.047846049070358276,
-4.267361930487823e-8,
-0.11657095700502396,
0.047806739807128906,
-0.018464157357811928,
0.1346563994884491,
-0.01063267607241869,
-0.016651831567287445,
-0.0676732286810875,
0.10849549621343613,
0.027452129870653152,
-0.07707442343235016,
0.07736968249082565,
-0.005273160990327597,
-0.0730176568031311,
-0.009275820106267929,
-0.015161647461354733,
0.045631516724824905,
0.023826394230127335,
0.024598877876996994,
-0.0599222294986248,
0.03363099694252014,
0.0017573230434209108,
0.07062966376543045,
-0.0364406555891037,
-0.03327149897813797,
0.06619665026664734,
-0.033005569130182266,
-0.048742566257715225,
0.15062381327152252,
0.05390239506959915,
-0.027199039235711098,
-0.032790545374155045,
-0.052586328238248825,
-0.021590765565633774,
-0.10143189877271652,
0.01407530065625906,
0.10548513382673264,
0.029871387407183647,
0.0231664776802063,
-0.0038819164037704468,
-0.04383617639541626,
-0.001992088509723544,
-0.02455255575478077,
-0.08539640158414841,
0.02693348564207554,
0.046545691788196564,
-0.052948374301195145,
-0.03247188404202461,
-0.07287349551916122,
-0.025582291185855865,
-0.06416621059179306,
0.024764837697148323,
-0.010969884693622589,
-0.048107702285051346,
-0.008072824217379093,
0.0022050754632800817,
0.03903646767139435,
-0.024823609739542007,
0.02937970869243145,
-0.015014193020761013,
0.032119013369083405,
0.014759146608412266,
-0.05930854380130768,
-0.10207276791334152,
-0.016960328444838524
] | 0 | 0 | 0 | 0 |
|
Herald: A Natural Language Annotated Lean 4 Dataset | https://openreview.net/forum?id=Se6MgCtRhz | [
"Guoxiong Gao",
"Yutong Wang",
"Jiedong Jiang",
"Qi Gao",
"Zihan Qin",
"Tianyi Xu",
"Bin Dong"
] | Poster | Verifiable formal languages like Lean have profoundly impacted mathematical reasoning, particularly through the use of large language models (LLMs) for automated reasoning. A significant challenge in training LLMs for these formal languages is the lack of parallel datasets that align natural language with formal language proofs. To address this challenge, this paper introduces a novel framework for translating the Mathlib4 corpus (a unified library of mathematics in formal language Lean 4) into natural language. Building upon this, we employ a dual augmentation strategy that combines tactic-based and informal-based approaches, leveraging the Lean-jixia system, a Lean 4 analyzer. We present the results of this pipeline on Mathlib4 as Herald (Hierarchy and Retrieval-based Translated Lean Dataset). We also propose the Herald Translator, which is fine-tuned on Herald. Herald translator achieves a 96.7\% accuracy (Pass@128) on formalizing statements in the miniF2F-test and a 23.5\% accuracy on our internal graduate-level textbook dataset, outperforming InternLM2-Math-Plus-7B (73.0\% and 7.5\%) and TheoremLlama (50.1\% and 4.0\%). Furthermore, we propose a section-level translation framework for real-world applications. As a direct application of Herald translator, we have successfully translated a template section in the Stack project, marking a notable progress in the automatic formalization of graduate-level mathematical literature. Our model, along with the datasets, are open-sourced to the public. | Lean 4, Autoformalizing, LLM, Retrieval Augmented Generation, Dataset | null | 13,870 | 2410.10878 | [
-0.05206345394253731,
-0.004329629708081484,
0.052246689796447754,
-0.012671319767832756,
0.021828528493642807,
-0.06889858096837997,
-0.04990352317690849,
0.059465959668159485,
-0.0626254752278328,
-0.052407339215278625,
-0.04583446681499481,
-0.10143531113862991,
0.03032338246703148,
0.02021639049053192,
0.050453536212444305,
0.04687491059303284,
0.002388766035437584,
0.08447488397359848,
-0.03130735456943512,
-0.08404070138931274,
-0.028347378596663475,
-0.015045775100588799,
0.0290367491543293,
-0.0005654536071233451,
0.052562639117240906,
-0.029192911460995674,
-0.017469896003603935,
-0.06555747985839844,
0.0008174472604878247,
-0.024241749197244644,
-0.02924937754869461,
0.02868834137916565,
-0.05605069920420647,
-0.004314880818128586,
0.021760955452919006,
0.09692524373531342,
0.0419103167951107,
0.017382895573973656,
-0.02924383245408535,
-0.03275587037205696,
-0.06230921298265457,
-0.04460729658603668,
0.08212822675704956,
-0.004403493367135525,
0.005745186470448971,
0.002085319487378001,
-0.0018829414620995522,
-0.0014393024612218142,
0.009382031857967377,
0.00835819449275732,
-0.11815652251243591,
-0.04847197234630585,
-0.005301847588270903,
0.03250714763998985,
0.007735383231192827,
0.0469818152487278,
0.08897644281387329,
0.002632339019328356,
0.009652115404605865,
-0.024480769410729408,
-0.025225944817066193,
-0.06260263174772263,
-0.019858628511428833,
0.013515852391719818,
-0.037719812244176865,
-0.007792280521243811,
0.0033411746844649315,
0.08475030213594437,
-0.025799497961997986,
0.05187921226024628,
-0.012214360758662224,
0.014419704675674438,
-0.051982417702674866,
0.04844517260789871,
0.0029802722856402397,
0.04361611232161522,
-0.00927602220326662,
-0.024385398253798485,
0.09465978294610977,
-0.02425210364162922,
-0.10133424401283264,
-0.031684283167123795,
0.02571735344827175,
0.030722567811608315,
0.09181730449199677,
-0.05030030757188797,
-0.007641106378287077,
-0.012609586119651794,
0.0010231786873191595,
0.09967052191495895,
0.05040090158581734,
0.006280597299337387,
0.05931459367275238,
0.07976992428302765,
0.049307700246572495,
0.075938880443573,
0.015426389873027802,
-0.03513799607753754,
0.03120223805308342,
0.1288665235042572,
-0.019109291955828667,
0.15906837582588196,
0.01625920459628105,
-0.08395431190729141,
-0.010838976129889488,
-0.05426878109574318,
0.06521192938089371,
-0.025606080889701843,
0.006194423884153366,
-0.10962275415658951,
0.08102307468652725,
-0.016089899465441704,
0.07363183796405792,
0.005360141862183809,
0.019065195694565773,
-0.022826069965958595,
-0.018868418410420418,
-0.014399499632418156,
-0.04434489831328392,
-0.016189448535442352,
-0.010413133539259434,
0.14116550981998444,
0.08529726415872574,
-0.04659629613161087,
0.10746580362319946,
0.03418741375207901,
-0.07262752950191498,
1.1032288152245719e-33,
0.08058161288499832,
0.10813484340906143,
0.0442499965429306,
0.04407166689634323,
0.00826247874647379,
-0.06487783789634705,
0.01309953723102808,
0.04209575429558754,
-0.08724340796470642,
-0.007071684114634991,
0.061226412653923035,
0.0394524447619915,
-0.06905410438776016,
0.06402534991502762,
0.015540806576609612,
-0.06749134510755539,
0.034158892929553986,
0.00997082144021988,
-0.0348689965903759,
0.02698425203561783,
0.026654623448848724,
-0.037586819380521774,
-0.06717076897621155,
-0.0296016875654459,
0.050105754286050797,
0.06746302545070648,
0.07651922851800919,
0.006309429183602333,
-0.06908630579710007,
0.029720742255449295,
-0.10341638326644897,
-0.07508765161037445,
-0.008408850058913231,
0.02437727339565754,
0.04276901111006737,
-0.14597484469413757,
-0.0010272754589095712,
-0.07837291806936264,
0.016519077122211456,
0.016004187986254692,
-0.03351415693759918,
-0.0018384420545771718,
-0.011530124582350254,
0.02987593226134777,
-0.030936820432543755,
0.011168534867465496,
0.011412374675273895,
0.05176549032330513,
0.12464995682239532,
-0.07086329907178879,
0.06214452162384987,
-0.029353901743888855,
-0.007203560322523117,
-0.004503623116761446,
0.04916851222515106,
0.018343478441238403,
-0.05818495154380798,
0.035479702055454254,
-0.0016291369684040546,
0.036813315004110336,
0.04989999532699585,
0.11232636868953705,
0.03581971675157547,
0.03430704399943352,
-0.013890828937292099,
0.02927778847515583,
-0.07693822681903839,
0.022652138024568558,
0.07814118266105652,
0.020864766091108322,
-0.019803082570433617,
-0.012340247631072998,
0.016866104677319527,
0.028472181409597397,
0.13350166380405426,
-0.04134861379861832,
0.05221937596797943,
-0.0753537267446518,
0.03783241659402847,
-0.037118107080459595,
0.04856115207076073,
0.018775533884763718,
0.05519506707787514,
-0.01531224511563778,
0.017611289396882057,
-0.06111450493335724,
0.0351838618516922,
-0.04698316752910614,
-0.022836921736598015,
-0.00819964986294508,
-0.07107923924922943,
-0.03848225250840187,
-0.06690824031829834,
-0.018191616982221603,
-0.08053702861070633,
-1.7630128558479979e-34,
-0.007908055558800697,
-0.011511271819472313,
-0.04236895218491554,
0.06320985406637192,
-0.06497922539710999,
-0.038018565624952316,
0.01619873382151127,
0.009357215836644173,
0.06889613717794418,
-0.03337075561285019,
-0.0034313853830099106,
-0.07992836087942123,
0.020876457914710045,
-0.0064673409797251225,
0.04499330744147301,
-0.017541654407978058,
-0.05302990600466728,
-0.07157811522483826,
-0.029752960428595543,
0.07833104580640793,
-0.025318684056401253,
0.04139174148440361,
-0.06696967035531998,
0.017786327749490738,
-0.0024807152803987265,
0.02069309912621975,
-0.02982250228524208,
0.05439561977982521,
-0.03417665883898735,
-0.010179976932704449,
0.03291942551732063,
-0.11722870171070099,
-0.031044194474816322,
-0.06712944805622101,
-0.092501699924469,
-0.02468143031001091,
0.12475189566612244,
-0.026392774656414986,
-0.0639675110578537,
0.02278674580156803,
0.07739581912755966,
-0.029533294960856438,
-0.02241240069270134,
-0.016363639384508133,
-0.05543157458305359,
0.014823664911091328,
-0.07890412956476212,
0.043076369911432266,
-0.035276517271995544,
-0.024960197508335114,
0.00033388557494618,
-0.04220632091164589,
0.023960351943969727,
0.014504510909318924,
0.010391916148364544,
-0.11779078096151352,
0.030501067638397217,
-0.04582763835787773,
-0.06822748482227325,
0.02666517160832882,
-0.058941226452589035,
-0.021073544397950172,
0.04514190927147865,
0.012904025614261627,
0.04527207463979721,
-0.04248853772878647,
-0.08784765005111694,
-0.009138542227447033,
-0.06801358610391617,
0.007268398068845272,
0.04000481590628624,
-0.05393047258257866,
-0.010532763786613941,
0.022596746683120728,
0.04179370030760765,
0.0608392134308815,
0.029798218980431557,
-0.03428134322166443,
0.008116215467453003,
-0.031630296260118484,
0.023399733006954193,
-0.024193942546844482,
0.08151650428771973,
-0.00007008295506238937,
0.008653197437524796,
-0.0425252765417099,
0.07887903600931168,
0.08655337989330292,
0.005154481157660484,
0.09053591638803482,
-0.03325895965099335,
0.035090479999780655,
-0.036100953817367554,
0.07165371626615524,
-0.031018493697047234,
-4.183280566394387e-8,
-0.009007975459098816,
0.08486427366733551,
0.012396798469126225,
0.008457757532596588,
-0.02791672945022583,
-0.019009895622730255,
-0.061699654906988144,
0.014157041907310486,
-0.037131596356630325,
-0.0026075744535773993,
0.03811958059668541,
0.041762471199035645,
-0.05101153254508972,
-0.07539181411266327,
-0.03394312039017677,
0.06645388901233673,
0.019778955727815628,
-0.019637180492281914,
-0.06469248980283737,
-0.031235741451382637,
0.071772001683712,
0.0638967975974083,
-0.06374631077051163,
0.01368979923427105,
0.01206847745925188,
-0.04251119866967201,
0.04917876049876213,
0.12493371218442917,
0.07406504452228546,
-0.062476254999637604,
-0.06199919432401657,
0.008623088710010052,
0.019035054370760918,
-0.10148347914218903,
0.022869396954774857,
0.061469245702028275,
-0.0001461587962694466,
0.005244819913059473,
0.05665475130081177,
0.036945607513189316,
0.0029066281858831644,
0.031871046870946884,
-0.08936811983585358,
0.008650971576571465,
0.052547238767147064,
-0.07833515107631683,
-0.07964511215686798,
-0.012227092869579792,
0.07146208733320236,
-0.024217383936047554,
-0.009836879558861256,
-0.03782429173588753,
-0.015742924064397812,
0.050529301166534424,
0.010029282420873642,
0.057346418499946594,
-0.06828711181879044,
0.02416156977415085,
-0.0320185162127018,
0.03095262497663498,
0.02161809429526329,
-0.028605179861187935,
0.04476328194141388,
0.014094159938395023
] | 0 | 0 | 0 | 0 |
|
Efficient Residual Learning with Mixture-of-Experts for Universal Dexterous Grasping | https://openreview.net/forum?id=BUj9VSCoET | [
"Ziye Huang",
"Haoqi Yuan",
"Yuhui Fu",
"Zongqing Lu"
] | Poster | Universal dexterous grasping across diverse objects presents a fundamental yet formidable challenge in robot learning. Existing approaches using reinforcement learning (RL) to develop policies on extensive object datasets face critical limitations, including complex curriculum design for multi-task learning and limited generalization to unseen objects.
To overcome these challenges, we introduce ResDex, a novel approach that integrates residual policy learning with a mixture-of-experts (MoE) framework. ResDex is distinguished by its use of geometry-agnostic base policies that are efficiently acquired on individual objects and capable of generalizing across a wide range of unseen objects. Our MoE framework incorporates several base policies to facilitate diverse grasping styles suitable for various objects. By learning residual actions alongside weights that combine these base policies, ResDex enables efficient multi-task RL for universal dexterous grasping.
ResDex achieves state-of-the-art performance on the DexGraspNet dataset comprising 3,200 objects with an 88.8% success rate. It exhibits no generalization gap with unseen objects and demonstrates superior training efficiency, mastering all tasks within only 12 hours on a single GPU. For further details and videos, visit our project page. | dexterous grasping, residual policy learning, reinforcement learning | null | 13,867 | 2410.02475 | [
-0.03011941909790039,
-0.06614486128091812,
0.052900899201631546,
0.03537767380475998,
-0.005608356557786465,
-0.012552088126540184,
0.021087005734443665,
-0.07804329693317413,
-0.062832310795784,
0.04146459698677063,
-0.017951996996998787,
-0.04411555081605911,
-0.03563821688294411,
0.06706181913614273,
0.024040548130869865,
0.030466046184301376,
0.10206150263547897,
0.048822786659002304,
-0.121601402759552,
0.045339327305555344,
0.029921218752861023,
0.01450375746935606,
0.09362533688545227,
-0.00561575498431921,
-0.03310186043381691,
0.06275437772274017,
0.021510984748601913,
-0.13038690388202667,
0.03857054188847542,
-0.05125919729471207,
0.00302450405433774,
-0.05318295583128929,
-0.01694669760763645,
-0.042399682104587555,
-0.07514480501413345,
0.09488937258720398,
-0.09456974267959595,
0.004629358649253845,
0.001191412564367056,
-0.035202641040086746,
-0.028912505134940147,
0.07601337134838104,
0.052310023456811905,
0.07427743822336197,
0.07643923163414001,
0.06177971512079239,
-0.033300939947366714,
-0.0545344203710556,
0.15378615260124207,
-0.0005726420204155147,
0.02909594587981701,
-0.07175169140100479,
0.05185089632868767,
0.04360048472881317,
0.0025435711722820997,
0.10542764514684677,
-0.012914528138935566,
-0.05730419233441353,
0.042887188494205475,
-0.005053012631833553,
-0.007481206674128771,
-0.0507500022649765,
-0.020450158044695854,
-0.02955794334411621,
-0.04532919079065323,
-0.06340342015028,
0.0778491348028183,
-0.0060882288962602615,
-0.00865845289081335,
0.017027679830789566,
0.010585951618850231,
0.05520370602607727,
0.0015363921411335468,
0.015611745417118073,
0.08351218700408936,
0.02012772485613823,
-0.01937401480972767,
-0.013634920120239258,
0.00006486219353973866,
-0.004778906237334013,
-0.008806824684143066,
0.022928716614842415,
0.04729306697845459,
0.03197401762008667,
0.11718840897083282,
-0.033313293009996414,
-0.024555297568440437,
0.08362502604722977,
0.011664915829896927,
0.005401731468737125,
-0.004423985257744789,
0.021848050877451897,
0.03151837736368179,
-0.02796592190861702,
-0.008316095918416977,
0.0030332168098539114,
-0.035321760922670364,
-0.10645352303981781,
-0.11231325566768646,
0.05750378221273422,
-0.007533403113484383,
-0.020806511864066124,
0.011137654073536396,
0.031117532402276993,
-0.018494630232453346,
-0.019540397450327873,
-0.05238525941967964,
-0.058965329080820084,
-0.01501265075057745,
-0.010508106090128422,
0.043324459344148636,
-0.08202652633190155,
0.019032688811421394,
0.01669357530772686,
-0.049984410405159,
0.0031377493869513273,
-0.06694619357585907,
-0.03284713253378868,
-0.03125952184200287,
0.04739861190319061,
-0.061179231852293015,
-0.031875573098659515,
0.043538760393857956,
0.02731052041053772,
0.0594351664185524,
-0.04225117713212967,
-0.04253716394305229,
1.7910258045830546e-33,
0.03522801026701927,
-0.002368955174461007,
0.07030344009399414,
-0.04570915177464485,
0.007642596028745174,
-0.025789804756641388,
0.06546631455421448,
0.04073444381356239,
-0.05910513922572136,
0.015081176534295082,
0.006786772515624762,
0.011640580371022224,
-0.07708238065242767,
0.10881992429494858,
0.05593571439385414,
-0.004431195557117462,
-0.015528646297752857,
0.08164278417825699,
-0.04216863960027695,
0.06521298736333847,
0.015984617173671722,
0.013419337570667267,
-0.034687694162130356,
-0.08207130432128906,
-0.01015013549476862,
-0.02405274100601673,
0.021383710205554962,
0.022201452404260635,
0.06037699803709984,
0.016137676313519478,
-0.07499143481254578,
-0.009276378899812698,
-0.053195126354694366,
0.13899460434913635,
0.005174442194402218,
0.011624855920672417,
-0.015046385116875172,
-0.05948209390044212,
-0.018281901255249977,
-0.023361559957265854,
0.011947862803936005,
0.04293764755129814,
0.05851129814982414,
-0.00646594911813736,
-0.06900054961442947,
-0.07099087536334991,
0.06482142210006714,
0.04055578634142876,
-0.04922131448984146,
-0.01039286982268095,
0.01601252518594265,
0.044645898044109344,
-0.036765556782484055,
-0.09816712886095047,
-0.07239628583192825,
0.004584577865898609,
-0.021366795524954796,
0.11422364413738251,
-0.005482942797243595,
-0.03908412531018257,
0.056925032287836075,
-0.0029867247212678194,
-0.06367793679237366,
0.09949585795402527,
-0.025214804336428642,
0.08551590889692307,
-0.007654566317796707,
0.06703443825244904,
0.08217776566743851,
-0.011305552907288074,
-0.021254384890198708,
-0.021139351651072502,
0.04622107371687889,
-0.055405765771865845,
0.06054593622684479,
0.042926207184791565,
0.04833770915865898,
-0.08115766942501068,
-0.011402703821659088,
0.03459581732749939,
-0.01885317638516426,
0.08210296928882599,
-0.0185407567769289,
-0.02721891924738884,
-0.05183912441134453,
-0.02359256148338318,
0.09018128365278244,
-0.05391918122768402,
-0.02211093157529831,
0.03246641159057617,
-0.10684777796268463,
-0.08780454099178314,
-0.10311497002840042,
0.012738021090626717,
-0.014428411610424519,
-2.132050491081024e-33,
0.061228640377521515,
0.013658557087182999,
-0.008948792703449726,
0.053919170051813126,
-0.02711944840848446,
-0.02923387475311756,
-0.025383517146110535,
-0.08760172873735428,
-0.029170216992497444,
-0.11837681382894516,
-0.026872579008340836,
-0.006861042231321335,
0.017884178087115288,
-0.025668172165751457,
0.0004080737999174744,
0.036657337099313736,
-0.0814947783946991,
0.026410315185785294,
-0.030984632670879364,
-0.04039348289370537,
0.06538717448711395,
0.04761986806988716,
-0.003178191604092717,
0.05507885292172432,
-0.09390474110841751,
-0.03892211243510246,
-0.056222740560770035,
0.042539529502391815,
-0.014504645951092243,
0.0821104496717453,
-0.030116897076368332,
-0.019071286544203758,
-0.030234307050704956,
0.00611162930727005,
-0.050255186855793,
0.05818641185760498,
0.032845985144376755,
0.06324589997529984,
0.00644097663462162,
0.08773794025182724,
0.004820633679628372,
-0.11361812800168991,
0.012852768413722515,
0.07542651891708374,
0.030447592958807945,
-0.012714101001620293,
-0.05792167782783508,
-0.005246696528047323,
-0.048633504658937454,
0.03981420770287514,
-0.07881277054548264,
0.056440163403749466,
-0.11681003123521805,
-0.08804783225059509,
-0.04097531735897064,
-0.06058947741985321,
-0.04167189076542854,
-0.031020572409033775,
0.1046077162027359,
-0.02480522356927395,
-0.009203356690704823,
0.012120385654270649,
-0.03631914407014847,
0.11091281473636627,
0.0038849569391459227,
0.03960396721959114,
0.008134089410305023,
-0.03208635747432709,
0.012509127147495747,
0.020019426941871643,
0.022087860852479935,
0.012301660142838955,
0.060973599553108215,
-0.05296815559267998,
0.04762522503733635,
0.030948100611567497,
0.022773249074816704,
-0.07160526514053345,
0.010689684189856052,
0.021302996203303337,
0.04345051944255829,
0.011099442839622498,
0.01773042045533657,
0.11797559261322021,
0.05871877074241638,
0.06728793680667877,
-0.0034857552964240313,
0.0889592170715332,
-0.024147043004631996,
0.04058205708861351,
-0.010532728396356106,
-0.0076717594638466835,
0.024848610162734985,
0.029445108026266098,
-0.03638074919581413,
-3.818469451744022e-8,
-0.006067521404474974,
0.07736671715974808,
0.015398881398141384,
0.008493505418300629,
0.023973077535629272,
0.03263033926486969,
-0.06048930436372757,
-0.009836501441895962,
-0.04180426523089409,
0.03329375386238098,
0.08525937795639038,
-0.00723288906738162,
0.012585073709487915,
-0.04718220978975296,
-0.03752949461340904,
0.031766775995492935,
0.014948082156479359,
-0.03524824604392052,
-0.016898183152079582,
-0.020425627008080482,
0.04812298342585564,
-0.0507451668381691,
0.06795188784599304,
-0.08371686935424805,
0.017281144857406616,
-0.07851210236549377,
-0.07170069962739944,
0.04798201844096184,
0.05036100000143051,
0.0411113016307354,
-0.003588686464354396,
-0.060943324118852615,
-0.001135057769715786,
0.0005056234076619148,
0.04617223143577576,
0.030337169766426086,
0.008854907937347889,
0.009991227649152279,
0.02715734951198101,
-0.02450498938560486,
0.0197941605001688,
0.07508863508701324,
0.01087227463722229,
0.007947207428514957,
-0.043007366359233856,
0.07884974032640457,
-0.05425862595438957,
-0.08744455873966217,
0.0441037155687809,
-0.014450320042669773,
0.02784121036529541,
-0.016235915943980217,
-0.05363345146179199,
0.10302369296550751,
0.10272631794214249,
0.027418870478868484,
-0.01447632908821106,
-0.05011685937643051,
0.0030288673005998135,
0.07892938703298569,
0.001044289325363934,
-0.02890203148126602,
-0.11733891069889069,
0.030897261574864388
] | 0 | 0 | 0 | 0 |
|
DPLM-2: A Multimodal Diffusion Protein Language Model | https://openreview.net/forum?id=5z9GjHgerY | [
"Xinyou Wang",
"Zaixiang Zheng",
"Fei YE",
"Dongyu Xue",
"Shujian Huang",
"Quanquan Gu"
] | Poster | Proteins are essential macromolecules defined by their amino acid sequences, which determine their three-dimensional structures and, consequently, their functions in all living organisms. Therefore, generative protein modeling necessitates a multimodal approach to simultaneously model, understand, and generate both sequences and structures. However, existing methods typically use separate models for each modality, limiting their ability to capture the intricate relationships between sequence and structure. This results in suboptimal performance in tasks that requires joint understanding and generation of both modalities.
In this paper, we introduce DPLM-2, a multimodal protein foundation model that extends discrete diffusion protein language model (DPLM) to accommodate both sequences and structures.
To enable structural learning with the language model, 3D coordinates are converted to discrete tokens using a lookup-free quantization-based tokenizer.
By training on both experimental and high-quality synthetic structures, DPLM-2 learns the joint distribution of sequence and structure, as well as their marginals and conditionals.
We also implement an efficient warm-up strategy to exploit the connection between large-scale evolutionary data and structural inductive biases from pre-trained sequence-based protein language models.
Empirical evaluation shows that DPLM-2 can simultaneously generate highly compatible amino acid sequences and their corresponding 3D structures eliminating the need for a two-stage generation approach.
Moreover, DPLM-2 demonstrates competitive performance in various conditional generation tasks, including folding, inverse folding, and scaffolding with multimodal motif inputs. | protein foundation model, diffusion language model, multimodal language model | null | 13,865 | null | [
-0.03223249316215515,
-0.19001510739326477,
0.03749983385205269,
-0.03960143402218819,
0.05507953464984894,
-0.01565188728272915,
-0.056605905294418335,
0.04025895521044731,
0.10675597190856934,
-0.05535805597901344,
0.037467461079359055,
-0.06944368779659271,
-0.028175748884677887,
0.09988220781087875,
0.03833619877696037,
-0.011308323591947556,
-0.02813839726150036,
0.11019471287727356,
-0.08412384241819382,
-0.06519394367933273,
0.08198148757219315,
0.07305891066789627,
0.014211839996278286,
-0.013632655143737793,
0.030140405520796776,
0.10372185707092285,
0.06014405936002731,
-0.011224476620554924,
0.051133960485458374,
-0.03268047049641609,
0.047847848385572433,
-0.037275947630405426,
-0.012842373922467232,
0.005854128859937191,
0.026516765356063843,
0.06435196846723557,
-0.06936411559581757,
-0.06404872983694077,
-0.01243748888373375,
-0.11145924776792526,
0.05479199066758156,
0.08042510598897934,
0.011008656583726406,
0.004592276178300381,
0.09171296656131744,
0.03220817446708679,
-0.04517386853694916,
0.0034311432391405106,
0.061126526445150375,
-0.02198019251227379,
-0.032850466668605804,
-0.05042172968387604,
-0.05280240625143051,
0.10142115503549576,
-0.023465072736144066,
-0.027551405131816864,
-0.004766071680933237,
-0.004938741214573383,
0.009737322106957436,
-0.06349165737628937,
-0.05341335013508797,
-0.08193143457174301,
0.03269026428461075,
-0.044018443673849106,
0.028796978294849396,
-0.053851839154958725,
0.015351939015090466,
0.09028245508670807,
-0.03846888616681099,
0.05219276249408722,
-0.038749076426029205,
0.05021735653281212,
-0.06804850697517395,
0.06632473319768906,
-0.013589081354439259,
0.03328903391957283,
0.03780994191765785,
-0.009577952325344086,
0.022694604471325874,
-0.0827450081706047,
0.04096555337309837,
0.03874494507908821,
0.10948553681373596,
0.02413470856845379,
0.06304161995649338,
-0.0075795468874275684,
-0.0053262184374034405,
0.049886345863342285,
0.005506178829818964,
0.000667561253067106,
-0.009160912595689297,
0.03043113276362419,
-0.005878226365894079,
0.005081681068986654,
-0.015264848247170448,
0.05251407250761986,
0.0038140160031616688,
-0.023642446845769882,
0.015134099870920181,
0.07447121292352676,
0.006436576135456562,
0.01540965586900711,
0.021006206050515175,
-0.07535836100578308,
0.03546227514743805,
-0.022762099280953407,
-0.008696275763213634,
0.03472115099430084,
0.050552211701869965,
0.024644378572702408,
0.04005249962210655,
0.05624194070696831,
0.015572730451822281,
0.08949651569128036,
0.02923484891653061,
-0.05656065046787262,
0.008599565364420414,
-0.027254993095993996,
0.10997128486633301,
-0.0410604402422905,
-0.005091571249067783,
-0.011700988747179508,
-0.05689967796206474,
0.023211263120174408,
-0.05934487283229828,
-0.036868855357170105,
-0.0848286971449852,
-2.0171495917079916e-33,
0.04217216372489929,
0.01970117911696434,
-0.011048832908272743,
0.11798103898763657,
0.07176194339990616,
-0.06371909379959106,
-0.020808439701795578,
0.01887177675962448,
-0.017216729000210762,
-0.005241030361503363,
-0.038883741945028305,
0.0029505921993404627,
-0.07562969624996185,
0.09838654100894928,
-0.006062714383006096,
-0.06741267442703247,
-0.008200565353035927,
0.0585675947368145,
0.022117819637060165,
0.021007465198636055,
-0.003915092907845974,
0.004186123143881559,
-0.04776673763990402,
0.048625413328409195,
0.041808657348155975,
0.0016204677522182465,
0.027022667229175568,
-0.09149482101202011,
-0.034674499183893204,
0.007970965467393398,
-0.20759151875972748,
-0.0790424570441246,
-0.05627638101577759,
0.00884899590164423,
0.06706037372350693,
-0.026738043874502182,
-0.034279145300388336,
-0.04351629316806793,
-0.012967314571142197,
-0.05061228573322296,
0.011362222954630852,
-0.009357544593513012,
-0.04972980171442032,
-0.005561502650380135,
-0.11025908589363098,
0.03075718693435192,
0.02819708175957203,
-0.026271270588040352,
-0.02959735319018364,
-0.14018888771533966,
0.04091028869152069,
-0.02182704769074917,
0.02945287525653839,
-0.060461655259132385,
-0.019817475229501724,
0.005288835149258375,
-0.03896772861480713,
-0.0035707741044461727,
0.0849190354347229,
0.021618442609906197,
-0.023426994681358337,
0.08808160573244095,
0.018677128478884697,
0.09738323092460632,
0.06592787802219391,
-0.05838618054986,
-0.14147716760635376,
0.05271739885210991,
0.06989631056785583,
0.09699687361717224,
-0.03663337975740433,
0.04765687882900238,
0.017302939668297768,
-0.04720070958137512,
0.02919967845082283,
-0.0495726577937603,
0.0673246756196022,
-0.03834645450115204,
0.01167042925953865,
0.09581059217453003,
0.014548525214195251,
0.027436349540948868,
-0.048250824213027954,
-0.07705458253622055,
-0.0360000804066658,
-0.049841031432151794,
0.05429910123348236,
-0.08836083114147186,
-0.08328408747911453,
-0.08011341840028763,
0.013629753142595291,
-0.033180370926856995,
0.003667947370558977,
0.04607972502708435,
-0.013962063938379288,
-7.557001092999533e-34,
-0.01950957253575325,
0.07199057191610336,
0.06634987145662308,
-0.0008013146580196917,
-0.05812060087919235,
-0.028510119765996933,
0.017847685143351555,
0.05204961448907852,
0.07217589765787125,
-0.04070206359028816,
0.021346326917409897,
-0.02286124788224697,
0.00673426641151309,
-0.01949359104037285,
0.06023568660020828,
0.04023405909538269,
-0.019191332161426544,
0.027622591704130173,
0.01878906786441803,
0.0570482462644577,
-0.03929112106561661,
0.05505305528640747,
-0.1290738731622696,
0.024655688554048538,
-0.09184092283248901,
0.0853809341788292,
-0.05847226083278656,
0.035786449909210205,
0.009406238794326782,
0.038995757699012756,
-0.060026172548532486,
0.019935501739382744,
-0.01934140920639038,
-0.047470103949308395,
-0.0928010642528534,
0.03460237383842468,
-0.021353168413043022,
-0.03292705863714218,
0.012067156843841076,
0.024751370772719383,
0.07240276783704758,
-0.012900757603347301,
-0.08434651046991348,
0.06873127073049545,
0.02595384605228901,
-0.01721923239529133,
0.0001776931603671983,
-0.006894032470881939,
-0.0020618701819330454,
0.029927514493465424,
0.05481678619980812,
0.048648685216903687,
-0.0632907897233963,
0.015985120087862015,
-0.03607422858476639,
-0.10520662367343903,
-0.0332912802696228,
-0.021428182721138,
0.01621198281645775,
-0.01985044591128826,
-0.05247761681675911,
-0.0881568044424057,
0.06346828490495682,
-0.032874271273612976,
-0.059198860079050064,
-0.048048846423625946,
-0.014424513094127178,
-0.05750511959195137,
-0.0028937046881765127,
-0.001503680949099362,
-0.018785977736115456,
0.0025804198812693357,
0.03047584556043148,
0.03361096233129501,
0.03524502366781235,
-0.09497840702533722,
-0.027578022330999374,
-0.017219163477420807,
-0.029463838785886765,
-0.028461303561925888,
-0.0014690851094201207,
0.03033597394824028,
-0.019868604838848114,
0.019871758297085762,
0.065249964594841,
0.021171607077121735,
0.06359107792377472,
0.093070849776268,
0.04439388960599899,
0.053943611681461334,
0.03258218616247177,
-0.024036675691604614,
0.02347363531589508,
0.1271292269229889,
-0.04571651294827461,
-4.280346033169735e-8,
-0.023436201736330986,
0.012683805078268051,
0.02444380894303322,
-0.02315773069858551,
0.0000955145078478381,
-0.019683413207530975,
-0.008865087293088436,
0.04499310255050659,
0.014750625938177109,
-0.027395037934184074,
0.008526665158569813,
0.02292361482977867,
-0.09179801493883133,
-0.031414855271577835,
-0.03341013938188553,
0.12162646651268005,
-0.055053066462278366,
-0.009362428449094296,
0.010300730355083942,
0.02721734531223774,
0.03501870855689049,
0.07713666558265686,
-0.016104336827993393,
0.015477319248020649,
0.07307583838701248,
-0.043343737721443176,
-0.046265099197626114,
-0.03107510507106781,
0.03009011782705784,
-0.06313109397888184,
-0.025047551840543747,
0.03650874271988869,
-0.06646981835365295,
-0.04114362969994545,
-0.03395524248480797,
0.07206539064645767,
0.028370385989546776,
-0.08278744667768478,
-0.03816252946853638,
-0.03482743352651596,
0.005271379370242357,
0.03959871828556061,
-0.06513560563325882,
0.02274935506284237,
-0.028215913102030754,
-0.01082163117825985,
-0.0018378038657829165,
0.004720629192888737,
-0.007510547526180744,
-0.0324653722345829,
0.005863358732312918,
0.016278602182865143,
-0.004718550946563482,
-0.00016319187125191092,
-0.024710319936275482,
0.009995535016059875,
-0.032880544662475586,
-0.014449388720095158,
0.03926817700266838,
-0.0409662090241909,
0.003701432840898633,
-0.0346049927175045,
0.005408772267401218,
0.01576131023466587
] | 0 | 0 | 0 | 0 |
|
Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation | https://openreview.net/forum?id=moWiYJuSGF | [
"Hyungjoo Chae",
"Namyoung Kim",
"Kai Tzu-iunn Ong",
"Minju Gwak",
"Gwanwoo Song",
"Jihoon Kim",
"Sunghwan Kim",
"Dongha Lee",
"Jinyoung Yeo"
] | Poster | Large language models (LLMs) have recently gained much attention in building autonomous agents. However, performance of current LLM-based web agents in long-horizon tasks is far from optimal, often yielding errors such as repeatedly buying a non-refundable flight ticket. By contrast, humans can avoid such an irreversible mistake, as we have an awareness of the potential outcomes (e.g., losing money) of our actions, also known as the "world model". Motivated by this, our study first starts with preliminary analyses, confirming the absence of world models in current LLMs (e.g., GPT-4o, Claude-3.5-Sonnet, etc.). Then, we present a World-model-augmented (WMA) web agent, which simulates the outcomes of its actions for better decision-making. To overcome the challenges in training LLMs as world models predicting next observations, such as repeated elements across observations and long HTML inputs, we propose a transition-focused observation abstraction, where the prediction objectives are free-form natural language descriptions exclusively highlighting important state differences between time steps. Experiments on WebArena and Mind2Web show that our world models improve agents' policy selection without training and demonstrate our agents' cost- and time-efficiency compared to recent tree-search-based agents. | Web Agent, World Model, Digital Agent, Planning, LLM | null | 13,861 | 2410.13232 | [
-0.021860409528017044,
-0.10408786684274673,
0.021082010120153427,
0.015046801418066025,
0.11558248847723007,
-0.03336762636899948,
0.01925739087164402,
-0.05536707490682602,
0.055428583174943924,
0.015157999470829964,
-0.0333310067653656,
-0.06285010278224945,
0.031050320714712143,
0.06777960062026978,
0.05135903134942055,
-0.03271912410855293,
0.0812130719423294,
-0.01224123127758503,
-0.06874746084213257,
-0.04119080677628517,
0.03691653162240982,
0.05585699528455734,
0.01524605043232441,
0.02641095407307148,
-0.05541073530912399,
0.04798588901758194,
0.00289409002289176,
-0.0456046499311924,
0.03488960117101669,
-0.011714814230799675,
0.00987048540264368,
-0.02143249660730362,
-0.04414873570203781,
0.029142824932932854,
-0.024080535396933556,
0.026674864813685417,
-0.12328728288412094,
-0.017751187086105347,
-0.05067574232816696,
-0.049623504281044006,
-0.0067545752972364426,
-0.04710398241877556,
-0.02670527622103691,
0.008763890713453293,
0.05297624319791794,
-0.02909221686422825,
-0.02485106699168682,
-0.011176995933055878,
-0.055091530084609985,
0.047576162964105606,
-0.12126626074314117,
-0.07351706176996231,
0.036471132189035416,
-0.03969339653849602,
-0.040097225457429886,
0.027549250051379204,
0.030104000121355057,
0.0029017601627856493,
-0.015615341253578663,
-0.0707860067486763,
-0.02227863296866417,
-0.09463182091712952,
-0.00007412001286866143,
-0.021266210824251175,
-0.036533173173666,
0.007866844534873962,
-0.02997758984565735,
0.05317489430308342,
-0.0036900655832141638,
-0.0016165833221748471,
-0.029200183227658272,
0.02348518557846546,
0.028023751452565193,
0.026542332023382187,
0.01622685417532921,
-0.049077317118644714,
0.03742671757936478,
-0.015568109229207039,
0.030128931626677513,
0.03595324978232384,
-0.005800690967589617,
-0.01862599141895771,
0.0027401098050177097,
0.04069110006093979,
0.0827656164765358,
-0.014174596406519413,
0.02626759186387062,
0.09058474749326706,
0.0491962656378746,
-0.037300366908311844,
-0.016543908044695854,
-0.09110674262046814,
0.07869881391525269,
0.030644599348306656,
0.05197324976325035,
0.08813238888978958,
-0.00993392150849104,
-0.06009802594780922,
0.010697434656322002,
0.08631246536970139,
0.030171586200594902,
0.06381908804178238,
0.024518663063645363,
-0.014183469116687775,
-0.005968399345874786,
-0.016407910734415054,
0.06825446337461472,
0.07873012125492096,
0.021405288949608803,
-0.044354476034641266,
-0.06761209666728973,
0.02794574573636055,
0.053636398166418076,
0.01958363689482212,
-0.018936049193143845,
-0.02640296146273613,
0.03584327921271324,
0.0039655244909226894,
-0.0020807706750929356,
0.06005279719829559,
0.04739826172590256,
0.033546071499586105,
0.10020486265420914,
0.01749560795724392,
0.08491383492946625,
-0.020473049953579903,
0.04398757591843605,
-1.1953831642252237e-33,
0.1068742573261261,
-0.00023768025857862085,
-0.021998004987835884,
0.051878735423088074,
0.09311774373054504,
0.05659567937254906,
0.0014405190013349056,
0.023448769003152847,
0.016881011426448822,
-0.023342015221714973,
0.004906590096652508,
0.05542997270822525,
-0.022162284702062607,
0.06441053003072739,
0.07377199083566666,
0.010558157227933407,
-0.04096384719014168,
0.05726979672908783,
0.015284005552530289,
-0.0566784106194973,
0.06908871233463287,
-0.011696970090270042,
0.04965503141283989,
-0.050030894577503204,
0.018206113949418068,
0.032442279160022736,
-0.018668336793780327,
-0.01535834465175867,
-0.0011572482762858272,
-0.006954897195100784,
-0.0420401506125927,
0.011000302620232105,
-0.09409896284341812,
0.08297667652368546,
-0.043775007128715515,
-0.025936361402273178,
-0.07745721936225891,
-0.0853230208158493,
0.01544200535863638,
-0.06845944374799728,
-0.06743375957012177,
0.03475351631641388,
0.008181076496839523,
-0.012610982172191143,
-0.07429347932338715,
-0.10715040564537048,
0.008069287054240704,
0.029870443046092987,
-0.04353852942585945,
0.005395726766437292,
0.0333794429898262,
0.02752063237130642,
-0.030535060912370682,
-0.06263692677021027,
-0.06896460801362991,
0.006307556759566069,
0.03971550241112709,
0.05328189954161644,
-0.039765916764736176,
-0.020132169127464294,
-0.00035599226248450577,
-0.021555988118052483,
0.027681700885295868,
0.08000603318214417,
0.029356297105550766,
0.06359322369098663,
-0.03154929354786873,
0.11084242910146713,
0.07314222306013107,
-0.06199174374341965,
0.007443833630532026,
0.017233707010746002,
0.0444256030023098,
0.004397690296173096,
0.025106212124228477,
-0.0426112562417984,
0.07170287519693375,
-0.10833840072154999,
0.01579258218407631,
0.01449530478566885,
0.05060385912656784,
-0.013436038047075272,
-0.00626504048705101,
-0.009283305145800114,
0.011663555167615414,
-0.06519945710897446,
0.008346782997250557,
-0.10156819224357605,
-0.003276226343587041,
-0.030587485060095787,
-0.1262153834104538,
-0.000607855326961726,
-0.032470300793647766,
0.03391209617257118,
0.00936437863856554,
-7.797436983023212e-34,
-0.049334414303302765,
-0.09487758576869965,
-0.015403129160404205,
0.03150864690542221,
-0.03522571548819542,
-0.08999428153038025,
-0.04692728444933891,
0.05555027350783348,
0.007546443957835436,
-0.06158158555626869,
-0.04661853238940239,
0.004969635047018528,
0.0730721652507782,
0.0283796526491642,
0.02607148140668869,
-0.06713522225618362,
0.05294693633913994,
-0.1057502031326294,
0.04066649079322815,
0.05227038264274597,
0.07352038472890854,
0.04637870937585831,
-0.16918408870697021,
-0.035464465618133545,
0.012104631401598454,
0.03105800785124302,
0.05537407845258713,
0.09271277487277985,
-0.006587504874914885,
0.04121321067214012,
-0.006467932369560003,
-0.008226651698350906,
-0.02097642607986927,
-0.032405197620391846,
-0.044285938143730164,
0.07219630479812622,
0.05131273344159126,
0.003011221531778574,
-0.0657692477107048,
0.09672495722770691,
0.06397615373134613,
-0.10922128707170486,
-0.015564368106424809,
-0.031055843457579613,
-0.069511778652668,
0.05106980353593826,
-0.09782539308071136,
0.06954745948314667,
0.008874374441802502,
-0.029487183317542076,
0.007931960746645927,
0.02237388677895069,
-0.07055893540382385,
-0.014282633550465107,
-0.06393881142139435,
-0.02082509733736515,
-0.03162268176674843,
-0.0556928850710392,
0.029419057071208954,
-0.012679388746619225,
-0.10766292363405228,
-0.002197379944846034,
0.01708242855966091,
0.030478717759251595,
-0.02394939213991165,
-0.06555258482694626,
-0.06580513715744019,
0.02729007788002491,
0.020127976313233376,
-0.02896248921751976,
0.06991859525442123,
-0.003975751344114542,
-0.07657070457935333,
-0.020349593833088875,
0.07858745753765106,
0.0006043547764420509,
0.022078026086091995,
-0.009188788011670113,
0.024594146758317947,
-0.06466340273618698,
0.017113670706748962,
0.03996342793107033,
0.0654715895652771,
0.007333576679229736,
0.03472867235541344,
0.06795921921730042,
-0.01195413339883089,
0.08483393490314484,
-0.017394479364156723,
-0.011795243248343468,
-0.07112080603837967,
0.005368408281356096,
-0.04114344343543053,
0.07196735590696335,
-0.14024731516838074,
-4.294747668609489e-8,
-0.08337977528572083,
0.05872267484664917,
0.061281900852918625,
0.08816653490066528,
0.018729083240032196,
-0.013836757279932499,
0.019010797142982483,
-0.012401989661157131,
-0.02554255910217762,
-0.013007547706365585,
0.07690870761871338,
0.014731147326529026,
-0.013936827890574932,
0.002875646110624075,
-0.0017313277348876,
0.06382711976766586,
-0.020259151235222816,
-0.07636994123458862,
0.007870778441429138,
0.007779193576425314,
0.06276416778564453,
0.02009996771812439,
0.0007273218361660838,
-0.026636892929673195,
0.04931168258190155,
-0.006237465422600508,
-0.06528691202402115,
0.08940524607896805,
0.00023953354684635997,
-0.009117761626839638,
0.0030172530096024275,
0.017360474914312363,
0.005949168931692839,
0.019859567284584045,
0.020064620301127434,
0.011217506602406502,
-0.010149098001420498,
-0.09801395237445831,
0.0008059694082476199,
-0.04913480207324028,
0.05683588609099388,
0.0766189768910408,
-0.057864703238010406,
0.012734862975776196,
0.05475802347064018,
0.007960126735270023,
-0.0924156978726387,
-0.1347847282886505,
0.07626228034496307,
-0.09264916181564331,
-0.0669013261795044,
-0.05181615427136421,
-0.020645713433623314,
0.055252671241760254,
0.12313462793827057,
0.02636772394180298,
0.03918219730257988,
-0.05144176259636879,
0.06986445188522339,
0.09799698740243912,
0.02049967832863331,
-0.024486536160111427,
-0.08030123263597488,
-0.00944235548377037
] | https://github.com/kyle8581/wma-agents | 6 | 0 | 0 | 0 |
HyperFace: Generating Synthetic Face Recognition Datasets by Exploring Face Embedding Hypersphere | https://openreview.net/forum?id=4YzVF9isgD | [
"Hatef Otroshi Shahreza",
"Sébastien Marcel"
] | Poster | Face recognition datasets are often collected by crawling Internet and without individuals' consents, raising ethical and privacy concerns. Generating synthetic datasets for training face recognition models has emerged as a promising alternative. However, the generation of synthetic datasets remains challenging as it entails adequate inter-class and intra-class variations. While advances in generative models have made it easier to increase intra-class variations in face datasets (such as pose, illumination, etc.), generating sufficient inter-class variation is still a difficult task. In this paper, we formulate the dataset generation as a packing problem on the embedding space (represented on a hypersphere) of a face recognition model and propose a new synthetic dataset generation approach, called HyperFace. We formalize our packing problem as an optimization problem and solve it with a gradient descent-based approach. Then, we use a conditional face generator model to synthesize face images from the optimized embeddings. We use our generated datasets to train face recognition models and evaluate the trained models on several benchmarking real datasets. Our experimental results show that models trained with HyperFace achieve state-of-the-art performance in training face recognition using synthetic datasets. Project page: https://www.idiap.ch/paper/hyperface | Face Recognition, Hypersphere Optimization, Privacy, Synthetic Data | We formulate the dataset generation as a packing problem on the embedding space (represented on a hypersphere) of a face recognition model and propose a new synthetic dataset generation approach. | 13,859 | 2411.08470 | [
-0.08422822505235672,
-0.04884840548038483,
0.03281238302588463,
-0.045451123267412186,
0.06592454761266708,
0.011305902153253555,
-0.01680644042789936,
-0.06386496871709824,
-0.060625385493040085,
-0.020669175311923027,
0.0030195889994502068,
-0.08941604942083359,
0.0507160909473896,
0.05036827176809311,
0.009525155648589134,
-0.057376373559236526,
0.13043269515037537,
0.08993025869131088,
-0.05575692653656006,
0.08459343761205673,
0.03502880409359932,
-0.005093313753604889,
0.055782005190849304,
-0.046339597553014755,
0.01825810968875885,
0.013112748973071575,
0.11893189698457718,
-0.021675964817404747,
0.011963116936385632,
-0.002868094714358449,
0.08343883603811264,
-0.03428799659013748,
-0.051178816705942154,
0.047486986964941025,
-0.01555759459733963,
0.08503095805644989,
-0.06096168980002403,
0.09153442084789276,
-0.038843825459480286,
-0.009766812436282635,
-0.10990163683891296,
-0.026717791333794594,
-0.031076673418283463,
0.10329481959342957,
0.057374756783246994,
0.02118457295000553,
-0.013828673399984837,
-0.014273550361394882,
0.034666143357753754,
-0.0815974548459053,
-0.044169459491968155,
-0.13284625113010406,
-0.06776881963014603,
0.020123712718486786,
-0.0783412903547287,
-0.024768276140093803,
0.07536674290895462,
-0.020699970424175262,
-0.01939181424677372,
0.048871662467718124,
-0.022119268774986267,
-0.0973857119679451,
0.02904277667403221,
-0.021565835922956467,
-0.07875639200210571,
-0.02122803032398224,
0.012140645645558834,
0.012331992387771606,
-0.0117415189743042,
0.03869711235165596,
-0.06243795156478882,
0.08041936159133911,
-0.13605602085590363,
-0.013685710728168488,
-0.03127613663673401,
0.03496534377336502,
0.04408933222293854,
-0.001729635871015489,
0.07138872891664505,
-0.06155531853437424,
0.027093760669231415,
0.0257707592099905,
0.036861322820186615,
-0.012508346699178219,
0.021900782361626625,
-0.044641170650720596,
-0.015307869762182236,
0.0597396120429039,
-0.04321020469069481,
0.03191661089658737,
-0.08499889075756073,
0.04461701959371567,
0.045182593166828156,
-0.05073118582367897,
0.024371152743697166,
-0.01460541132837534,
0.020099956542253494,
0.005214888136833906,
0.044303517788648605,
0.06614591181278229,
-0.06740595400333405,
0.028182435780763626,
0.007898762822151184,
-0.008022839203476906,
0.049952857196331024,
-0.07158776372671127,
-0.012939088977873325,
0.030914761126041412,
0.05427788197994232,
-0.0878928080201149,
0.06374437361955643,
0.0048035490326583385,
-0.07659215480089188,
0.0003415601095184684,
0.0023561303969472647,
-0.023206491023302078,
0.05673801153898239,
0.019319972023367882,
-0.0000607299261901062,
0.02358134277164936,
-0.004767066799104214,
0.047477446496486664,
0.05697254091501236,
-0.008870972320437431,
0.009707482531666756,
-0.01678226701915264,
-0.061026476323604584,
1.0431913596112852e-33,
0.05861438438296318,
0.06503832340240479,
-0.016573172062635422,
0.06879720091819763,
0.01375542115420103,
0.011318895034492016,
-0.005153272766619921,
0.046445365995168686,
-0.035066019743680954,
0.055250268429517746,
-0.065909743309021,
-0.056169576942920685,
-0.0255422443151474,
0.15540353953838348,
0.052575480192899704,
0.03670743852853775,
-0.017455488443374634,
-0.0031611237209290266,
-0.007203544955700636,
-0.01816932111978531,
0.014134560711681843,
-0.04565609246492386,
0.04696442559361458,
0.0360763780772686,
-0.09875016659498215,
0.015526995994150639,
0.059609781950712204,
-0.06424064934253693,
-0.008549834601581097,
-0.017569977790117264,
-0.07845153659582138,
0.02925482764840126,
0.05085934326052666,
0.036084480583667755,
0.12127774953842163,
-0.0017219100845977664,
0.00009082665201276541,
-0.0302279032766819,
0.01751251146197319,
-0.020904945209622383,
0.044362880289554596,
0.07412311434745789,
0.037331026047468185,
-0.010947275906801224,
-0.027287272736430168,
0.044960249215364456,
0.022758357226848602,
0.04657099023461342,
-0.023129789158701897,
0.020855970680713654,
0.03931111469864845,
0.0662154108285904,
-0.032593272626399994,
-0.04725588858127594,
-0.08684206753969193,
-0.01597326062619686,
0.014429835602641106,
0.05490396171808243,
0.03062850795686245,
-0.004429372493177652,
-0.015236482955515385,
0.007574890274554491,
0.050954390317201614,
0.08921343088150024,
-0.06440126150846481,
-0.049470044672489166,
0.013629846274852753,
-0.05343532934784889,
-0.048232853412628174,
0.03981291130185127,
0.0696994811296463,
0.030288754031062126,
-0.04988847300410271,
-0.05380535125732422,
0.032664358615875244,
0.02116205170750618,
0.09938579797744751,
0.018490426242351532,
-0.032896582037210464,
0.041784726083278656,
0.025164969265460968,
0.09314241260290146,
-0.012897677719593048,
-0.15201601386070251,
-0.04503290727734566,
-0.08705730736255646,
-0.0026428133714944124,
-0.03881888836622238,
0.021704459562897682,
-0.002421117853373289,
-0.06431134790182114,
-0.0540764220058918,
-0.05183960124850273,
-0.004223630763590336,
-0.03646865114569664,
-1.5405708455995309e-33,
-0.014095824211835861,
0.013058790937066078,
-0.049628846347332,
-0.053838130086660385,
0.012458288110792637,
0.024381885305047035,
0.05142810940742493,
0.05598882585763931,
-0.038516443222761154,
0.026839297264814377,
0.0006583051872439682,
-0.017531348392367363,
0.03804050758481026,
-0.031152108684182167,
-0.017395872622728348,
0.049107763916254044,
-0.012871899642050266,
-0.053588610142469406,
-0.07700926810503006,
0.0343041867017746,
-0.02944907173514366,
0.08410490304231644,
-0.08332867920398712,
0.05308925732970238,
-0.056316979229450226,
0.04676872864365578,
-0.018711702898144722,
0.08050238341093063,
-0.025138182565569878,
0.05911608785390854,
0.008580655790865421,
-0.0417754203081131,
-0.08784706890583038,
0.017190193757414818,
-0.10266482084989548,
0.013821463100612164,
0.024481981992721558,
0.0013476114254444838,
-0.05782438814640045,
0.11083509773015976,
0.014767163433134556,
-0.030002327635884285,
-0.13041406869888306,
0.08011019974946976,
-0.02118007466197014,
-0.11691217124462128,
-0.06912340223789215,
-0.05357058346271515,
0.0017883400432765484,
0.03920724242925644,
-0.058975737541913986,
0.016279879957437515,
-0.02469256892800331,
0.07865475118160248,
-0.028856739401817322,
-0.10008380562067032,
-0.05763166770339012,
0.03519425168633461,
0.08401849120855331,
-0.018801862373948097,
-0.004855853971093893,
-0.0713493824005127,
-0.07533445954322815,
0.05465532839298248,
-0.04999057203531265,
-0.037710510194301605,
-0.0449489951133728,
-0.04403797164559364,
-0.009712773375213146,
0.062290627509355545,
0.08606147021055222,
-0.05300629511475563,
-0.004729086998850107,
0.01001636404544115,
0.0007978452485986054,
0.0006100664613768458,
0.06685101240873337,
0.0673876702785492,
-0.017552385106682777,
0.016583332791924477,
0.024199161678552628,
-0.035235822200775146,
0.09461855888366699,
0.05692831799387932,
0.05919722095131874,
0.029925527051091194,
0.06667738407850266,
0.011675928719341755,
-0.02881755493581295,
0.0032829437404870987,
-0.042101502418518066,
0.020688338205218315,
-0.027965787798166275,
0.05729975923895836,
0.055016640573740005,
-3.693697081530445e-8,
-0.022010060027241707,
-0.02346370741724968,
0.04301443323493004,
-0.06086632236838341,
-0.023180773481726646,
0.02232680283486843,
-0.01500210165977478,
0.06927570700645447,
-0.10845601558685303,
0.00997349712997675,
0.08082360029220581,
-0.007073931396007538,
0.011977354995906353,
-0.0040488336235284805,
0.00364487711340189,
0.10183507949113846,
-0.049105994403362274,
0.05096270143985748,
0.01037216279655695,
0.034922946244478226,
-0.06770928204059601,
0.025906383991241455,
0.07239966094493866,
-0.02286536805331707,
0.08563252538442612,
-0.050923869013786316,
-0.00740227522328496,
0.05679146200418472,
-0.00805522222071886,
0.008294102735817432,
-0.07799125462770462,
0.06146491691470146,
-0.026472408324480057,
-0.054907627403736115,
-0.02611444517970085,
0.01806398667395115,
-0.017527960240840912,
-0.05699657276272774,
-0.012422792613506317,
-0.00642777606844902,
0.0010863899951800704,
0.05292497202754021,
-0.04538451135158539,
-0.009607185609638691,
0.07318461686372757,
0.07926015555858612,
0.05938517674803734,
-0.04096071049571037,
0.015545275993645191,
0.07335410267114639,
-0.028563514351844788,
-0.014619662426412106,
-0.01878676563501358,
0.03843843936920166,
0.0010188831947743893,
-0.007243819069117308,
-0.018466562032699585,
-0.022808676585555077,
0.06276532262563705,
0.05050204321742058,
-0.015112899243831635,
0.04298178479075432,
-0.029257675632834435,
-0.05968008190393448
] | 0 | 0 | 0 | 0 |
|
Language Imbalance Driven Rewarding for Multilingual Self-improving | https://openreview.net/forum?id=Kak2ZH5Itp | [
"Wen Yang",
"Junhong Wu",
"Chen Wang",
"Chengqing Zong",
"Jiajun Zhang"
] | Poster | Large Language Models (LLMs) have achieved state-of-the-art performance across numerous tasks. However, these advancements have predominantly benefited "first-class" languages such as English and Chinese, leaving many other languages underrepresented. This imbalance, while limiting broader applications, generates a natural preference ranking between languages, offering an opportunity to bootstrap the multilingual capabilities of LLM in a self-improving manner. Thus, we propose $\textit{Language Imbalance Driven Rewarding}$, where the inherent imbalance between dominant and non-dominant languages within LLMs is leveraged as a reward signal. Iterative DPO training demonstrates that this approach not only enhances LLM performance in non-dominant languages but also improves the dominant language's capacity, thereby yielding an iterative reward signal. Fine-tuning Meta-Llama-3-8B-Instruct over two iterations of this approach results in continuous improvements in multilingual performance across instruction-following and arithmetic reasoning tasks, evidenced by an average improvement of 7.46\% win rate on the X-AlpacaEval leaderboard and 13.9\% accuracy on the MGSM benchmark. This work serves as an initial exploration, paving the way for multilingual self-improvement of LLMs. | Large Language Model, Self-Improving, Multilinguality | This paper proposes Language Imbalance Driven Rewarding, which leverages the inherent imbalance in LLMs as a reward signal to bootstrap LLMs’ multilingual capabilities in a self-improving manner. | 13,855 | 2410.08964 | [
-0.02707487903535366,
-0.06783077120780945,
0.05976999178528786,
-0.03586019203066826,
-0.03848688304424286,
0.0111272307112813,
0.01555640622973442,
0.0868663415312767,
0.052177365869283676,
-0.00006945480708964169,
-0.08314650505781174,
-0.07050227373838425,
0.050891704857349396,
0.04209364950656891,
0.05903102084994316,
-0.002664862433448434,
0.06661558151245117,
0.040832798928022385,
-0.05947299674153328,
-0.10242363810539246,
0.05357450991868973,
0.006681379396468401,
0.07924651354551315,
0.025875836610794067,
0.004312312230467796,
-0.015804961323738098,
-0.05074600502848625,
-0.007426861673593521,
0.07544032484292984,
-0.056805383414030075,
0.01169262919574976,
0.05738697946071625,
0.05003499239683151,
0.03356252610683441,
-0.13163603842258453,
0.030487503856420517,
-0.08003006875514984,
-0.03128098323941231,
-0.009408661164343357,
-0.02922901138663292,
-0.012370653450489044,
0.022775836288928986,
0.07979843765497208,
-0.05211644247174263,
0.08070424944162369,
0.02913147583603859,
-0.05457403138279915,
-0.0028477369341999292,
-0.0349278524518013,
0.010246443562209606,
-0.07490875571966171,
0.0017993804067373276,
0.031689465045928955,
0.039279114454984665,
-0.007158322259783745,
-0.007515147794038057,
0.08045828342437744,
0.06440026313066483,
-0.023876909166574478,
-0.07800320535898209,
-0.0657610148191452,
-0.08319246768951416,
0.008255167864263058,
0.022860093042254448,
0.02204381301999092,
-0.0571393258869648,
0.01715547777712345,
0.05779128149151802,
-0.09974842518568039,
0.047793809324502945,
0.02817966416478157,
0.01789039373397827,
0.007216581609100103,
0.08597680926322937,
0.07294652611017227,
0.09239863604307175,
0.00788184441626072,
-0.06907236576080322,
0.05440644174814224,
-0.04185498505830765,
-0.02101815678179264,
0.046944767236709595,
0.008177762851119041,
0.02058384381234646,
0.07835608720779419,
-0.05120307579636574,
-0.025694122537970543,
0.005199093837291002,
0.06244390457868576,
0.033208269625902176,
0.002918277168646455,
0.009898609481751919,
-0.010577323846518993,
0.052616823464632034,
-0.011830272153019905,
0.07065621763467789,
-0.006012832745909691,
-0.02281353808939457,
-0.09301106631755829,
0.06161412596702576,
0.08778372406959534,
0.07868973165750504,
-0.06327057629823685,
0.016174020245671272,
-0.028231725096702576,
-0.06921742856502533,
0.04907980188727379,
0.015431330539286137,
0.05585620924830437,
-0.14309175312519073,
0.019599957391619682,
0.04120523855090141,
-0.00619844626635313,
0.03297455236315727,
-0.02208755351603031,
0.021518681198358536,
0.029364347457885742,
0.0022936458699405193,
0.07106287777423859,
0.04897238686680794,
-0.05599289759993553,
-0.053820110857486725,
0.03942164033651352,
0.0035991526674479246,
0.009384125471115112,
-0.028579048812389374,
-0.04068543016910553,
1.7778292274561435e-33,
0.07853204756975174,
0.057124458253383636,
0.014022459276020527,
-0.004618676844984293,
0.020763099193572998,
-0.0414109043776989,
0.03788245469331741,
0.05631927028298378,
-0.1347326785326004,
-0.030055886134505272,
-0.021257000043988228,
0.09823745489120483,
-0.03029416874051094,
0.06802075356245041,
0.07861075550317764,
-0.025605812668800354,
-0.025522733107209206,
0.016137320548295975,
-0.02272341027855873,
-0.0702243521809578,
0.13888998329639435,
0.007494538091123104,
0.05415491387248039,
-0.08765547722578049,
0.03101758100092411,
0.021517863497138023,
0.054800279438495636,
-0.029344432055950165,
-0.0837443396449089,
0.012306391261518002,
-0.06510106474161148,
-0.0508413203060627,
-0.017440244555473328,
0.016231590881943703,
-0.02622893638908863,
-0.09049681574106216,
-0.011103152297437191,
-0.020722275599837303,
0.06169866397976875,
0.02474391646683216,
-0.018494732677936554,
-0.008971024304628372,
0.0288055632263422,
0.009584457613527775,
-0.027189720422029495,
-0.008686901070177555,
0.008501298725605011,
0.001992110861465335,
-0.07136678695678711,
-0.006798925809562206,
0.010135140269994736,
0.014698667451739311,
-0.0013301849830895662,
-0.014844154007732868,
0.04006965458393097,
0.06591861695051193,
0.019172390922904015,
0.05580024793744087,
0.011964895762503147,
-0.0083578210324049,
-0.01599915884435177,
0.04034056514501572,
-0.04210180789232254,
0.11599151045084,
0.021756883710622787,
0.01665983349084854,
0.02475172095000744,
0.012960793450474739,
0.022799700498580933,
-0.009100092574954033,
-0.0026168511249125004,
-0.042359303683042526,
-0.045136526226997375,
-0.05345612019300461,
0.033112067729234695,
0.013079721480607986,
0.05943794548511505,
-0.1130920946598053,
0.06519947201013565,
-0.0022598295472562313,
0.03179219737648964,
-0.023675091564655304,
-0.019941255450248718,
-0.04492836073040962,
-0.009257343597710133,
-0.030694397166371346,
0.07718417048454285,
-0.11937543749809265,
0.0977209135890007,
0.007929776795208454,
-0.002498115412890911,
-0.08733242005109787,
0.030452551320195198,
-0.013226461596786976,
-0.016862140968441963,
-1.1143871953497525e-33,
-0.07280118763446808,
-0.018126588314771652,
-0.03715120255947113,
0.06898434460163116,
-0.01859636977314949,
-0.04677242413163185,
0.04741436988115311,
-0.0040182629600167274,
-0.04617074504494667,
-0.04989805445075035,
-0.04147101938724518,
0.01734839752316475,
0.04952698573470116,
0.08666229248046875,
0.007423109374940395,
-0.021287990733981133,
0.02222917601466179,
0.09237680584192276,
0.0036809719167649746,
0.01759503223001957,
0.014810130931437016,
0.10035835951566696,
-0.10373220592737198,
0.046035490930080414,
-0.037096887826919556,
0.03224002942442894,
-0.042074501514434814,
0.050805989652872086,
-0.05149569362401962,
0.023250620812177658,
0.018200090155005455,
0.00998659897595644,
0.00612421752884984,
-0.03431588411331177,
-0.04011649638414383,
0.03593318536877632,
-0.019905203953385353,
-0.09304425865411758,
-0.045545365661382675,
0.1357937902212143,
0.016837328672409058,
-0.02347865328192711,
-0.006100006401538849,
-0.0016307217301800847,
-0.007878969423472881,
-0.015282342210412025,
-0.13381317257881165,
-0.017158804461359978,
0.03649509325623512,
0.047498978674411774,
0.028213107958436012,
0.016745684668421745,
-0.0029476385097950697,
0.07915833592414856,
-0.010564974509179592,
-0.11262691020965576,
0.007846124470233917,
-0.049820758402347565,
-0.011696279980242252,
0.0035232037771493196,
-0.08515429496765137,
0.024442845955491066,
0.08386711031198502,
-0.11919938027858734,
0.0030744823161512613,
-0.031170537695288658,
-0.041172370314598083,
-0.008072887547314167,
0.025242965668439865,
-0.016904249787330627,
0.0212552547454834,
0.016786539927124977,
-0.03057129867374897,
-0.05360936000943184,
-0.03575543314218521,
0.03746192157268524,
-0.007325960323214531,
0.02078968845307827,
-0.0281926728785038,
-0.08211886137723923,
-0.05269365385174751,
0.0224007498472929,
0.03286776691675186,
-0.04037029668688774,
-0.041529010981321335,
0.023412523791193962,
0.03243963047862053,
0.10283775627613068,
0.0043154130689799786,
0.0172688290476799,
0.017598962411284447,
0.0292851310223341,
0.01039667148143053,
0.03317173197865486,
-0.05997126176953316,
-4.6550905352660266e-8,
-0.08736038953065872,
0.000036959634599043056,
0.008234597742557526,
0.11881514638662338,
0.03956156224012375,
-0.0500158816576004,
-0.06666882336139679,
0.02816045843064785,
0.023526430130004883,
-0.0577407032251358,
0.060017943382263184,
0.010934433899819851,
-0.024700237438082695,
-0.06566190719604492,
0.03342118114233017,
0.07104402035474777,
0.023412618786096573,
-0.015893692150712013,
-0.02325660176575184,
-0.04792739823460579,
0.053902558982372284,
0.02302004024386406,
0.024701761081814766,
-0.12771598994731903,
-0.035840850323438644,
-0.04801364243030548,
-0.06995576620101929,
0.04017424210906029,
0.08801745623350143,
-0.0242350772023201,
-0.004143074620515108,
0.0021618292666971684,
-0.05325078219175339,
-0.04446016624569893,
0.0713827982544899,
0.06633368879556656,
-0.01904154196381569,
-0.021614598110318184,
-0.000850701704621315,
0.006432381924241781,
0.08796016871929169,
0.08058255165815353,
-0.006659246049821377,
0.09385454654693604,
-0.017434414476156235,
-0.028734195977449417,
-0.11209870874881744,
-0.04546822980046272,
0.008775078691542149,
-0.09806908667087555,
-0.012913304381072521,
-0.044252026826143265,
-0.010939233005046844,
0.06263473629951477,
0.0601724311709404,
0.046887632459402084,
-0.06294408440589905,
-0.04986455291509628,
-0.020287837833166122,
0.10973411798477173,
0.05965466424822807,
-0.0341247096657753,
-0.04183758422732353,
-0.0026607420295476913
] | https://github.com/znlp/language-imbalance-driven-rewarding | 17 | 0 | 0 | 0 |
Quantum-PEFT: Ultra parameter-efficient fine-tuning | https://openreview.net/forum?id=dgR6i4TSng | [
"Toshiaki Koike-Akino",
"Francesco Tonin",
"Yongtao Wu",
"Zhengqing Wu",
"Leyla Naz Candogan",
"Volkan Cevher"
] | Poster | This paper introduces Quantum-PEFT that leverages quantum computations for parameter-efficient fine-tuning (PEFT). Unlike other additive PEFT methods, such as low-rank adaptation (LoRA), Quantum-PEFT exploits an underlying full-rank yet surprisingly parameter efficient _quantum unitary parameterization_. With the use of Pauli parameterization, the number of trainable parameters grows only logarithmically with the ambient dimension, as opposed to linearly as in LoRA-based PEFT methods. Quantum-PEFT achieves vanishingly smaller number of trainable parameters than the lowest-rank LoRA as dimensions grow, enhancing parameter efficiency while maintaining a competitive performance. We apply Quantum-PEFT to several transfer learning benchmarks in language and vision, demonstrating significant advantages in parameter efficiency. | parameter-efficient fine-tuning, lora, quantum machine learning, orthogonality constraints | null | 13,846 | null | [
-0.05355680733919144,
-0.05116339772939682,
-0.01024998165667057,
0.0568389892578125,
-0.021366925910115242,
0.04054100811481476,
0.025393519550561905,
-0.07661622762680054,
0.0009399394039064646,
-0.04634653031826019,
-0.021270666271448135,
-0.05928783491253853,
-0.024191323667764664,
0.028194600716233253,
-0.020859336480498314,
0.043462447822093964,
0.04567212238907814,
0.08488821983337402,
-0.07153647392988205,
0.013627744279801846,
0.032695744186639786,
-0.010025492869317532,
0.06511145085096359,
-0.019572827965021133,
0.09300374984741211,
-0.015493541955947876,
-0.0177924744784832,
-0.00772758899256587,
0.07663814723491669,
-0.07967136800289154,
-0.03552708402276039,
0.09616799652576447,
-0.04702771455049515,
-0.04570259526371956,
-0.08037552237510681,
0.035873137414455414,
-0.01540137454867363,
-0.003448608797043562,
0.0012412706855684519,
-0.04877393692731857,
0.08024045825004578,
0.0830913856625557,
-0.03693746030330658,
0.03932445868849754,
0.05908939987421036,
0.04552987962961197,
0.059443067759275436,
0.008011551573872566,
-0.03313199058175087,
-0.1012752577662468,
-0.024997172877192497,
0.03306794911623001,
-0.008585023693740368,
0.08097263425588608,
-0.009075487032532692,
0.008776305243372917,
0.05911244451999664,
-0.031744081526994705,
-0.0190072450786829,
-0.08743041753768921,
-0.09757248312234879,
-0.03493889048695564,
0.006705018226057291,
-0.024929331615567207,
0.050166480243206024,
-0.05696418508887291,
-0.022687990218400955,
0.01413444522768259,
-0.05682532489299774,
0.035449445247650146,
-0.09243828803300858,
0.0016450252151116729,
-0.05688194930553436,
0.0217104684561491,
0.07984504103660583,
0.04371502622961998,
0.057368505746126175,
0.028570333495736122,
-0.0001921556977322325,
-0.0023961090482771397,
0.04492798075079918,
-0.09749363362789154,
-0.004852629266679287,
-0.01985565386712551,
0.18795067071914673,
0.039397209882736206,
0.00332591007463634,
-0.01793990284204483,
-0.026575032621622086,
-0.05238065868616104,
0.002296741586178541,
-0.018875859677791595,
0.011180228553712368,
-0.044454582035541534,
-0.0036173765547573566,
0.04829571768641472,
-0.0076077887788414955,
-0.0575423389673233,
-0.0020101568661630154,
0.09811089187860489,
0.07641001790761948,
0.05680128559470177,
-0.0020192195661365986,
-0.004181170370429754,
-0.05490119010210037,
-0.05069569870829582,
0.06883037090301514,
0.03606486693024635,
-0.018445175141096115,
-0.10435682535171509,
0.07101663202047348,
0.03725014626979828,
0.01922600530087948,
0.0411907322704792,
-0.0052845231257379055,
0.04153003543615341,
-0.0441981703042984,
0.001059450558386743,
0.022086963057518005,
0.02266068384051323,
-0.1167643591761589,
0.019280031323432922,
-0.05114957317709923,
0.12887230515480042,
-0.0471099317073822,
-0.002198119880631566,
-0.08592890202999115,
2.2425745107071663e-33,
0.06067246198654175,
0.11575043201446533,
-0.0013989113504067063,
-0.03927457705140114,
-0.013850340619683266,
0.03465251997113228,
0.08893002569675446,
-0.026674902066588402,
-0.052650175988674164,
-0.012877724133431911,
-0.04691736400127411,
0.0003791951749008149,
-0.004690323956310749,
0.04294008016586304,
0.037909455597400665,
-0.012844989076256752,
-0.057093679904937744,
0.03978210687637329,
0.0412050224840641,
-0.07697855681180954,
0.09071589261293411,
-0.020710840821266174,
-0.032921578735113144,
-0.03568192943930626,
-0.0390109047293663,
0.01005788054317236,
0.06929091364145279,
-0.010120758786797523,
-0.04568222537636757,
-0.009647512808442116,
-0.05745508521795273,
0.04860794544219971,
-0.010458149015903473,
-0.03123781457543373,
0.03892570361495018,
-0.07860589772462845,
-0.01587652415037155,
0.05189584195613861,
-0.010814341716468334,
-0.07013989239931107,
-0.024326849728822708,
0.004740003030747175,
0.07152745872735977,
0.028951141983270645,
-0.04296588525176048,
-0.0456489659845829,
-0.019307352602481842,
0.01933443173766136,
-0.03533157333731651,
-0.036504387855529785,
-0.04944111034274101,
-0.08789517730474472,
-0.09519529342651367,
-0.02857680805027485,
0.007701550144702196,
-0.05048036202788353,
0.13095660507678986,
0.06295187771320343,
-0.042957853525877,
0.05698636919260025,
0.0014294899301603436,
0.05908547714352608,
-0.05996502935886383,
0.027108145877718925,
-0.017325839027762413,
0.006285710260272026,
-0.019373930990695953,
0.05346740782260895,
0.0015482159797102213,
0.08661676943302155,
-0.019705604761838913,
0.026330284774303436,
-0.010043916292488575,
-0.02359291911125183,
0.12852422893047333,
-0.07707031816244125,
0.017962388694286346,
-0.07753539830446243,
0.051206231117248535,
0.04223420098423958,
0.014853245578706264,
0.016700047999620438,
-0.01753806881606579,
-0.0026853764429688454,
-0.03727269545197487,
-0.18425461649894714,
-0.06936097890138626,
-0.07457093894481659,
0.025551440194249153,
-0.030071066692471504,
-0.025676460936665535,
-0.04340996965765953,
0.010695121251046658,
-0.022494053468108177,
-0.04587569087743759,
-2.8311123897402303e-33,
0.016375672072172165,
0.017174270004034042,
0.045812927186489105,
0.10991498827934265,
-0.03688102215528488,
-0.04379357397556305,
0.0010193174239248037,
0.03697189688682556,
0.00667230412364006,
0.021859711036086082,
-0.05962761491537094,
-0.007255460135638714,
0.0378965400159359,
0.00908151175826788,
0.07600542157888412,
0.009510851465165615,
-0.12405651062726974,
-0.00984472967684269,
0.0284634567797184,
-0.01943875662982464,
0.015420865267515182,
0.03468289226293564,
0.013664830476045609,
0.0815964862704277,
-0.051135942339897156,
0.042432013899087906,
0.037597816437482834,
0.04030398279428482,
-0.00922357477247715,
-0.04224793240427971,
-0.055296458303928375,
-0.09525731950998306,
-0.08065664768218994,
0.041976235806941986,
-0.09946653246879578,
0.0273415707051754,
0.04069455340504646,
0.03045468032360077,
-0.053493842482566833,
0.08232177048921585,
-0.02772546373307705,
-0.03600626811385155,
-0.01162570621818304,
-0.010698007419705391,
0.02109896019101143,
0.02276129089295864,
-0.06999113410711288,
-0.05656260997056961,
-0.020813701674342155,
0.064567930996418,
0.018915655091404915,
0.06641384214162827,
0.0003657383203972131,
-0.0006616024766117334,
-0.06226455792784691,
-0.036196496337652206,
-0.0677051767706871,
0.029627859592437744,
0.11183688789606094,
0.011001311242580414,
0.0013165243435651064,
-0.016920436173677444,
0.007950742729008198,
-0.019942235201597214,
-0.03396730497479439,
-0.025789279490709305,
-0.021712400019168854,
-0.0014370650751516223,
0.033773038536310196,
0.02836613729596138,
-0.04562051594257355,
-0.045707110315561295,
0.007534158881753683,
-0.02908864989876747,
-0.04845893010497093,
-0.012007412500679493,
0.03397700935602188,
-0.04212114214897156,
-0.002661140402778983,
-0.032137371599674225,
0.04427928850054741,
0.009240275248885155,
-0.04198913276195526,
0.059393297880887985,
0.014941628091037273,
0.11172453314065933,
0.07657650858163834,
0.06777945160865784,
0.029915906488895416,
-0.05468744784593582,
0.038261786103248596,
0.0714055597782135,
0.09724468737840652,
-0.003582334378734231,
-0.04197162389755249,
-3.660711200836886e-8,
-0.1055779904127121,
-0.005220269784331322,
-0.07552354782819748,
0.07400622218847275,
0.03137531876564026,
-0.056037575006484985,
0.036611177027225494,
0.07649704813957214,
-0.04542529955506325,
-0.04476960003376007,
-0.030353190377354622,
-0.023209430277347565,
-0.025557728484272957,
-0.0014812280423939228,
0.06967715173959732,
0.05475426837801933,
-0.015890447422862053,
-0.0033991446252912283,
0.006872810889035463,
-0.040488481521606445,
-0.0012744840933009982,
0.10613798350095749,
0.014577852562069893,
0.022260276600718498,
-0.03209912031888962,
-0.08370815962553024,
-0.004074912518262863,
0.054276466369628906,
0.0716516450047493,
-0.020285913720726967,
0.0742860957980156,
0.0032434130553156137,
0.003347293473780155,
0.050667475908994675,
-0.042221538722515106,
0.04860812798142433,
-0.05175663158297539,
-0.048514317721128464,
-0.020684247836470604,
0.03967691957950592,
-0.00635051354765892,
0.09218881279230118,
-0.07493618130683899,
0.032209962606430054,
0.041898421943187714,
-0.04836525395512581,
0.004039570223540068,
-0.06775207817554474,
0.05868799611926079,
0.04594568908214569,
0.059853192418813705,
-0.03357568383216858,
-0.0029823260847479105,
-0.03802347183227539,
-0.011242964304983616,
0.030072834342718124,
-0.037158042192459106,
-0.08596019446849823,
0.038951270282268524,
0.07971348613500595,
0.04417583718895912,
-0.03988730534911156,
-0.036131732165813446,
0.06299940496683121
] | 0 | 0 | 0 | 0 |
|
Think Then React: Towards Unconstrained Action-to-Reaction Motion Generation | https://openreview.net/forum?id=UxzKcIZedp | [
"Wenhui Tan",
"Boyuan Li",
"Chuhao Jin",
"Wenbing Huang",
"Xiting Wang",
"Ruihua Song"
] | Poster | Modeling human-like action-to-reaction generation has significant real-world applications, like human-robot interaction and games.
Despite recent advancements in single-person motion generation, it is still challenging to well handle action-to-reaction generation, due to the difficulty of directly predicting reaction from action sequence without prompts, and the absence of a unified representation that effectively encodes multi-person motion.
To address these challenges, we introduce Think-Then-React (TTR), a large language-model-based framework designed to generate human-like reactions.
First, with our fine-grained multimodal training strategy, TTR is capable to unify two processes during inference: a thinking process that explicitly infers action intentions and reasons corresponding reaction description, which serve as semantic prompts, and a reacting process that predicts reactions based on input action and the inferred semantic prompts.
Second, to effectively represent multi-person motion in language models, we propose a unified motion tokenizer by decoupling egocentric pose and absolute space features, which effectively represents action and reaction motion with same encoding.
Extensive experiments demonstrate that TTR outperforms existing baselines, achieving significant improvements in evaluation metrics, such as reducing FID from 3.988 to 1.942. | Human Reaction Generation, 3D Human Motion, Large Language Model | null | 13,835 | null | [
-0.04025209695100784,
-0.13413363695144653,
0.06818313896656036,
-0.03902159631252289,
0.08758310228586197,
0.017192207276821136,
0.014344302006065845,
0.05227001756429672,
0.07927066087722778,
-0.020301533862948418,
-0.03414109721779823,
-0.10620859265327454,
-0.021334519609808922,
0.058675266802310944,
0.053936202079057693,
-0.024016469717025757,
0.04018501192331314,
0.06303179264068604,
-0.04593944549560547,
-0.0006373125361278653,
0.05905534327030182,
-0.019308479502797127,
0.08956431597471237,
0.02451837807893753,
-0.0019260572735220194,
0.029290013015270233,
0.03687503933906555,
-0.06979690492153168,
0.024289671331644058,
-0.005834728479385376,
0.04978979006409645,
0.013995288871228695,
0.010856146924197674,
0.013870415277779102,
-0.08792678266763687,
0.10059840232133865,
-0.08140038698911667,
-0.01996993087232113,
-0.056516002863645554,
-0.11619316041469574,
-0.06713200360536575,
-0.016539786010980606,
0.03859758749604225,
0.015929825603961945,
0.09408910572528839,
0.06951697915792465,
0.0028759187553077936,
-0.0694732666015625,
0.015133371576666832,
0.022605981677770615,
-0.0782521516084671,
-0.015487387776374817,
0.059458523988723755,
-0.014707163907587528,
0.014185179956257343,
-0.04975736886262894,
0.050904154777526855,
-0.01984800584614277,
0.07160542905330658,
-0.0614798478782177,
-0.03346272557973862,
-0.10285686701536179,
0.03040899708867073,
-0.03331360965967178,
-0.030237233266234398,
-0.03658212348818779,
0.051305629312992096,
0.04473673179745674,
0.00936766155064106,
0.038406144827604294,
-0.01350035984069109,
-0.0260323379188776,
0.04824626073241234,
-0.014299028553068638,
0.014007918536663055,
0.008242033421993256,
0.013453750871121883,
-0.046247150748968124,
0.03173333406448364,
0.00667447317391634,
0.018193960189819336,
-0.0770951360464096,
0.029860183596611023,
-0.03320848196744919,
0.05768604949116707,
0.0344252735376358,
-0.002266371389850974,
0.08357597887516022,
-0.003967417869716883,
0.05265386402606964,
-0.14665500819683075,
-0.06878647208213806,
0.06476067751646042,
-0.029623938724398613,
-0.0003070762031711638,
0.0312892310321331,
-0.08497639000415802,
-0.06354725360870361,
-0.007567163091152906,
0.045238904654979706,
0.027008242905139923,
0.07988343387842178,
-0.02921142242848873,
-0.008019596338272095,
0.006957514677196741,
-0.0023786562960594893,
0.026757633313536644,
0.018653618171811104,
0.037769753485918045,
-0.033940304070711136,
0.0071942987851798534,
-0.04682276397943497,
-0.01320178247988224,
0.03690456598997116,
0.043827272951602936,
0.00048083686851896346,
0.004596243146806955,
-0.06280449032783508,
0.06293176859617233,
-0.00607266603037715,
0.017930537462234497,
0.005415564402937889,
-0.02372010238468647,
0.04852317273616791,
0.0750439465045929,
-0.06780171394348145,
-0.051169294863939285,
-2.268455957576396e-33,
0.055513378232717514,
0.06044726073741913,
0.12654279172420502,
0.03999404236674309,
0.059708837419748306,
-0.008539646863937378,
-0.010305878706276417,
0.010499212890863419,
-0.016855472698807716,
-0.026713144034147263,
0.016814149916172028,
-0.04545116424560547,
-0.02526066079735756,
0.10583480447530746,
-0.03446025401353836,
0.026159221306443214,
-0.05404604598879814,
-0.02757427841424942,
-0.03861650079488754,
0.04062783718109131,
-0.002985269296914339,
0.05648428946733475,
-0.055846743285655975,
0.014789866283535957,
-0.009210797026753426,
0.046984847635030746,
0.060067009180784225,
-0.01502396259456873,
-0.027368564158678055,
-0.04518386721611023,
-0.11162662506103516,
-0.024221254512667656,
-0.036571402102708817,
0.06900236010551453,
0.06783396005630493,
-0.08876515924930573,
-0.05627783387899399,
-0.028919745236635208,
0.0005117440014146268,
0.0015797923551872373,
-0.01936478540301323,
0.045042186975479126,
-0.05294787138700485,
0.012607205659151077,
-0.06849070638418198,
-0.08480989933013916,
0.020009111613035202,
0.0894058495759964,
-0.06571965664625168,
0.029474053531885147,
0.06645920127630234,
0.060663290321826935,
0.0550139956176281,
-0.06790061295032501,
-0.033282745629549026,
-0.04461710527539253,
0.02815823443233967,
-0.02357783354818821,
0.04207180440425873,
-0.01871340721845627,
-0.041630350053310394,
0.03719738498330116,
0.03888457268476486,
0.01786895841360092,
0.01833263784646988,
0.004590297117829323,
-0.03525147959589958,
0.10861518234014511,
0.09987229108810425,
0.016618825495243073,
-0.021760957315564156,
0.03023652546107769,
0.021891985088586807,
-0.058677732944488525,
0.011220098473131657,
0.05312497168779373,
0.05516821891069412,
-0.10325679928064346,
-0.01573760434985161,
0.09268400818109512,
0.01572866551578045,
-0.06371695548295975,
-0.031966496258974075,
0.03009418398141861,
0.014849018305540085,
-0.03915896266698837,
0.02114471234381199,
-0.12550227344036102,
-0.054420750588178635,
0.011902641505002975,
-0.04158274829387665,
-0.049739543348550797,
-0.011882814578711987,
0.03284502401947975,
-0.049310799688100815,
9.219209338957269e-34,
0.05388091504573822,
0.030309228226542473,
-0.08451904356479645,
0.01912534236907959,
-0.02010815590620041,
-0.05223478376865387,
0.014596252702176571,
-0.010815589688718319,
0.03247124329209328,
-0.04730657488107681,
-0.058460135012865067,
-0.045005518943071365,
0.1112084910273552,
-0.0236176960170269,
0.051826201379299164,
-0.015295526944100857,
0.01692824624478817,
0.02975662425160408,
0.006326287053525448,
0.025694990530610085,
0.035857416689395905,
0.08825187385082245,
-0.08887502551078796,
0.013736889697611332,
-0.07070618867874146,
0.06304961442947388,
0.05557882413268089,
0.013425514101982117,
-0.010260256007313728,
0.01634669303894043,
-0.012668673880398273,
-0.0567140094935894,
0.021307459101080894,
-0.019581938162446022,
-0.020923182368278503,
0.033201683312654495,
0.02451152727007866,
-0.02337845228612423,
0.0151869747787714,
0.02897731401026249,
0.05493665486574173,
-0.07998925447463989,
-0.03934339061379433,
0.07584020495414734,
-0.0912925973534584,
0.004477574024349451,
-0.0707087516784668,
0.027412796393036842,
-0.033769864588975906,
0.08520082384347916,
-0.06232934817671776,
0.06528452038764954,
-0.10281707346439362,
-0.11011454463005066,
-0.05826294422149658,
-0.10898774117231369,
-0.007876304909586906,
-0.056118112057447433,
0.06385272741317749,
-0.02960624173283577,
-0.024746175855398178,
-0.04905995726585388,
0.03644725680351257,
-0.02034817636013031,
-0.006873857229948044,
-0.004512439947575331,
-0.05323639512062073,
-0.00827508233487606,
0.011525807902216911,
-0.013589024543762207,
0.08008502423763275,
0.02497638389468193,
0.021735090762376785,
0.07547510415315628,
0.06069309636950493,
-0.04013950377702713,
0.03751020133495331,
-0.06771144270896912,
0.036686789244413376,
-0.06801335513591766,
-0.023539401590824127,
-0.021876441314816475,
0.04789760708808899,
0.03832084685564041,
0.027715960517525673,
0.08527349680662155,
-0.02838989533483982,
0.14062465727329254,
-0.0025864022318273783,
0.018024994060397148,
-0.028320463374257088,
0.11747013032436371,
0.02485644817352295,
0.08045529574155807,
-0.08370744436979294,
-3.740368370586111e-8,
-0.10754946619272232,
0.019216066226363182,
0.045500244945287704,
0.047266799956560135,
-0.024793753400444984,
-0.0007200272521004081,
0.004051362629979849,
-0.005881277844309807,
-0.0038546882569789886,
-0.004854713566601276,
0.05432674661278725,
0.00028218133957125247,
-0.0034546738024801016,
-0.05273412540555,
0.006214532535523176,
0.03396490588784218,
-0.02613758109509945,
-0.011676552705466747,
-0.05595441162586212,
0.0017501829424872994,
0.023541513830423355,
0.029633162543177605,
-0.049619097262620926,
-0.007309382781386375,
0.02166558988392353,
-0.021101001650094986,
-0.006803093943744898,
0.07961226254701614,
-0.008177259936928749,
-0.05480773001909256,
0.003089473582804203,
-0.014598004519939423,
-0.025116700679063797,
-0.08684566617012024,
-0.0049686855636537075,
0.028237150982022285,
0.07008123397827148,
-0.03234000504016876,
0.009935700334608555,
-0.05778917297720909,
-0.024604378268122673,
0.08222469687461853,
-0.06273142993450165,
0.061368491500616074,
-0.02619398757815361,
0.02679724432528019,
-0.018778346478939056,
-0.166483074426651,
0.02975902147591114,
-0.006187073420733213,
0.0366593673825264,
-0.024587146937847137,
-0.021795041859149933,
0.1034093052148819,
0.06600495427846909,
0.08795809000730515,
-0.008025169372558594,
0.02198023535311222,
0.07676108926534653,
0.06631406396627426,
0.021768877282738686,
0.024419575929641724,
-0.08386855572462082,
-0.015726596117019653
] | 0 | 0 | 0 | 0 |
|
Rapid Selection and Ordering of In-Context Demonstrations via Prompt Embedding Clustering | https://openreview.net/forum?id=1Iu2Yte5N6 | [
"Kha Pham",
"Hung Le",
"Man Ngo",
"Truyen Tran"
] | Poster | While Large Language Models (LLMs) excel at in-context learning (ICL) using just a few demonstrations, their performances are sensitive to demonstration orders. The reasons behind this sensitivity remain poorly understood. In this paper, we investigate the prompt embedding space to bridge the gap between the order sensitivity of ICL with inner workings of decoder-only LLMs, uncovering the clustering property: prompts sharing the first and last demonstrations have closer embeddings, with first-demonstration clustering usually being stronger in practice. We explain this property through extensive theoretical analyses and empirical evidences. Our finding suggests that the positional encoding and the causal attention mask are key contributors to the clustering phenomenon. Leveraging this clustering insight, we introduce Cluster-based Search, a novel method that accelerates the selection and ordering of demonstrations in self-adaptive ICL settings. Our approach substantially decreases the time complexity from factorial to quadratic, saving 92% to nearly 100% execution time while maintaining comparable performance to exhaustive search. | in-context learning, order sensitivity, LLMs, clustering, cluster-based search, positional encoding, attention mask, serial-position effect, cluster-based search | We accelerate selection and ordering of in-context demonstrations in self-adaptive ICL settings by leveraging our newfound clustering property in prompt embedding spaces. | 13,824 | null | [
-0.005059466231614351,
-0.07706286013126373,
0.06309213489294052,
0.020593201741576195,
0.11482738703489304,
0.042591191828250885,
-0.011230440810322762,
-0.01643461361527443,
0.16073322296142578,
-0.028058381751179695,
-0.04586222395300865,
-0.043575137853622437,
0.07077477127313614,
-0.02430742047727108,
-0.017387982457876205,
-0.10499085485935211,
0.12978322803974152,
-0.01276656985282898,
-0.08660576492547989,
-0.08384481072425842,
-0.0012280388036742806,
-0.01721012592315674,
0.11302121728658676,
-0.01989690400660038,
0.028422513976693153,
0.05029908940196037,
-0.02155310846865177,
-0.0013014550786465406,
0.09133210778236389,
0.032961539924144745,
0.034645773470401764,
0.08411046862602234,
0.06091950088739395,
0.03182065486907959,
-0.03549868240952492,
0.0534890778362751,
-0.06181435286998749,
0.03859816864132881,
-0.02450275421142578,
-0.04005757346749306,
-0.027674805372953415,
-0.014904523268342018,
0.027424760162830353,
0.003068427322432399,
0.030778877437114716,
-0.003958896268159151,
-0.001381828566081822,
0.015043361112475395,
-0.010842623189091682,
0.013695118948817253,
-0.06677200645208359,
-0.025460319593548775,
-0.009251680225133896,
-0.04680868983268738,
-0.022825682535767555,
0.03411360830068588,
0.08369536697864532,
0.019480517134070396,
0.025848787277936935,
-0.00816951785236597,
-0.0068208458833396435,
-0.13207389414310455,
0.001723192515783012,
0.02101714164018631,
-0.00802660547196865,
0.030284404754638672,
0.020855797454714775,
0.07406429201364517,
0.022176219150424004,
0.021304665133357048,
0.04558420926332474,
-0.02468010038137436,
0.02389681525528431,
0.08533721417188644,
0.06105663254857063,
0.02431624010205269,
-0.013139143586158752,
-0.10257366299629211,
0.0010493319714441895,
-0.05510132759809494,
-0.00458852993324399,
-0.049589402973651886,
0.02641790173947811,
-0.0024828873574733734,
0.09114900231361389,
0.014578697271645069,
0.0552721843123436,
-0.01696907915174961,
0.01945073902606964,
0.031485047191381454,
-0.05743791162967682,
-0.11607012152671814,
-0.04608950391411781,
0.024223757907748222,
-0.013934767805039883,
-0.02216540277004242,
0.02412228472530842,
-0.07822710275650024,
0.021223288029432297,
0.04438258334994316,
0.02801329642534256,
0.028453078120946884,
0.017693854868412018,
-0.031423211097717285,
-0.09729458391666412,
-0.059740256518125534,
-0.025554126128554344,
0.06732914596796036,
0.07498155534267426,
-0.045203063637018204,
-0.019534964114427567,
0.05285963416099548,
0.06406719237565994,
0.06756258010864258,
0.0002545551396906376,
-0.05384480953216553,
0.059032659977674484,
-0.06487026810646057,
0.07383211702108383,
0.01188461109995842,
0.04520297050476074,
-0.03043944388628006,
0.018332133069634438,
0.01785833016037941,
0.02906476892530918,
-0.0760650485754013,
-0.06563276797533035,
-1.9866327481701873e-34,
0.07816255837678909,
0.030848097056150436,
0.04939240589737892,
0.02140515297651291,
0.01573246717453003,
0.0063233464024960995,
0.06925557553768158,
0.008434321731328964,
0.021397089585661888,
-0.001547671272419393,
0.050974730402231216,
0.007430011872202158,
0.02849484607577324,
0.03965774551033974,
0.002373192925006151,
0.03525039553642273,
-0.03900891914963722,
0.02476688101887703,
-0.13600991666316986,
-0.06522674113512039,
0.004449216648936272,
-0.014471309259533882,
-0.01816978119313717,
-0.09033829718828201,
-0.0018161038169637322,
0.03340495377779007,
0.027490004897117615,
-0.05990694463253021,
-0.020290467888116837,
0.007875475101172924,
-0.10664044320583344,
-0.041071757674217224,
-0.05215868353843689,
0.059808697551488876,
0.04011764004826546,
0.043801963329315186,
0.033964529633522034,
-0.07177270203828812,
0.020310739055275917,
0.01019237656146288,
-0.029399532824754715,
-0.028512656688690186,
0.05790162459015846,
-0.036323897540569305,
-0.06677860021591187,
-0.02760555036365986,
-0.06410016864538193,
-0.012389329262077808,
-0.04639154672622681,
-0.03869608789682388,
0.04635899141430855,
0.05107079818844795,
0.011124969460070133,
-0.05312937870621681,
0.06041840463876724,
0.01587650738656521,
-0.010714094154536724,
0.06168856844305992,
0.02115674875676632,
0.05085856840014458,
-0.011528218165040016,
-0.000699188734870404,
-0.009571079164743423,
0.11338268965482712,
0.027695484459400177,
0.06736353039741516,
-0.058517057448625565,
0.0013687117025256157,
0.06443008780479431,
0.0444234237074852,
-0.061379410326480865,
0.03756484389305115,
-0.041113290935754776,
-0.019043931737542152,
0.07938909530639648,
0.0006119358586147428,
-0.019071754068136215,
-0.0732044130563736,
0.029612919315695763,
0.010867488570511341,
0.03177991509437561,
-0.06377313286066055,
-0.011382754892110825,
-0.005091493017971516,
-0.060398779809474945,
-0.07952218502759933,
0.04799031838774681,
-0.09587366133928299,
-0.019362524151802063,
-0.09016811102628708,
-0.010790265165269375,
-0.022501830011606216,
0.015224885195493698,
0.0772186890244484,
0.011595714837312698,
-1.2049809837641915e-33,
0.033697403967380524,
-0.05122796446084976,
0.013445599935948849,
0.0030747156124562025,
-0.04067585989832878,
-0.04241521656513214,
0.029405683279037476,
-0.0014184769243001938,
-0.05527866631746292,
-0.03940704092383385,
-0.015880970284342766,
-0.022026868537068367,
0.04446689411997795,
0.01787545345723629,
0.018665175884962082,
0.003725687274709344,
0.0682697668671608,
0.03665937855839729,
0.07210660725831985,
0.01690504513680935,
0.00844357255846262,
0.017690783366560936,
-0.029915468767285347,
0.017368294298648834,
-0.03339235112071037,
-0.009152278304100037,
-0.024653427302837372,
0.073712557554245,
-0.02723553031682968,
-0.005477242637425661,
-0.03908594325184822,
0.06411296874284744,
-0.011316534131765366,
0.013304962776601315,
-0.07857578992843628,
0.13480915129184723,
0.00765189016237855,
-0.003912623040378094,
-0.09538933634757996,
0.04301648959517479,
0.010629527270793915,
-0.02639961428940296,
-0.047870613634586334,
0.02673153579235077,
-0.04641231149435043,
-0.0480230413377285,
-0.05356946215033531,
-0.019798900932073593,
0.03733636438846588,
0.02487998828291893,
-0.04705467447638512,
0.07638967037200928,
-0.0028930373955518007,
-0.03626898676156998,
-0.0629548504948616,
-0.02560211345553398,
-0.007309101056307554,
-0.04324471205472946,
0.04086292162537575,
-0.051467590034008026,
-0.016270646825432777,
-0.013973142951726913,
0.08439699560403824,
-0.08128314465284348,
-0.001871539163403213,
-0.0751720666885376,
-0.026556482538580894,
0.031041942536830902,
0.00355910649523139,
-0.057724837213754654,
0.03203780576586723,
0.0016691596247255802,
-0.02984483540058136,
-0.008671073243021965,
0.001807689550332725,
-0.015769170597195625,
0.03864995390176773,
-0.037110619246959686,
-0.0712139680981636,
-0.07908429950475693,
0.004468736704438925,
0.011396795511245728,
0.03802573308348656,
0.06586121022701263,
0.030770527198910713,
0.12301985174417496,
-0.012492409907281399,
0.14601370692253113,
-0.010530047118663788,
0.034174371510744095,
-0.0016797746066004038,
-0.01299248170107603,
0.05252012982964516,
0.07283597439527512,
-0.04317285865545273,
-4.17242560502018e-8,
-0.16050982475280762,
0.018901579082012177,
0.025332024320960045,
0.004373874049633741,
0.04070745036005974,
-0.004781546536833048,
-0.03781169280409813,
0.0525602251291275,
-0.05205971747636795,
-0.08296553790569305,
0.009168482385575771,
0.02633870765566826,
-0.027024725452065468,
-0.01587793231010437,
-0.010505718179047108,
0.09026961028575897,
-0.015789510682225227,
-0.04806704446673393,
-0.013141369447112083,
-0.04649927839636803,
0.009022886864840984,
0.043804582208395004,
0.030509157106280327,
-0.04499290883541107,
0.007930253632366657,
-0.037934720516204834,
0.010681051760911942,
0.09439908713102341,
0.015891699120402336,
-0.04507165774703026,
-0.06016348674893379,
0.003527154680341482,
-0.04137025773525238,
0.002117884811013937,
0.026877205818891525,
0.10688646137714386,
0.02845611982047558,
-0.044981326907873154,
-0.037899550050497055,
-0.07599090039730072,
0.010393924079835415,
-0.023027490824460983,
-0.02686372399330139,
0.035406287759542465,
0.05997796729207039,
0.03186194226145744,
0.01698850840330124,
-0.17132364213466644,
0.03002200275659561,
-0.13599909842014313,
0.03545384481549263,
-0.09209868311882019,
-0.060887157917022705,
0.057376429438591,
0.08445481210947037,
0.12463577836751938,
-0.07239248603582382,
0.018757658079266548,
0.018697666004300117,
0.1022881343960762,
0.0396759994328022,
0.05625006556510925,
-0.06206188350915909,
-0.025806095451116562
] | 0 | 0 | 0 | 0 |
|
Asymptotic Analysis of Two-Layer Neural Networks after One Gradient Step under Gaussian Mixtures Data with Structure | https://openreview.net/forum?id=tNn6Hskmti | [
"Samet Demir",
"Zafer Dogan"
] | Poster | In this work, we study the training and generalization performance of two-layer neural networks (NNs) after one gradient descent step under structured data modeled by Gaussian mixtures. While previous research has extensively analyzed this model under isotropic data assumption, such simplifications overlook the complexities inherent in real-world datasets. Our work addresses this limitation by analyzing two-layer NNs under Gaussian mixture data assumption in the asymptotically proportional limit, where the input dimension, number of hidden neurons, and sample size grow with finite ratios. We characterize the training and generalization errors by leveraging recent advancements in Gaussian universality. Specifically, we prove that a high-order polynomial model performs equivalent to the non-linear neural networks under certain conditions. The degree of the equivalent model is intricately linked to both the "data spread" and the learning rate employed during one gradient step. Through extensive simulations, we demonstrate the equivalence between the original model and its polynomial counterpart across various regression and classification tasks. Additionally, we explore how different properties of Gaussian mixtures affect learning outcomes. Finally, we illustrate experimental results on Fashion-MNIST classification, indicating that our findings can translate to realistic data. | deep learning theory, random features, Gaussian equivalence, universality, high-dimensional asymptotics | We study the impacts of Gaussian mixtures data assumption to feature learning in neural networks trained with one gradient step in order to bridge the gap between isotropic data assumption and real datasets. | 13,819 | 2503.00856 | [
-0.05726712569594383,
-0.14971768856048584,
0.06948845088481903,
0.06783599406480789,
0.07747609913349152,
0.016944631934165955,
-0.03140394017100334,
-0.06308829039335251,
0.012994781136512756,
-0.12363434582948685,
0.07320225983858109,
-0.009920349344611168,
0.033366844058036804,
0.023649858310818672,
-0.03521217405796051,
-0.05200530216097832,
0.05710548907518387,
0.07374896109104156,
-0.11492615193128586,
0.0017538049723953009,
0.027961691841483116,
0.0007842393242754042,
-0.00479529146105051,
-0.02006981335580349,
0.09014736115932465,
0.006299992091953754,
-0.03051825799047947,
-0.03128429874777794,
0.002717849100008607,
0.007688706275075674,
0.07739462703466415,
-0.07849214226007462,
0.01306107547134161,
0.07469776272773743,
-0.0687447115778923,
-0.008473682217299938,
-0.013016019016504288,
0.1012994572520256,
0.012298673391342163,
0.012125262059271336,
0.07369764894247055,
-0.008053626865148544,
-0.016645271331071854,
0.08418098092079163,
0.09251237660646439,
0.022042347118258476,
-0.004026209469884634,
-0.09359896928071976,
0.00153568375390023,
-0.016739284619688988,
0.01297968439757824,
-0.004900304134935141,
-0.027715902775526047,
0.10425876826047897,
-0.011995023116469383,
-0.009119894355535507,
0.008905363269150257,
0.10803581774234772,
-0.07682973146438599,
0.07678305357694626,
-0.018879041075706482,
-0.07892431318759918,
-0.030140584334731102,
-0.010276573710143566,
0.04648752510547638,
0.00871272198855877,
0.027364570647478104,
0.0072756255976855755,
-0.003174074459820986,
0.10269513726234436,
0.038569364696741104,
0.12418273091316223,
-0.08828775584697723,
-0.010484426282346249,
0.01624072715640068,
-0.026751743629574776,
0.001504214247688651,
0.051838818937540054,
-0.03639119118452072,
-0.08693188428878784,
-0.004794360604137182,
0.00851425901055336,
0.007867313921451569,
-0.05809294432401657,
0.05440526083111763,
-0.004531179554760456,
0.0029281130991876125,
0.08703365921974182,
-0.010176914744079113,
-0.0388273261487484,
-0.07986240833997726,
0.06614157557487488,
-0.009502715431153774,
-0.026599865406751633,
0.07705036550760269,
0.00869451928883791,
-0.018736395984888077,
-0.0388234443962574,
0.042642150074243546,
0.06335553526878357,
-0.04545934125781059,
-0.00667209830135107,
0.03994369134306908,
-0.030907858163118362,
0.0778074711561203,
0.006873763632029295,
0.01810147613286972,
0.010840951465070248,
0.0818287804722786,
-0.05687082186341286,
0.03856221213936806,
0.0369284488260746,
0.0050451927818357944,
0.007782513741403818,
0.00907512754201889,
-0.01863880455493927,
0.10902433097362518,
-0.016204340383410454,
-0.0356217660009861,
0.07794472575187683,
-0.03823316469788551,
0.03630468249320984,
-0.039103440940380096,
0.020321698859333992,
-0.007335224188864231,
-0.014562878757715225,
-0.0923374742269516,
3.5577442227010274e-33,
-0.013127210550010204,
0.008510191924870014,
0.01583784818649292,
-0.00437456090003252,
0.01663505658507347,
-0.0015739677473902702,
0.06704144924879074,
0.012218193151056767,
0.08896021544933319,
0.09703466296195984,
-0.04678478464484215,
0.030763089656829834,
-0.07940909266471863,
0.11051425337791443,
0.04935980588197708,
0.07356959581375122,
-0.01802005246281624,
0.015001211315393448,
0.08915594220161438,
-0.07269862294197083,
0.07506142556667328,
-0.03726071119308472,
-0.0021203223150223494,
0.021026045083999634,
-0.048106223344802856,
-0.024523522704839706,
0.034430842846632004,
-0.0034411491360515356,
-0.025974148884415627,
-0.014629808254539967,
-0.0370025709271431,
0.0008427326683886349,
-0.04105600714683533,
0.0034320210106670856,
0.053424835205078125,
-0.011911618523299694,
0.06950024515390396,
-0.020455826073884964,
0.04300909861922264,
-0.039515603333711624,
-0.024229852482676506,
0.031129274517297745,
0.07183050364255905,
-0.027391009032726288,
-0.08883116394281387,
-0.10326452553272247,
0.008523305878043175,
0.020276686176657677,
0.02350943349301815,
-0.08482835441827774,
-0.03533795103430748,
0.01656581275165081,
-0.035856518894433975,
-0.022937657311558723,
0.03629302605986595,
0.02562674693763256,
-0.007489388342946768,
0.06999173015356064,
0.0173636507242918,
0.037172041833400726,
0.01488199271261692,
0.0020652576349675655,
0.006322827655822039,
0.033385373651981354,
-0.02310279943048954,
0.031130854040384293,
-0.03926555812358856,
0.04431740567088127,
-0.0008304226212203503,
0.0012978189624845982,
0.018623612821102142,
0.017374346032738686,
-0.07354963570833206,
-0.0532853901386261,
0.07445277273654938,
0.010296790860593319,
-0.029798636212944984,
-0.14257195591926575,
-0.013428018428385258,
0.02391190640628338,
-0.08183515071868896,
0.02830994315445423,
-0.05926569178700447,
-0.06198662519454956,
-0.09834105521440506,
0.019510354846715927,
0.03182248771190643,
0.006562874652445316,
-0.051990482956171036,
0.01648815907537937,
-0.09837664663791656,
-0.0015459104906767607,
-0.04461877793073654,
-0.034976307302713394,
0.024231085553765297,
-3.6973925842367305e-33,
-0.08394745737314224,
0.04799025133252144,
-0.040218621492385864,
0.04586699977517128,
0.0066512273624539375,
0.01834261417388916,
-0.019262127578258514,
0.05122573301196098,
-0.04073787480592728,
0.03351207822561264,
0.049636490643024445,
0.02585131861269474,
0.05765007808804512,
0.024341417476534843,
0.043763987720012665,
-0.002151690423488617,
-0.05090754106640816,
0.12848110496997833,
0.10253602266311646,
-0.052180953323841095,
0.019238242879509926,
0.03592268377542496,
-0.09693670272827148,
0.07745437324047089,
-0.04372940585017204,
-0.018331436440348625,
-0.08386082202196121,
0.11011742055416107,
0.014489226043224335,
0.0354694165289402,
-0.08293742686510086,
-0.016703538596630096,
-0.05710819363594055,
-0.018257271498441696,
-0.0244178157299757,
0.07827004790306091,
0.0331336185336113,
-0.0005820172955282032,
-0.05054882541298866,
-0.01857117749750614,
-0.010498631745576859,
-0.0459679514169693,
-0.04655227065086365,
-0.009004123508930206,
0.053380295634269714,
-0.1023784652352333,
-0.005224245134741068,
-0.04053107649087906,
-0.03248993679881096,
0.01588127389550209,
-0.01569875329732895,
0.06041404604911804,
-0.11539582908153534,
0.03181733563542366,
-0.046711526811122894,
0.012473254464566708,
-0.04442579671740532,
-0.005628181155771017,
0.02566344290971756,
-0.06146320700645447,
0.0037160757929086685,
-0.05354666709899902,
-0.07105126231908798,
-0.028732914477586746,
-0.05008504167199135,
-0.07170948386192322,
0.003545180894434452,
0.06453749537467957,
0.036429621279239655,
0.06499692052602768,
0.0008692163391970098,
-0.022959904745221138,
0.05302463099360466,
0.02960273250937462,
-0.09471169859170914,
-0.017406431958079338,
0.0666503980755806,
-0.014428233727812767,
-0.023737821727991104,
0.01012787502259016,
0.08333680033683777,
0.03936358541250229,
0.012065342627465725,
0.000464748649392277,
0.10851992666721344,
0.10231824219226837,
0.0757036805152893,
0.06124202907085419,
-0.009953228756785393,
-0.0610664077103138,
0.008258147165179253,
-0.013384872116148472,
-0.016731208190321922,
0.029854563996195793,
-0.058578915894031525,
-4.5613990806714355e-8,
-0.039490748196840286,
0.051268722862005234,
0.056963350623846054,
-0.0349460244178772,
0.05594887584447861,
-0.06481696665287018,
-0.012398114427924156,
0.06373923271894455,
-0.061305806040763855,
-0.008417194709181786,
0.052812427282333374,
0.022941257804632187,
-0.013008029200136662,
-0.04655436426401138,
0.01228063739836216,
0.08893951028585434,
-0.035513848066329956,
-0.004401535727083683,
0.03711753711104393,
-0.06623794883489609,
-0.02952711097896099,
0.002145078731700778,
-0.022536924108862877,
-0.06953002512454987,
0.032491590827703476,
-0.07900744676589966,
-0.008184447884559631,
0.11767985671758652,
0.027596257627010345,
0.00820282381027937,
0.0029754454735666513,
-0.025606367737054825,
-0.060678355395793915,
-0.032414793968200684,
0.02401326410472393,
0.038022588938474655,
0.06346362084150314,
-0.00923895463347435,
0.008186628110706806,
-0.07418911159038544,
-0.048024095594882965,
-0.0024839849211275578,
0.060950152575969696,
0.007248324807733297,
0.0497121624648571,
0.0035268834326416254,
0.017362559214234352,
-0.10914785414934158,
-0.021964022889733315,
0.005578251089900732,
0.07611652463674545,
0.00023990344197954983,
-0.01812286302447319,
0.061617132276296616,
0.05591674521565437,
-0.07612981647253036,
-0.06554180383682251,
-0.08778446167707443,
-0.024912124499678612,
0.028831040486693382,
-0.0030371947214007378,
0.017326720058918,
-0.0644744262099266,
-0.01053075585514307
] | https://github.com/KU-MLIP/2-Layer-NNs-with-Gaussian-Mixtures-Data | 0 | 0 | 0 | 0 |
OCEAN: Offline Chain-of-thought Evaluation and Alignment in Large Language Models | https://openreview.net/forum?id=rlgplAuN2p | [
"Junda Wu",
"Xintong Li",
"Ruoyu Wang",
"Yu Xia",
"Yuxin Xiong",
"Jianing Wang",
"Tong Yu",
"Xiang Chen",
"Branislav Kveton",
"Lina Yao",
"Jingbo Shang",
"Julian McAuley"
] | Poster | Offline evaluation of LLMs is crucial in understanding their capacities, though current methods remain underexplored in existing research. In this work, we focus on the offline evaluation of the chain-of-thought capabilities and show how to optimize LLMs based on the proposed evaluation method. To enable offline feedback with rich knowledge and reasoning paths, we use knowledge graphs (KGs) (e.g., Wikidata5M) to provide feedback on the generated chain of thoughts. Due to the heterogeneity between LLM reasoning and KG structures, direct interaction and feedback from knowledge graphs on LLM behavior are challenging, as they require accurate entity linking and grounding of LLM-generated chains of thought in the KG. To address the above challenge, we propose an offline chain-of-thought evaluation framework, OCEAN, which models chain-of-thought reasoning in LLMs as a Markov Decision Process (MDP), and evaluate the policy’s alignment with KG preference modeling. To overcome the reasoning heterogeneity and grounding problems, we leverage on-policy KG exploration and reinforcement learning to model a KG policy that generates token-level likelihood distributions for LLM-generated chain-of-thought reasoning paths, simulating KG reasoning preference. Then we incorporate the knowledge-graph feedback on the validity and alignment of the generated reasoning paths into inverse propensity scores and propose KG-IPS estimator. Theoretically, we prove the unbiasedness of the proposed KG-IPS estimator and provide a lower bound on its variance. With the off-policy evaluated value function, we can directly enable off-policy optimization to further enhance chain-of-thought alignment. Our empirical study shows that OCEAN can be efficiently optimized for generating chain-of-thought reasoning paths with higher estimated values without affecting LLMs’ general abilities in downstream tasks or their internal knowledge. | chain-of-thought, large language models, offline policy evaluation, agentic | null | 13,812 | 2410.23703 | [
-0.029236232861876488,
-0.042078275233507156,
0.03282322734594345,
-0.02987612597644329,
0.03304777666926384,
-0.041763536632061005,
0.01266450248658657,
0.031011633574962616,
0.043801199644804,
0.07104091346263885,
-0.04894436150789261,
-0.040065232664346695,
0.09259472787380219,
0.03305448219180107,
0.07041707634925842,
0.09851421415805817,
0.06977064162492752,
-0.018668334931135178,
-0.09538570791482925,
-0.04139363765716553,
0.0726683959364891,
0.02536773309111595,
0.013447683304548264,
-0.020918531343340874,
-0.04919210076332092,
-0.031490541994571686,
0.013842670246958733,
0.03231311962008476,
-0.009392095729708672,
0.04288625344634056,
0.004949299152940512,
0.04778280109167099,
-0.019555222243070602,
-0.01591385528445244,
0.0009910756489261985,
0.0903191938996315,
-0.10313557833433151,
0.021425137296319008,
-0.05150432512164116,
0.01246853917837143,
-0.05578143522143364,
-0.023141996935009956,
0.008449826389551163,
0.020558610558509827,
0.041589025408029556,
0.033528849482536316,
-0.03941939026117325,
-0.014609990641474724,
-0.05367414653301239,
0.06691579520702362,
-0.09299160540103912,
-0.04222534969449043,
0.0008876014617271721,
-0.029471980407834053,
0.057168252766132355,
0.05422337353229523,
-0.03145384043455124,
0.04189636558294296,
0.02632431499660015,
-0.11720243841409683,
-0.033332936465740204,
-0.10593678802251816,
-0.04192657768726349,
-0.03987335041165352,
-0.046290017664432526,
-0.03280939161777496,
-0.018074562773108482,
0.09151861071586609,
-0.020307479426264763,
0.019713807851076126,
-0.010811557993292809,
0.0625365823507309,
-0.0067313555628061295,
0.020119810476899147,
-0.056153830140829086,
0.037426918745040894,
0.03349781036376953,
-0.03167751058936119,
-0.008472226560115814,
-0.07076861709356308,
-0.027841519564390182,
0.0875243991613388,
0.004595817066729069,
0.004622862208634615,
0.07603387534618378,
-0.0037604488898068666,
0.027762413024902344,
0.08922076225280762,
0.10186146199703217,
0.0414372980594635,
-0.02283598855137825,
-0.05384920910000801,
0.07612430304288864,
-0.00106830894947052,
0.07650835067033768,
0.10983334481716156,
-0.03841264173388481,
-0.04495019093155861,
-0.04921308904886246,
0.09007430076599121,
0.06621675938367844,
0.11639849096536636,
0.026732079684734344,
-0.08622913062572479,
0.012655362486839294,
-0.04973169043660164,
-0.02689017355442047,
0.10992588102817535,
0.025699682533740997,
-0.09167306125164032,
-0.018419606611132622,
0.0688609778881073,
0.05567853897809982,
0.026492806151509285,
-0.04619987681508064,
-0.0294018667191267,
0.09826799482107162,
-0.0902327224612236,
0.04170291870832443,
-0.025838157162070274,
0.014319061301648617,
-0.02155313454568386,
0.06307794898748398,
-0.032126035541296005,
0.0037262130063027143,
-0.04942139610648155,
0.01813599094748497,
-1.4825374241834038e-33,
0.059231072664260864,
0.017429420724511147,
0.12043172866106033,
-0.0021330025047063828,
0.05081210657954216,
-0.04282940924167633,
-0.000798675580881536,
0.013581302016973495,
-0.09886778146028519,
-0.002896697260439396,
-0.08907151222229004,
0.024922862648963928,
-0.0022694061044603586,
-0.014771130867302418,
0.02796570584177971,
0.026592368260025978,
-0.011971239931881428,
0.010790483094751835,
-0.03379801660776138,
-0.02060210518538952,
0.08478257060050964,
0.01651414856314659,
0.03196582943201065,
-0.033768609166145325,
0.05656563863158226,
-0.006315040402114391,
0.01453500334173441,
-0.024820048362016678,
-0.04521125927567482,
-0.004801315255463123,
-0.021831385791301727,
-0.030069228261709213,
-0.0350639708340168,
0.030557112768292427,
-0.03697878494858742,
-0.02700020745396614,
-0.09012295305728912,
-0.04802233353257179,
-0.0022705940064042807,
-0.1072646751999855,
0.0019517269684001803,
0.023941215127706528,
0.020940037444233894,
0.08710639178752899,
-0.07784061878919601,
-0.010846647433936596,
0.04674004763364792,
-0.033847734332084656,
0.024955619126558304,
-0.03613976016640663,
0.0931876078248024,
-0.0042108497582376,
0.008947065100073814,
-0.026394661515951157,
-0.04150645434856415,
-0.025012245401740074,
-0.009308482520282269,
0.09162438660860062,
-0.013006589375436306,
0.02505188062787056,
-0.02757134661078453,
0.041326019912958145,
0.04463697224855423,
0.07804747670888901,
0.052618686109781265,
0.043210189789533615,
-0.11994507908821106,
0.018656887114048004,
0.0730351135134697,
-0.11128035932779312,
0.00929860956966877,
0.05619942396879196,
0.08406614512205124,
-0.06775988638401031,
0.004442252218723297,
-0.018092334270477295,
-0.0727200135588646,
-0.10101612657308578,
0.020005322992801666,
0.06152739375829697,
0.0413275882601738,
-0.03932517394423485,
-0.08875921368598938,
-0.04379069060087204,
-0.004091646522283554,
-0.05057445168495178,
0.04497913271188736,
-0.07536367326974869,
0.017355162650346756,
-0.014328845776617527,
-0.0733649879693985,
-0.04656902328133583,
-0.03208528086543083,
0.006021740846335888,
0.04383833333849907,
2.3735724850794525e-35,
-0.04348010569810867,
-0.04439995437860489,
0.08970708400011063,
0.036145731806755066,
-0.009783515706658363,
-0.04235897213220596,
0.023434022441506386,
0.022773824632167816,
-0.058280572295188904,
-0.029272114858031273,
-0.05784514173865318,
-0.07104392349720001,
0.001489462098106742,
0.00748645793646574,
0.020650800317525864,
-0.02768474817276001,
-0.053087517619132996,
-0.014826346188783646,
0.028347181156277657,
0.097635418176651,
-0.003756717313081026,
0.03281494602560997,
-0.13923224806785583,
0.09797625243663788,
0.01381171215325594,
0.0530458427965641,
0.04528287425637245,
0.039415355771780014,
-0.028683694079518318,
0.05728688836097717,
-0.04665243253111839,
-0.0611145980656147,
-0.04473399743437767,
-0.008660094812512398,
-0.09829394519329071,
0.028835779055953026,
0.01712569035589695,
-0.016186214983463287,
-0.08021572232246399,
0.04493214935064316,
-0.015670787543058395,
-0.07082820683717728,
-0.021830011159181595,
-0.08345407992601395,
-0.08842630684375763,
0.057747140526771545,
-0.07313135266304016,
0.03967231884598732,
0.007513466756790876,
-0.07035461813211441,
-0.01503362413495779,
0.04387972131371498,
-0.028966257348656654,
0.036502301692962646,
0.007534665986895561,
-0.025968480855226517,
-0.029620610177516937,
0.000933838018681854,
-0.0099312923848629,
-0.004843563307076693,
-0.012864603661000729,
-0.01976529136300087,
0.027026858180761337,
-0.021229688078165054,
0.015904659405350685,
-0.00006752842455171049,
-0.004354504402726889,
0.04079584777355194,
-0.023640433326363564,
-0.016499124467372894,
-0.023555738851428032,
0.0052229976281523705,
-0.027689572423696518,
-0.05297549068927765,
0.1548398733139038,
-0.018117258325219154,
0.02577885426580906,
-0.05634783208370209,
-0.015026788227260113,
-0.08521170914173126,
0.037360768765211105,
-0.006483678240329027,
0.027516042813658714,
0.0649004727602005,
0.06875447183847427,
-0.0459006205201149,
0.056643497198820114,
0.09060318022966385,
0.0036312593147158623,
-0.0059135910123586655,
-0.04662704840302467,
-0.03125800937414169,
-0.03080693632364273,
0.08387622982263565,
-0.08451247960329056,
-3.9221209391371303e-8,
-0.08792704343795776,
-0.030903730541467667,
0.09889689832925797,
0.07888947427272797,
0.03266678750514984,
-0.01982611045241356,
0.02141452766954899,
0.07471951842308044,
0.01588442362844944,
-0.002773632761090994,
0.0644138902425766,
0.044168870896101,
-0.07583734393119812,
-0.06887052953243256,
-0.004977433942258358,
0.007932324893772602,
0.023188801482319832,
-0.018725518137216568,
0.0058243172243237495,
0.013207790441811085,
0.03517040237784386,
-0.019260231405496597,
-0.041449591517448425,
0.06830911338329315,
0.06743143498897552,
-0.03329925239086151,
-0.03138938918709755,
0.12034487724304199,
0.03434120863676071,
-0.03879629448056221,
0.004828485660254955,
-0.045838840305805206,
0.01727166958153248,
0.03298419713973999,
0.06795860826969147,
0.014367773197591305,
-0.008336781524121761,
-0.004311066586524248,
0.04049922898411751,
-0.04555797204375267,
0.04983925446867943,
0.05869697779417038,
-0.0792798101902008,
0.04851105064153671,
0.05585501715540886,
-0.009534121491014957,
-0.12433487921953201,
-0.05173913389444351,
-0.010982215404510498,
-0.018568964675068855,
0.016164451837539673,
-0.05100128427147865,
0.001948830089531839,
0.005400948692113161,
0.05873071402311325,
0.014605877920985222,
-0.004890163894742727,
0.002227046526968479,
0.017353424802422523,
0.07318621873855591,
0.06774818152189255,
-0.02075894922018051,
-0.044760022312402725,
0.0476452112197876
] | 0 | 0 | 0 | 0 |
|
Distribution-Free Data Uncertainty for Neural Network Regression | https://openreview.net/forum?id=pDDODPtpx9 | [
"Domokos M. Kelen",
"Ádám Jung",
"Péter Kersch",
"Andras A Benczur"
] | Poster | Quantifying uncertainty is an essential part of predictive modeling, especially in the context of high-stakes decision-making. While classification output includes data uncertainty by design in the form of class probabilities, the regression task generally aims only to predict the expected value of the target variable. Probabilistic extensions often assume parametric distributions around the expected value, optimizing the likelihood over the resulting explicit densities. However, using parametric distributions can limit practical applicability, making it difficult for models to capture skewed, multi-modal, or otherwise complex distributions. In this paper, we propose optimizing a novel nondeterministic neural network regression architecture for loss functions derived from a sample-based approximation of the continuous ranked probability score (CRPS), enabling a truly distribution-free approach by learning to sample from the target's aleatoric distribution, rather than predicting explicit densities. Our approach allows the model to learn well-calibrated, arbitrary uni- and multivariate output distributions. We evaluate the method on a variety of synthetic and real-world tasks, including uni- and multivariate problems, function inverse approximation, and standard regression uncertainty benchmarks. Finally, we make all experiment code publicly available. | deep learning, uncertainty quantification, regression uncertainty, aleatoric uncertainty, scoring rules, continuous ranked probability score | We propose a distribution-free neural network regression approach that learns aleatoric uncertainty through sample-based CRPS optimization. | 13,804 | null | [
-0.08733492344617844,
-0.08407498896121979,
0.018770672380924225,
0.03355725482106209,
0.033634405583143234,
0.032816097140312195,
0.001146017573773861,
0.04306219145655632,
0.040199585258960724,
-0.028377939015626907,
-0.021723207086324692,
-0.08271288126707077,
0.07383311539888382,
0.027686113491654396,
-0.11912021785974503,
-0.04599757492542267,
0.029698451980948448,
0.08346406370401382,
-0.05236831679940224,
0.009157592430710793,
0.048496730625629425,
0.06937433779239655,
-0.039575815200805664,
0.02697698213160038,
0.0902881920337677,
-0.07624678313732147,
0.00016473198775202036,
0.011535268276929855,
-0.10082076489925385,
0.010058779269456863,
0.07511238753795624,
-0.06653596460819244,
-0.04235794395208359,
-0.011213205754756927,
-0.03182929381728172,
0.0023438171483576298,
-0.0834207609295845,
0.025854747742414474,
0.08473976701498032,
-0.03745679929852486,
0.030303608626127243,
-0.003956171218305826,
-0.0025236725341528654,
0.041289322078228,
0.05837308615446091,
-0.01990920677781105,
0.015463818795979023,
-0.1059260293841362,
0.001786277280189097,
0.09915776550769806,
0.0057900394313037395,
0.005066379904747009,
-0.006416622549295425,
-0.017199812456965446,
-0.008655844256281853,
0.026713065803050995,
-0.0055476101115345955,
0.0017125556478276849,
-0.02805141545832157,
0.012927569448947906,
-0.0400938056409359,
-0.11359506100416183,
0.0210314579308033,
-0.013699514791369438,
-0.021949902176856995,
-0.0020145438611507416,
-0.03565194085240364,
0.0684179738163948,
0.014740177430212498,
0.02426258660852909,
-0.011386443860828876,
0.08646027743816376,
-0.08671937137842178,
0.05977654084563255,
0.04324881359934807,
-0.022072259336709976,
0.11045772582292557,
-0.006295090541243553,
-0.03542647138237953,
0.022208141162991524,
0.014644944109022617,
0.05261934921145439,
0.00107712356839329,
0.032257020473480225,
0.17703863978385925,
0.07266254723072052,
0.0162628386169672,
0.11157253384590149,
-0.004100295249372721,
-0.010497979819774628,
-0.0815393328666687,
0.014690546318888664,
0.013890111818909645,
-0.0020635435357689857,
0.017010727897286415,
0.0652356669306755,
-0.03095400333404541,
-0.14019858837127686,
-0.003077558008953929,
0.07920774072408676,
-0.03300030529499054,
0.07102574408054352,
0.017807167023420334,
-0.08842699229717255,
0.012448170222342014,
-0.05630747228860855,
-0.0008309738477692008,
0.0064661577343940735,
-0.02575533837080002,
-0.08667033165693283,
-0.007665691897273064,
-0.028625771403312683,
-0.008291346952319145,
-0.054371174424886703,
-0.0042341952212154865,
0.03212923929095268,
-0.025011520832777023,
0.021140584722161293,
-0.015305301174521446,
0.0426640510559082,
-0.09977353364229202,
0.01902967132627964,
0.004139882978051901,
0.021188674494624138,
0.018734022974967957,
0.031353700906038284,
-0.05649052932858467,
3.5531436316854966e-33,
-0.020064381882548332,
-0.041121646761894226,
0.03856746852397919,
-0.010465477593243122,
0.039341576397418976,
-0.011610516346991062,
-0.021364860236644745,
0.048268068581819534,
0.02444181777536869,
0.020636681467294693,
-0.016107697039842606,
0.003508311230689287,
-0.08268101513385773,
0.017828956246376038,
0.08561281859874725,
0.009124833159148693,
-0.04318615049123764,
0.04816615954041481,
0.009759771637618542,
-0.011414939537644386,
-0.018739568069577217,
-0.03366916999220848,
-0.022151101380586624,
-0.023400403559207916,
-0.002272492041811347,
0.015146375633776188,
0.06256067007780075,
-0.01330060325562954,
-0.049707669764757156,
-0.010206331498920918,
-0.07992082834243774,
0.031283602118492126,
0.03810811787843704,
-0.11493618786334991,
0.041422080248594284,
-0.05720202624797821,
-0.061486631631851196,
0.0022242700215429068,
-0.05352180823683739,
0.02336772158741951,
-0.02539358288049698,
-0.010790922679007053,
0.0312524288892746,
0.0830598846077919,
-0.054385919123888016,
-0.08031274378299713,
0.04549264907836914,
-0.042331863194704056,
0.00214822287671268,
-0.09921740740537643,
-0.007797271013259888,
0.011083469726145267,
0.028618060052394867,
-0.038926709443330765,
-0.056886907666921616,
-0.020141540095210075,
0.02767830528318882,
-0.0034113749861717224,
-0.025158805772662163,
-0.005874237045645714,
-0.11725710332393646,
0.04420923441648483,
-0.02085425890982151,
0.024258116260170937,
0.0009245065157301724,
0.03193620592355728,
-0.06380350142717361,
0.04608859866857529,
0.04598057270050049,
0.005180235952138901,
0.047569938004016876,
0.04786849021911621,
0.03268891200423241,
-0.027355656027793884,
0.05536545068025589,
0.08484325557947159,
0.04406236112117767,
-0.04701601341366768,
0.0317871980369091,
-0.004113574977964163,
0.05980145186185837,
-0.01843986101448536,
-0.035054922103881836,
-0.010999959893524647,
0.05494484305381775,
-0.012517448514699936,
0.07496225088834763,
0.039543963968753815,
0.017810262739658356,
0.004546749405562878,
-0.09209546446800232,
0.003675702726468444,
-0.0566113255918026,
0.003276607021689415,
-0.029438013210892677,
-1.911371630523294e-33,
-0.11696584522724152,
0.07312876731157303,
-0.05095207318663597,
0.00983355287462473,
0.03093278780579567,
0.024847114458680153,
-0.043792665004730225,
0.02482997067272663,
0.05478135123848915,
-0.049830202013254166,
-0.08350946009159088,
-0.029136329889297485,
0.03415590524673462,
0.028856651857495308,
0.03879889100790024,
0.030566347762942314,
-0.09180667251348495,
0.03997143357992172,
-0.029792597517371178,
-0.026765162125229836,
0.051895420998334885,
0.09947973489761353,
-0.053201932460069656,
0.05415600165724754,
-0.0674758106470108,
0.03697420284152031,
-0.06200381740927696,
0.12727226316928864,
-0.010100537911057472,
0.004500267561525106,
-0.018648266792297363,
-0.046632055193185806,
-0.0038237806875258684,
-0.02766554243862629,
-0.014167285524308681,
0.0032429469283670187,
0.01348591223359108,
-0.025705687701702118,
-0.0730004608631134,
0.022592680528759956,
0.0014672302640974522,
-0.014760826714336872,
-0.09003322571516037,
-0.09458030015230179,
0.0399523489177227,
0.00851060263812542,
0.04034185782074928,
-0.05477564036846161,
0.0277464147657156,
0.06750766932964325,
-0.07091294974088669,
0.06830048561096191,
-0.11290708929300308,
0.07266455143690109,
0.004518226720392704,
-0.06135513260960579,
-0.10351550579071045,
-0.040208764374256134,
0.06990872323513031,
0.103702612221241,
-0.006106541957706213,
-0.020508790388703346,
0.046420883387327194,
0.023569179698824883,
-0.05525285750627518,
0.005859729368239641,
-0.03943805396556854,
0.0771486908197403,
0.001315005705691874,
0.06388552486896515,
-0.03224417194724083,
0.00496545247733593,
0.06953532993793488,
-0.013460959307849407,
-0.06605198234319687,
-0.06238807737827301,
0.005724998656660318,
0.07643341273069382,
0.02405235357582569,
0.0887821838259697,
0.13231723010540009,
-0.007898484356701374,
0.034652601927518845,
0.06568975746631622,
0.11615067720413208,
0.04047255590558052,
0.043257225304841995,
0.032813508063554764,
0.11841613799333572,
-0.008513382636010647,
0.018141625449061394,
0.05240919813513756,
0.02266956865787506,
0.022159812971949577,
0.018176905810832977,
-4.793666619207215e-8,
-0.029619678854942322,
0.10707132518291473,
0.07663189619779587,
0.04661833494901657,
0.046185120940208435,
0.0010732142254710197,
0.04899197444319725,
-0.08325168490409851,
-0.02693590335547924,
-0.03843589127063751,
0.0458424836397171,
-0.05159740522503853,
-0.032591670751571655,
-0.06865312159061432,
-0.022879155352711678,
0.08462372422218323,
0.025510134175419807,
0.05528124421834946,
0.04884271323680878,
-0.012371735647320747,
0.05277884006500244,
0.0004153384652454406,
0.0034263732377439737,
-0.03849291428923607,
0.01879332959651947,
-0.11109733581542969,
-0.007642952725291252,
0.017304526641964912,
0.04127196595072746,
0.0667325034737587,
-0.029484594240784645,
-0.025518620386719704,
0.017479009926319122,
0.0032012267038226128,
-0.027782026678323746,
0.11442258954048157,
0.07153144478797913,
-0.046855777502059937,
-0.008361163549125195,
0.01046088244765997,
-0.03892643377184868,
-0.016689613461494446,
-0.06934141367673874,
0.046593084931373596,
0.005755083169788122,
-0.03903889283537865,
0.0686001181602478,
-0.05945521220564842,
0.07251294702291489,
-0.036449264734983444,
0.06723348051309586,
-0.005567480344325304,
-0.0008138903067447245,
0.0550508126616478,
0.0690198764204979,
0.02092568948864937,
-0.035481467843055725,
-0.05061997473239899,
-0.030201861634850502,
0.013546353206038475,
0.028805680572986603,
0.03002702072262764,
-0.05601843446493149,
0.0067511615343391895
] | 0 | 0 | 0 | 0 |
|
SOO-Bench: Benchmarks for Evaluating the Stability of Offline Black-Box Optimization | https://openreview.net/forum?id=bqf0aCF3Dd | [
"Hong Qian",
"Yiyi Zhu",
"Xiang Shu",
"Shuo Liu",
"Yaolin Wen",
"Xin An",
"Huakang Lu",
"Aimin Zhou",
"Ke Tang",
"Yang Yu"
] | Poster | Black-box optimization aims to find the optima through building a model close to the black-box objective function based on function value evaluation. However, in many real-world tasks, such as the design of molecular formulas and mechanical structures, it is perilous, costly, or even infeasible to evaluate the objective function value of an actively sampled solution. In this situation, optimization can only be conducted via utilizing offline historical data, which yields offline black-box optimization. Different from the traditional goal that is to pursue the optimal solution, this paper emphasizes that the goal of offline optimization is to stably surpass the offline dataset during optimization procedure. Although benchmarks called Design-Bench already exist in this emerging field, it can hardly evaluate the stability of offline optimization and mainly provides real-world offline tasks and the corresponding offline datasets. To this end, this paper proposes benchmarks named SOO-Bench (i.e., Stable Offline Optimization Benchmarks) for offline black-box optimization algorithms, so as to systematically evaluate the stability of surpassing the offline dataset under different data distributions. Along with SOO-Bench, we also propose a stability indicator to measure the degree of stability. Specifically, SOO-Bench includes various real-world offline optimization tasks and offline datasets under different data distributions, involving the fields of satellites, materials science, structural mechanics, and automobile manufacturing. Empirically, baseline and state-of-the-art algorithms are tested and analyzed on SOO-Bench. Hopefully, SOO-Bench is expected to serve as a catalyst for the rapid developments of more novel and stable offline optimization methods. The code is available at \url{https://github.com/zhuyiyi-123/SOO-Bench}. | Offline Optimization, Black-Box Optimization, Stability, Benchmarks | null | 13,800 | null | [
-0.11064593493938446,
-0.012248966842889786,
-0.06399036943912506,
0.036939285695552826,
0.010671804659068584,
-0.0540107898414135,
-0.0707552582025528,
0.054436009377241135,
-0.047588735818862915,
-0.04273024946451187,
-0.06590348482131958,
0.006030144169926643,
-0.044026050716638565,
-0.007635999005287886,
-0.052254706621170044,
0.029004380106925964,
0.07312384992837906,
0.055719006806612015,
-0.09837637096643448,
-0.0468391515314579,
-0.0049162376672029495,
0.022914474830031395,
0.014538352377712727,
-0.027528665959835052,
0.013733173720538616,
0.041199687868356705,
-0.006474100053310394,
0.011681360192596912,
-0.0004992556641809642,
-0.0327908992767334,
0.030496418476104736,
-0.004319456871598959,
-0.07270585000514984,
-0.07217739522457123,
0.020221106708049774,
0.028342073783278465,
-0.11820638924837112,
0.008272460661828518,
-0.011381031014025211,
-0.02553595043718815,
-0.08187660574913025,
0.021282227709889412,
-0.01316334679722786,
-0.013460145331919193,
0.04741737246513367,
-0.01530388928949833,
0.029386987909674644,
-0.000009349633728561457,
-0.000951176683884114,
0.001311934320256114,
-0.07888054102659225,
-0.016871782019734383,
-0.06315456330776215,
0.05683080106973648,
0.04623536393046379,
0.07413442432880402,
0.06562204658985138,
-0.04739820584654808,
0.05234494060277939,
-0.013001934625208378,
0.047009922564029694,
-0.05229593813419342,
0.012667975388467312,
-0.008936497382819653,
0.0003846906474791467,
0.027881629765033722,
0.07596398144960403,
-0.03437187895178795,
0.05074393004179001,
0.047619421035051346,
-0.07412001490592957,
0.11757203191518784,
0.0006484886980615556,
-0.008738715201616287,
0.02031063102185726,
0.04086054116487503,
0.0777224600315094,
-0.034080781042575836,
0.03800974786281586,
-0.030893782153725624,
0.03435385972261429,
0.02372351661324501,
-0.0335799902677536,
0.04643303528428078,
0.04941321909427643,
0.027774205431342125,
0.02310766465961933,
0.06186889484524727,
0.10926306992769241,
0.024940554052591324,
0.010994663462042809,
0.10863722115755081,
0.03978431597352028,
-0.021389028057456017,
-0.06877458095550537,
-0.02756999433040619,
0.04014153406023979,
0.038372114300727844,
0.031131494790315628,
0.04784572869539261,
-0.010777886025607586,
0.04997294396162033,
0.04138587787747383,
-0.031279586255550385,
0.0051402319222688675,
-0.0727669969201088,
0.07626267522573471,
0.07529589533805847,
0.0590653158724308,
-0.0031009677331894636,
0.02625274285674095,
0.019732199609279633,
0.008598966524004936,
-0.011147397570312023,
-0.037500880658626556,
-0.0027892363723367453,
0.029101045802235603,
-0.07787971943616867,
-0.01606873795390129,
0.08011983335018158,
-0.022604214027523994,
-0.021959995850920677,
0.017158936709165573,
0.03222960978746414,
0.010955935344099998,
0.03854198381304741,
-0.06582096964120865,
3.586309470109978e-33,
-0.04620435833930969,
-0.056251924484968185,
0.056089892983436584,
-0.056849006563425064,
0.03240647539496422,
-0.10695634037256241,
0.01275464054197073,
-0.014853055588901043,
-0.10298945754766464,
0.030488628894090652,
-0.04451689124107361,
-0.036527715623378754,
-0.10658618062734604,
-0.005903503391891718,
0.11376558989286423,
-0.04392436891794205,
-0.00303093739785254,
0.015885425731539726,
0.02807985059916973,
0.019504787400364876,
0.05694526433944702,
-0.08961772918701172,
0.006340000778436661,
-0.026983246207237244,
0.06740479171276093,
0.020112987607717514,
0.07000410556793213,
0.029403304681181908,
-0.10384664684534073,
0.0006682109087705612,
0.004422557074576616,
-0.06546536833047867,
-0.028583908453583717,
0.017665620893239975,
0.07948426902294159,
-0.04226142540574074,
-0.07666991651058197,
-0.05178157612681389,
-0.007806127890944481,
0.0527765229344368,
0.0010420953622087836,
0.07131581008434296,
0.046195611357688904,
-0.030499905347824097,
-0.0309051051735878,
-0.1107499971985817,
0.01571972481906414,
0.025542501360177994,
0.038946934044361115,
-0.0783652663230896,
-0.02161741815507412,
0.020033180713653564,
0.027576018124818802,
0.013339242897927761,
-0.0389813557267189,
-0.011613093316555023,
-0.012094604782760143,
0.02300094999372959,
0.024985386058688164,
0.13319189846515656,
-0.05386081710457802,
-0.050377633422613144,
-0.00017006396956276149,
-0.032799430191516876,
0.010879253037273884,
0.05018869787454605,
-0.05519604682922363,
0.0017649463843554258,
-0.018558640033006668,
0.03428282588720322,
0.012220785021781921,
-0.002231156686320901,
0.0541347898542881,
-0.04994351044297218,
0.0714145302772522,
0.048638373613357544,
0.004105820786207914,
0.0112773347645998,
0.04437803849577904,
-0.005402762442827225,
0.05649310350418091,
0.12146033346652985,
-0.018835585564374924,
-0.027062127366662025,
-0.02549697831273079,
-0.13256508111953735,
-0.014535708352923393,
0.002467229263857007,
-0.10200735926628113,
-0.01861824467778206,
-0.11691828072071075,
-0.0651683658361435,
-0.100124292075634,
-0.08586081862449646,
-0.0022024493664503098,
-3.115095661851571e-33,
-0.04134973883628845,
-0.051807254552841187,
0.0463036485016346,
0.09000289440155029,
-0.0001427095412509516,
0.010944419540464878,
-0.011392205953598022,
-0.05965209752321243,
0.03295845538377762,
0.04611218720674515,
0.006933659315109253,
-0.0078561557456851,
0.001417343388311565,
-0.01443262118846178,
0.038440242409706116,
0.052750591188669205,
-0.100033700466156,
-0.04573175683617592,
0.03928312286734581,
0.06566981971263885,
-0.006493318825960159,
0.05115605890750885,
-0.062121398746967316,
0.03179061785340309,
-0.10041669011116028,
0.01012521144002676,
-0.07956476509571075,
0.020656384527683258,
-0.01673395372927189,
0.026117734611034393,
0.011858982965350151,
0.03772813454270363,
-0.1066647469997406,
0.029460176825523376,
-0.07046135514974594,
0.07138600200414658,
0.019639810547232628,
0.047906506806612015,
-0.00012680537474807352,
0.051135607063770294,
0.058165293186903,
0.04240883141756058,
-0.0960756242275238,
-0.04038437455892563,
0.08447828888893127,
0.08058473467826843,
-0.07725092768669128,
-0.011778869666159153,
0.019714441150426865,
-0.02403634414076805,
-0.023200811818242073,
0.03289739787578583,
0.019558265805244446,
0.023383408784866333,
-0.02178061380982399,
-0.06137101724743843,
-0.07587256282567978,
-0.015440844930708408,
0.05245653912425041,
0.039632443338632584,
0.02179158292710781,
-0.07553218305110931,
-0.03854753077030182,
0.009625297971069813,
-0.03498128801584244,
-0.024942142888903618,
0.03154025599360466,
-0.007817547768354416,
-0.045339323580265045,
-0.03544149175286293,
-0.014486369676887989,
-0.03470330685377121,
0.07672639191150665,
0.052646003663539886,
-0.05762406066060066,
-0.021864237263798714,
0.04128438979387283,
-0.027633029967546463,
0.016884583979845047,
0.09074278175830841,
0.03007548302412033,
0.06216675788164139,
0.01443594228476286,
0.06622551381587982,
0.0319732204079628,
0.01550764124840498,
0.01476488821208477,
0.11131276190280914,
0.0250936858355999,
-0.04551362991333008,
-0.019642231985926628,
0.006124615203589201,
-0.013377154245972633,
0.0427006296813488,
0.008136735297739506,
-4.499056416307212e-8,
0.008339796215295792,
-0.009006456471979618,
0.11232881247997284,
-0.008408510126173496,
-0.006612819619476795,
0.0381893552839756,
0.029215235263109207,
0.010566337034106255,
0.05590144917368889,
-0.0888184979557991,
0.0919448733329773,
-0.05105705186724663,
-0.025159139186143875,
-0.024921970441937447,
-0.04974541440606117,
0.06184881925582886,
-0.02858174405992031,
-0.054101429879665375,
0.03214126080274582,
-0.031336650252342224,
-0.01926310360431671,
0.030884835869073868,
0.033035993576049805,
-0.008384007029235363,
0.07047376781702042,
-0.08980889618396759,
0.018841369077563286,
0.0004883493529632688,
0.082774318754673,
-0.027614912018179893,
-0.045530304312705994,
-0.003568915417417884,
0.12509500980377197,
0.026898346841335297,
0.09300440549850464,
-0.003150997683405876,
-0.006294180639088154,
0.05194973573088646,
-0.02023239992558956,
0.0725889727473259,
0.03247176855802536,
0.09931164234876633,
-0.11021251976490021,
-0.04084300994873047,
0.13232602179050446,
0.0044429418630898,
-0.07128776609897614,
0.03596988692879677,
0.012014791369438171,
0.01580238714814186,
0.0522085539996624,
-0.047047849744558334,
-0.06857984513044357,
0.027355214580893517,
0.04095291718840599,
-0.06944394111633301,
-0.07248931378126144,
-0.04989704117178917,
0.03662233427166939,
-0.004225290846079588,
0.03253303840756416,
-0.04639691859483719,
-0.11416470259428024,
-0.04075092449784279
] | 0 | 0 | 0 | 0 |
|
Understanding and Enhancing Safety Mechanisms of LLMs via Safety-Specific Neuron | https://openreview.net/forum?id=yR47RmND1m | [
"Yiran Zhao",
"Wenxuan Zhang",
"Yuxi Xie",
"Anirudh Goyal",
"Kenji Kawaguchi",
"Michael Shieh"
] | Poster | Safety alignment for large language models (LLMs) has become a critical issue due to their rapid progress. However, our understanding of effective safety mechanisms in LLMs remains limited, leading to safety alignment training that mainly focuses on improving optimization, data-level enhancement, or adding extra structures to intentionally block harmful outputs. To address this gap, we develop a neuron detection method to identify safety neurons—those consistently crucial for handling and defending against harmful queries. Our findings reveal that these safety neurons constitute less than $1\%$ of all parameters, are language-specific and are predominantly located in self-attention layers. Moreover, safety is collectively managed by these neurons in the first several layers. Based on these observations, we introduce a $\underline{S}$afety $\underline{N}$euron $\underline{Tun}$ing method, named $\texttt{SN-Tune}$, that exclusively tune safety neurons without compromising models' general capabilities. $\texttt{SN-Tune}$ significantly enhances the safety of instruction-tuned models, notably reducing the harmful scores of Llama3-8B-Instruction from $65.5$ to $2.0$, Mistral-7B-Instruct-v0.2 from $70.8$ to $4.5$, and Vicuna-13B-1.5 from $93.5$ to $3.0$. Moreover, $\texttt{SN-Tune}$ can be applied to base models on efficiently establishing LLMs' safety mechanism. In addition, we propose $\underline{R}$obust $\underline{S}$afety $\underline{N}$euron $\underline{Tun}$ing method ($\texttt{RSN-Tune}$), which preserves the integrity of LLMs' safety mechanisms during downstream task fine-tuning by separating the safety neurons from models' foundation neurons. | Large Language Models, Alignment, Safety, Interpretability, Neuron Detection | null | 13,799 | null | [
-0.00973929651081562,
-0.08062111586332321,
-0.02745751291513443,
-0.06015424057841301,
0.030682949349284172,
0.08949345350265503,
0.04470439627766609,
0.03192010894417763,
0.0669102594256401,
0.020654963329434395,
-0.027463793754577637,
-0.02041708119213581,
0.05237288400530815,
-0.040374040603637695,
-0.08387687802314758,
-0.044790104031562805,
0.06408573687076569,
0.06586156040430069,
-0.12382909655570984,
-0.014466694556176662,
0.07459801435470581,
0.04579133540391922,
0.07942190766334534,
0.03733083978295326,
-0.06096724048256874,
-0.0068430607207119465,
-0.06273284554481506,
-0.03114304691553116,
0.015562077052891254,
-0.07935623824596405,
0.051834430545568466,
-0.08167169243097305,
0.0021273663733154535,
0.08260854333639145,
-0.07656803727149963,
-0.047493934631347656,
-0.07395721971988678,
-0.02642439305782318,
-0.015015779994428158,
-0.06956654787063599,
-0.0543762743473053,
0.010793635621666908,
0.02316443622112274,
-0.004569888114929199,
0.10168256610631943,
-0.007099148817360401,
0.013721460476517677,
-0.04618264362215996,
-0.039099615067243576,
-0.07411115616559982,
-0.04268168658018112,
-0.03883178532123566,
0.020812619477510452,
0.031274959444999695,
0.058793243020772934,
0.019831271842122078,
0.009084210731089115,
0.07036211341619492,
-0.06336970627307892,
0.04981289431452751,
0.0005634089466184378,
-0.034301891922950745,
-0.03620252385735512,
-0.05106009170413017,
0.020720047876238823,
0.028514960780739784,
-0.03334321826696396,
0.03109564445912838,
-0.02448449283838272,
0.08189772814512253,
-0.0033962682355195284,
0.021621547639369965,
0.009667770937085152,
0.012730211950838566,
0.06995262950658798,
0.07994629442691803,
0.07636778056621552,
-0.0003947522782254964,
0.07986103743314743,
-0.0787273496389389,
-0.0033413050696253777,
-0.004821414593607187,
0.09068150818347931,
0.06437384337186813,
0.10104382038116455,
0.044350165873765945,
0.0011909141903743148,
0.0006169523694552481,
0.10214265435934067,
-0.01097716111689806,
-0.07900411635637283,
-0.10183742642402649,
0.004799434449523687,
0.023721059784293175,
0.07611995935440063,
0.023835105821490288,
-0.01450173370540142,
-0.027533410117030144,
-0.02398425154387951,
0.004798691254109144,
0.09128023684024811,
0.07609392702579498,
0.010867342352867126,
0.0014369371347129345,
0.047460902482271194,
0.002276401733979583,
0.016587747260928154,
0.02708517201244831,
0.020089415833353996,
-0.046415235847234726,
-0.016516968607902527,
0.0653170496225357,
0.006174310576170683,
0.027845239266753197,
0.0074198259972035885,
-0.04469969496130943,
-0.029700176790356636,
-0.02728837914764881,
0.06325669586658478,
0.03726792335510254,
-0.005969064775854349,
0.038727760314941406,
-0.017031176015734673,
0.02942558564245701,
0.038167715072631836,
-0.04101530462503433,
-0.07106117159128189,
3.209906628085081e-33,
0.039135292172431946,
0.08255666494369507,
-0.09179063886404037,
-0.02688666246831417,
-0.012294323183596134,
-0.01588790863752365,
-0.014808827079832554,
0.014801880344748497,
0.007056557573378086,
0.10045073926448822,
-0.08561331033706665,
-0.029780447483062744,
-0.007788084447383881,
0.01383884809911251,
0.10393722355365753,
0.03104807250201702,
-0.038664866238832474,
0.006109343376010656,
-0.10581736266613007,
-0.09728984534740448,
0.07618816941976547,
-0.020910361781716347,
0.06975015252828598,
-0.03991645947098732,
0.010007830336689949,
0.0002958266413770616,
0.022794131189584732,
-0.02933610789477825,
-0.06951473653316498,
0.032949306070804596,
-0.13199807703495026,
-0.02379787713289261,
-0.01325126364827156,
-0.006828795652836561,
0.05912483483552933,
-0.025555530562996864,
-0.09894709289073944,
-0.006638721562922001,
0.04257539287209511,
0.012002309784293175,
0.0032014071475714445,
0.07900066673755646,
0.03934910520911217,
0.026880178600549698,
0.014399182982742786,
-0.0013836509315297008,
-0.05399545282125473,
0.01358993910253048,
-0.020571626722812653,
-0.10013452172279358,
-0.018277760595083237,
0.019822223111987114,
0.05054479092359543,
-0.04906269162893295,
-0.052638232707977295,
0.05702997371554375,
0.02851262129843235,
0.059582654386758804,
-0.031949032098054886,
0.05761780962347984,
-0.0597677119076252,
0.0036798466462641954,
0.005616288632154465,
0.04511219263076782,
0.10049741715192795,
-0.006500418297946453,
-0.014515111222863197,
0.04427823796868324,
0.07859066873788834,
-0.052699267864227295,
-0.06327022612094879,
-0.020898083224892616,
0.007287683896720409,
-0.05693789944052696,
-0.054278500378131866,
-0.037624821066856384,
0.08642130345106125,
-0.052326809614896774,
0.030069729313254356,
-0.02118806727230549,
0.08066515624523163,
0.007310966961085796,
-0.016993194818496704,
-0.002590073039755225,
-0.06031285226345062,
-0.09156569838523865,
0.052555281668901443,
-0.026172546669840813,
-0.0042993756942451,
0.08570398390293121,
0.01676337793469429,
-0.07158974558115005,
0.06117510795593262,
-0.024443548172712326,
-0.09867487847805023,
-2.290670964097014e-33,
-0.05689486861228943,
-0.06182518228888512,
-0.06727912276983261,
0.05347533896565437,
-0.10777387768030167,
-0.05422661080956459,
0.022278526797890663,
0.04021883010864258,
-0.09910878539085388,
-0.013635881245136261,
-0.06149962544441223,
-0.05635815113782883,
0.047591373324394226,
-0.004720914177596569,
0.016418572515249252,
-0.056361667811870575,
-0.06164814531803131,
0.11397309601306915,
0.06324277073144913,
-0.02036244235932827,
0.00813356228172779,
0.09058007597923279,
-0.1279059499502182,
0.06837564706802368,
-0.0667838454246521,
0.04289482906460762,
-0.048155251890420914,
0.05586561933159828,
0.019258473068475723,
0.0383576825261116,
-0.035114873200654984,
0.039385054260492325,
0.008128112182021141,
0.012228380888700485,
-0.022728538140654564,
0.04316338896751404,
0.05311822518706322,
-0.0344071201980114,
0.005432441830635071,
-0.03190729022026062,
0.0567014254629612,
0.05202633515000343,
-0.027121596038341522,
0.01073121465742588,
0.016085920855402946,
-0.023549657315015793,
-0.06610029190778732,
0.003359298687428236,
-0.04830595478415489,
0.016554027795791626,
0.0012992601841688156,
-0.06803704053163528,
-0.0016025126678869128,
-0.014989549294114113,
-0.02598307840526104,
-0.028658343479037285,
0.023159600794315338,
-0.05483183637261391,
0.04270521178841591,
0.0004408010863699019,
-0.05810787156224251,
-0.017388200387358665,
0.027209173887968063,
-0.06472453474998474,
0.011852368712425232,
0.06377965211868286,
-0.06432195752859116,
-0.0009597094031050801,
0.0982886254787445,
-0.07845765352249146,
0.09437795728445053,
0.004732104949653149,
0.013260343112051487,
-0.010775089263916016,
-0.09891185164451599,
-0.017619460821151733,
-0.035925719887018204,
-0.03380430489778519,
-0.018352920189499855,
-0.023955149576067924,
0.013095984235405922,
0.008759060874581337,
-0.0043477462604641914,
0.08500369638204575,
0.05096747726202011,
0.03719399869441986,
0.08269065618515015,
0.07949727773666382,
-0.04126882553100586,
-0.052462462335824966,
0.01867586374282837,
0.020877264440059662,
-0.04513595625758171,
0.025437450036406517,
-0.07007090747356415,
-4.7169425698712075e-8,
-0.039421938359737396,
-0.013445742428302765,
-0.009559131227433681,
0.13334958255290985,
0.020319877192378044,
-0.05601535737514496,
-0.003828437300398946,
0.026749899610877037,
-0.039318062365055084,
0.0017482800176367164,
0.07155399769544601,
0.03059689886868,
-0.059042006731033325,
-0.018305249512195587,
0.026694167405366898,
0.051144085824489594,
0.01842433214187622,
0.07852909713983536,
-0.07534705847501755,
-0.02661963552236557,
-0.006814966443926096,
0.08533865958452225,
0.006498759146779776,
-0.026353081688284874,
0.05518399924039841,
-0.023276809602975845,
-0.04602757841348648,
0.10953908413648605,
0.0675278753042221,
0.029312297701835632,
0.038494620472192764,
-0.022355351597070694,
-0.006815462373197079,
0.01574334315955639,
-0.011819942854344845,
0.12372662127017975,
0.00003314310743007809,
0.009814975783228874,
0.036890868097543716,
0.0014348552795127034,
0.02507716231048107,
0.03538099676370621,
0.035092905163764954,
0.08633455634117126,
-0.01780523732304573,
-0.054064907133579254,
-0.02156420424580574,
-0.0913209468126297,
0.012114876881241798,
-0.07163141667842865,
0.027361426502466202,
-0.004006646107882261,
-0.005603610072284937,
0.06695625185966492,
-0.013467919081449509,
0.006538732443004847,
-0.01961749978363514,
-0.06470062583684921,
-0.02205629087984562,
0.11908906698226929,
0.02822985127568245,
-0.020680226385593414,
0.013164094649255276,
-0.0337742455303669
] | 0 | 0 | 0 | 0 |
|
Long Context Compression with Activation Beacon | https://openreview.net/forum?id=1eQT9OzfNQ | [
"Peitian Zhang",
"Zheng Liu",
"Shitao Xiao",
"Ninglu Shao",
"Qiwei Ye",
"Zhicheng Dou"
] | Poster | Long context compression is a critical research problem due to its significance in reducing the high computational and memory costs associated with LLMs. In this paper, we propose Activation Beacon, a plug-in module for transformer-based LLMs that targets effective, efficient, and flexible compression of long contexts. To achieve this, our method introduces the following technical designs.
1) We directly compress the activations (i.e. keys and values at every layer), rather than leveraging soft prompts to relay information (which constitute a major bottleneck to encapsulate the complex information within long contexts).
2) We tailor the compression workflow, where each fine-grained input unit is progressively compressed, enabling high-quality compression and efficient computation during both training and inference.
3) We train the model through compression-based auto-regression, making full use of plain texts and instructional data to optimize the model's compression performance.
4) During training, we randomly sample a compression ratio at each step, teaching the model to support a wide range of compression configurations.
Extensive evaluations are conducted on various long-context tasks whose lengths (e.g., 128K) may far exceed the maximum training length (20K), such as document understanding, few-shot learning, and Needle-in-a-Haystack. Whilst existing methods struggle to handle these challenging tasks, Activation Beacon maintains a comparable performance to the uncompressed baseline across various scenarios,
achieving a 2x acceleration in inference time and an 8x reduction of memory costs for KV cache. | Context Compression, Long Context LLMs, LLM Memory | null | 13,798 | 2401.03462 | [
-0.041681572794914246,
0.02545608952641487,
0.028640951961278915,
0.03275708109140396,
0.06861938536167145,
-0.03278419375419617,
-0.00456579402089119,
-0.00826498493552208,
0.005472046788781881,
-0.048598844558000565,
-0.07137898355722427,
0.0033598076552152634,
0.03591172769665718,
0.002902831882238388,
0.03586757183074951,
0.023245492950081825,
0.06879603117704391,
0.11095750331878662,
-0.10203836113214493,
-0.07497157901525497,
0.07210348546504974,
0.00207554385997355,
0.03319267928600311,
-0.005246197339147329,
0.08977645635604858,
-0.03484375774860382,
-0.02233183942735195,
-0.07935022562742233,
0.07855599373579025,
0.0027395610231906176,
0.01619083620607853,
-0.0082151023671031,
0.06196582689881325,
0.06292571127414703,
-0.06605098396539688,
0.0484730526804924,
-0.04165143147110939,
0.010951180942356586,
-0.047855015844106674,
-0.03646664693951607,
-0.015574601478874683,
-0.04339103028178215,
0.011107315309345722,
0.09558485448360443,
0.008456691168248653,
0.019748955965042114,
0.07221433520317078,
-0.0112023139372468,
-0.016675394028425217,
0.025165867060422897,
-0.021247457712888718,
0.00031082413624972105,
-0.025888513773679733,
0.051240503787994385,
-0.07566247880458832,
0.01704367808997631,
0.036907684057950974,
0.06789087504148483,
-0.012190920300781727,
0.061340853571891785,
-0.05777207016944885,
-0.07075437158346176,
-0.034260060638189316,
0.0036137355491518974,
-0.0535111278295517,
-0.0020791110582649708,
0.09406574815511703,
0.07596928626298904,
0.020589562132954597,
0.003963784780353308,
-0.03945987671613693,
-0.0021043107844889164,
-0.05980202928185463,
0.08465218544006348,
-0.00653781509026885,
0.01152182649821043,
-0.0045646424405276775,
-0.04296866059303284,
0.01936979778110981,
-0.026530960574746132,
0.05398276448249817,
-0.0455460324883461,
0.06559107452630997,
-0.05101846531033516,
0.05001937597990036,
-0.02241809479892254,
0.0005116498796269298,
0.029562171548604965,
-0.033503416925668716,
0.0018957494758069515,
-0.03394582122564316,
-0.09159228950738907,
0.05365164950489998,
2.40363078773953e-7,
0.050235409289598465,
0.006770119071006775,
-0.010000910609960556,
-0.06361618638038635,
-0.07143213599920273,
0.07900463789701462,
0.010784880258142948,
0.03566240146756172,
-0.0542580671608448,
-0.02639533020555973,
-0.05664868652820587,
-0.07865725457668304,
0.027063751593232155,
0.08304807543754578,
-0.020825503394007683,
-0.0006745418650098145,
0.052138037979602814,
0.07748395204544067,
0.0040564145892858505,
-0.07208415865898132,
0.05335715040564537,
-0.024397239089012146,
-0.0022897913586348295,
-0.022072410210967064,
0.020712340250611305,
0.14579728245735168,
-0.02741047367453575,
-0.028173575177788734,
-0.12461847066879272,
0.014164549298584461,
0.007099408656358719,
-0.029937561601400375,
-0.03703979775309563,
9.467953127140881e-34,
0.05874999985098839,
0.10800660401582718,
-0.0364094078540802,
0.0743010863661766,
0.02231673337519169,
0.04504699632525444,
0.039345066994428635,
0.06412629038095474,
0.0035089021548628807,
0.07334819436073303,
-0.038522109389305115,
0.04398270696401596,
-0.03316975384950638,
0.09765132516622543,
0.06857097893953323,
-0.016636580228805542,
-0.02899831347167492,
0.09770280122756958,
-0.02118426188826561,
-0.010622117668390274,
0.023383626714348793,
-0.05613885447382927,
0.012879476882517338,
-0.07446026802062988,
0.04592791199684143,
-0.01431439071893692,
0.05003076046705246,
-0.04356249421834946,
-0.007073039188981056,
-0.010986581444740295,
-0.10163944214582443,
-0.0023227909114211798,
0.0009613207657821476,
-0.015061486512422562,
0.05972268059849739,
-0.03458604961633682,
0.006761470809578896,
-0.026119353249669075,
-0.025354739278554916,
-0.02699314057826996,
0.013328561559319496,
0.03428105637431145,
0.03548621013760567,
-0.013860008679330349,
-0.09422706067562103,
-0.10768201947212219,
-0.015289940871298313,
0.026307247579097748,
-0.057416267693042755,
0.021873045712709427,
0.11409095674753189,
-0.0002112514484906569,
-0.07669361680746078,
-0.0073628174141049385,
0.03428143262863159,
-0.027923593297600746,
0.040621183812618256,
-0.008281399495899677,
0.07190918922424316,
0.06471437960863113,
-0.06015556678175926,
-0.017693497240543365,
0.002379225566983223,
0.09190906584262848,
0.0632452592253685,
0.00669392803683877,
-0.049967844039201736,
0.05241403356194496,
0.0033488390035927296,
-0.044181160628795624,
-0.04705922305583954,
-0.016528137028217316,
-0.008666198700666428,
-0.03723979368805885,
-0.002956019714474678,
-0.0023570405319333076,
0.014472901821136475,
-0.06300247460603714,
-0.1120578870177269,
0.0155570013448596,
0.03595441207289696,
0.019607536494731903,
0.015008916147053242,
-0.012271260842680931,
0.02907385118305683,
0.001275626476854086,
-0.041050031781196594,
-0.10667378455400467,
-0.05950925499200821,
-0.033266883343458176,
-0.058177221566438675,
-0.020500941202044487,
-0.04481825977563858,
-0.05037080869078636,
-0.04351688548922539,
-3.4043170469927716e-34,
0.009496690705418587,
0.03668658062815666,
-0.0850709080696106,
0.09329631179571152,
0.03503831848502159,
-0.050243597477674484,
0.02866687811911106,
0.03504263609647751,
-0.08776403218507767,
-0.04073071479797363,
-0.04796471819281578,
-0.05608208850026131,
0.03691689670085907,
-0.010600319132208824,
0.00801520049571991,
-0.018370386213064194,
-0.060666147619485855,
-0.0086001455783844,
-0.019526267424225807,
-0.018388979136943817,
-0.01089899055659771,
0.07108159363269806,
-0.07399854809045792,
-0.011248487047851086,
-0.032331738620996475,
-0.022094862535595894,
-0.11344953626394272,
0.09623660892248154,
0.04512634873390198,
0.041448693722486496,
-0.049947164952754974,
0.001094477018341422,
-0.028178129345178604,
-0.052093636244535446,
-0.04237492009997368,
0.02765805833041668,
0.08255758881568909,
0.1036519929766655,
-0.005232640542089939,
0.04855826497077942,
0.1407777965068817,
-0.027602341026067734,
-0.03044501692056656,
0.028053421527147293,
-0.0032462086528539658,
-0.06585026532411575,
-0.07722779363393784,
-0.04363317787647247,
0.0084396256133914,
0.045274414122104645,
-0.007769560907036066,
0.003977731801569462,
-0.08272925764322281,
0.004626655951142311,
0.019264869391918182,
-0.11346495896577835,
-0.10807649046182632,
-0.022959310561418533,
0.010569927282631397,
-0.05100132152438164,
-0.009640968404710293,
-0.044865839183330536,
0.05085469409823418,
-0.030399665236473083,
0.0499102883040905,
-0.07383687794208527,
0.013380724936723709,
-0.015896394848823547,
-0.01969919726252556,
0.05888133868575096,
0.006989250425249338,
-0.0064420863054692745,
-0.0032392535358667374,
0.050848059356212616,
-0.017937347292900085,
0.014907581731677055,
0.02114318497478962,
-0.04376736283302307,
-0.02349015511572361,
0.034366827458143234,
0.011828158050775528,
0.0382271409034729,
-0.03322437405586243,
0.05471837893128395,
0.044635429978370667,
0.1044248715043068,
0.05488874390721321,
0.12295792996883392,
0.035292088985443115,
0.04594021663069725,
-0.06505259871482849,
0.09305192530155182,
0.005260063335299492,
0.01423433143645525,
-0.027572499588131905,
-4.759663241316048e-8,
-0.04304535686969757,
0.1058216542005539,
-0.004736732225865126,
0.02156667970120907,
0.013661039993166924,
0.005232265684753656,
0.03556373342871666,
0.11834971606731415,
0.035506319254636765,
-0.025134587660431862,
0.07861936837434769,
-0.07383455336093903,
-0.0520908460021019,
-0.028527209535241127,
0.026248449459671974,
0.0847800076007843,
0.007509750779718161,
0.006421801168471575,
-0.011724748648703098,
0.007084731012582779,
0.031107168644666672,
0.08431258797645569,
-0.02344539947807789,
0.04169386252760887,
0.055574364960193634,
-0.0743362084031105,
0.024559518322348595,
0.131653293967247,
0.0690484270453453,
-0.012828998267650604,
-0.056535009294748306,
0.007530918810516596,
-0.03863603621721268,
-0.08234719932079315,
-0.11957783997058868,
0.06319521367549896,
0.08114966750144958,
-0.01380646787583828,
0.03432343900203705,
-0.05941196531057358,
-0.006671702489256859,
-0.01922205276787281,
-0.08214367181062698,
0.01759730651974678,
0.01632346771657467,
-0.0007252516224980354,
0.04941709339618683,
-0.08857130259275436,
-0.05549348145723343,
0.012911983765661716,
0.012606690637767315,
-0.012983853928744793,
-0.020994633436203003,
0.03123919852077961,
0.09093715250492096,
0.015484780073165894,
0.002216736786067486,
-0.007669374812394381,
0.06345265358686447,
0.07558313757181168,
-0.018769579008221626,
0.023613251745700836,
-0.10192186385393143,
-0.03218352422118187
] | https://github.com/flagopen/flagembedding | 9,473 | 0 | 0 | 0 |
LASeR: Towards Diversified and Generalizable Robot Design with Large Language Models | https://openreview.net/forum?id=7mlvOHL6qJ | [
"Junru Song",
"Yang Yang",
"Huan Xiao",
"Wei Peng",
"Wen Yao",
"Feifei Wang"
] | Poster | Recent advances in Large Language Models (LLMs) have stimulated a significant paradigm shift in evolutionary optimization, where hand-crafted search heuristics are gradually replaced with LLMs serving as intelligent search operators. However, these studies still bear some notable limitations, including a challenge to balance exploitation with exploration, often leading to inferior solution diversity, as well as poor generalizability of problem solving across different task settings. These unsolved issues render the prowess of LLMs in robot design automation largely untapped. In this work, we present LASeR -- Large Language Model-Aided Evolutionary Search for Robot Design Automation. Leveraging a novel reflection mechanism termed DiRect, we elicit more knowledgeable exploratory behaviors from LLMs based on past search trajectories, reshaping the exploration-exploitation tradeoff with dual improvements in optimization efficiency and solution diversity. Additionally, with evolution fully grounded in task-related background information, we unprecedentedly uncover the inter-task reasoning capabilities of LLMs, facilitating generalizable design processes that effectively inspire zero-shot robot proposals for new applications. Our simulated experiments on voxel-based soft robots showcase distinct advantages of LASeR over competitive baselines. Code at https://github.com/WoodySJR/LASeR. | Robot Design Automation, Large Language Model, Voxel-Based Soft Robot | This work improves the diversity and inter-task generalizability of robot design processes with the aid of Large Language Models. | 13,792 | null | [
-0.03130219504237175,
-0.07642947882413864,
0.03780977055430412,
0.03589870035648346,
0.0425381101667881,
-0.045480865985155106,
-0.012123010121285915,
-0.0338507741689682,
-0.023795081302523613,
-0.014994404278695583,
-0.06763864308595657,
-0.09870653599500656,
0.03355984762310982,
0.07283545285463333,
0.04880448430776596,
0.015127074904739857,
0.10454210638999939,
0.006638354621827602,
-0.04994482174515724,
0.024780534207820892,
0.06914377212524414,
0.006752813700586557,
0.06422001868486404,
-0.009436697699129581,
-0.055170997977256775,
0.02629181556403637,
0.04352284595370293,
-0.07916039973497391,
0.050468623638153076,
0.008731751702725887,
0.06202574819326401,
-0.01162366010248661,
-0.028091954067349434,
0.009986990131437778,
0.044703930616378784,
0.05923397094011307,
-0.1126176044344902,
0.08021766692399979,
0.006235359236598015,
-0.09260651469230652,
-0.040257781744003296,
-0.010734358802437782,
0.04077104106545448,
0.012148856185376644,
0.02342371642589569,
0.05991389602422714,
-0.08469446003437042,
-0.07513921707868576,
0.005269670393317938,
-0.013730300590395927,
-0.09195855259895325,
-0.10727201402187347,
-0.007332665380090475,
-0.04913279786705971,
0.024503204971551895,
0.018460584804415703,
0.03919430077075958,
0.013641254976391792,
0.09180980920791626,
-0.09193260967731476,
0.002306267386302352,
-0.06817523390054703,
0.05534469708800316,
-0.04378903657197952,
0.0020410979632288218,
-0.017746368423104286,
0.01022463571280241,
0.009366709738969803,
-0.009823106229305267,
0.0037596796173602343,
0.03835081681609154,
0.0314425528049469,
0.0368831604719162,
0.04236304759979248,
0.051358554512262344,
0.03177957981824875,
-0.013310928829014301,
-0.024843603372573853,
-0.03159760311245918,
-0.016620568931102753,
0.017868543043732643,
-0.028587039560079575,
-0.053723130375146866,
0.11097701638936996,
0.04166104272007942,
-0.0061719766817986965,
-0.0012742351973429322,
-0.0010135553311556578,
0.07655617594718933,
-0.0006946313660591841,
-0.03110385127365589,
-0.03580458089709282,
-0.0015850968193262815,
-0.04279940575361252,
-0.005445191171020269,
-0.0012141059851273894,
0.024566754698753357,
-0.07989644259214401,
-0.025910263881087303,
0.010684967041015625,
-0.0016714523080736399,
0.005642299074679613,
0.05094069615006447,
-0.0374065563082695,
-0.053537819534540176,
-0.026681357994675636,
0.0020592433866113424,
-0.004140751902014017,
0.0033073609229177237,
-0.05659722536802292,
0.026447070762515068,
-0.04938674345612526,
0.08828475326299667,
0.0658876970410347,
0.0008133324445225298,
-0.019313836470246315,
-0.0030070857610553503,
-0.011693201027810574,
0.013317020609974861,
0.08305581659078598,
0.037600334733724594,
0.0026718147564679384,
-0.04157046228647232,
0.03890054672956467,
0.02849384769797325,
0.03585865721106529,
-0.0672876164317131,
3.910061093642079e-34,
0.068662628531456,
0.054630596190690994,
0.004921106155961752,
0.08398756384849548,
0.08585900813341141,
-0.0861799493432045,
0.09524514526128769,
0.1264950931072235,
-0.10073261708021164,
0.0037673208862543106,
-0.06546866148710251,
0.04622137174010277,
-0.10595957189798355,
0.06779816746711731,
0.02880614437162876,
0.012789096683263779,
-0.026280701160430908,
0.07290725409984589,
-0.10710178315639496,
-0.006399218458682299,
0.025683743879199028,
0.012038596905767918,
0.0469946414232254,
-0.005338688846677542,
0.020279955118894577,
-0.0030068266205489635,
0.055949702858924866,
-0.09442518651485443,
-0.06099601462483406,
-0.00880285445600748,
-0.17762356996536255,
0.05497187376022339,
0.0019902477506548166,
0.06550908088684082,
-0.007369547616690397,
0.033900536596775055,
-0.033184077590703964,
-0.04078889265656471,
-0.03329085186123848,
0.028159135952591896,
0.008427596651017666,
0.047475844621658325,
-0.0018039046553894877,
-0.04727695509791374,
-0.030294721946120262,
-0.05444997176527977,
-0.013297396712005138,
0.03856286779046059,
-0.02492363564670086,
-0.040849216282367706,
0.03340611234307289,
0.03910253196954727,
-0.0021212196443229914,
-0.002842274960130453,
0.04007555916905403,
-0.0637897178530693,
-0.014352940022945404,
0.04837740585207939,
-0.017063630744814873,
-0.0311528779566288,
0.06810850650072098,
-0.0133847426623106,
0.027075981721282005,
0.06010575219988823,
0.06239080801606178,
-0.015368371270596981,
0.035756610333919525,
0.08979687839746475,
0.141626238822937,
0.045092031359672546,
-0.015826083719730377,
-0.004406890366226435,
0.017589647322893143,
-0.06780858337879181,
0.060990843921899796,
-0.014989619143307209,
0.06147979572415352,
-0.04596226289868355,
0.025362882763147354,
0.007785451132804155,
0.004489418119192123,
0.06440529972314835,
-0.06392929702997208,
0.0052079493179917336,
0.004552381578832865,
-0.06311079114675522,
0.04765598475933075,
-0.043760452419519424,
-0.09138759225606918,
-0.043337006121873856,
-0.08024883270263672,
-0.16132999956607819,
-0.10445823520421982,
0.019928233698010445,
0.007150420919060707,
-2.6383704381267385e-33,
-0.00005807132038171403,
-0.08849916607141495,
-0.021958816796541214,
-0.008824169635772705,
-0.008741972036659718,
-0.05446920916438103,
-0.03618890792131424,
-0.07863679528236389,
-0.026871968060731888,
-0.010980862192809582,
-0.05522184818983078,
-0.0009625487145967782,
0.07363806664943695,
0.02571711875498295,
-0.014974411576986313,
0.00896783173084259,
-0.020111406221985817,
-0.004769514314830303,
0.004869409371167421,
0.037605468183755875,
0.019138120114803314,
0.16497015953063965,
-0.09000900387763977,
0.012065553106367588,
-0.04363740608096123,
0.003585566533729434,
-0.023567749187350273,
0.029058663174510002,
0.01718880608677864,
0.069324791431427,
-0.0430317260324955,
-0.009909025393426418,
-0.01837833784520626,
0.04064548760652542,
-0.008662872947752476,
0.06950586289167404,
-0.007957273162901402,
0.01830359548330307,
0.04224739223718643,
0.034176405519247055,
0.012254335917532444,
-0.08678935468196869,
-0.00780533067882061,
0.029736166819930077,
0.02521740458905697,
-0.04564591497182846,
-0.03687232360243797,
-0.058343786746263504,
-0.012995042838156223,
-0.009629704989492893,
0.016222523525357246,
0.016988806426525116,
-0.026799477636814117,
-0.05285944417119026,
-0.0752691850066185,
-0.08791421353816986,
-0.02823013998568058,
0.012834887020289898,
0.020910702645778656,
0.03765678033232689,
-0.061028048396110535,
0.025959210470318794,
0.023897729814052582,
0.005093955434858799,
-0.08852892369031906,
0.04400280490517616,
-0.06434350460767746,
-0.036387886852025986,
-0.02445966936647892,
-0.03893108665943146,
0.03372713178396225,
-0.04322506859898567,
0.027091240510344505,
0.031061530113220215,
0.007674555294215679,
0.040781911462545395,
-0.0024142623879015446,
-0.04386887326836586,
-0.010785241611301899,
-0.01974564418196678,
-0.07177364081144333,
0.027487536892294884,
0.03657703846693039,
0.12683844566345215,
0.004281970672309399,
0.040682289749383926,
-0.04851989075541496,
0.12221333384513855,
-0.021090539172291756,
-0.015624509193003178,
0.005714806262403727,
-0.045549988746643066,
0.03269588202238083,
0.027454568073153496,
-0.05251839756965637,
-4.273502440810262e-8,
-0.033303532749414444,
0.09353591501712799,
0.04490327462553978,
0.06909104436635971,
0.015712596476078033,
-0.008392902091145515,
-0.07353797554969788,
0.025752803310751915,
-0.08240509033203125,
-0.06052780523896217,
0.08212066441774368,
-0.039776645600795746,
0.06767594814300537,
0.049808256328105927,
0.038635626435279846,
0.0640290156006813,
-0.025978002697229385,
-0.027361012995243073,
-0.0011286091757938266,
-0.014084350317716599,
-0.011241184547543526,
-0.008402202278375626,
0.007539999671280384,
0.013427844271063805,
-0.06001422181725502,
-0.023620862513780594,
-0.08471157401800156,
-0.0000658997378195636,
0.11684795469045639,
-0.0033333334140479565,
0.018909741193056107,
-0.02139073610305786,
0.004606303293257952,
0.0651479959487915,
0.01611565612256527,
-0.021327590569853783,
-0.04461453855037689,
-0.056121762841939926,
-0.032496847212314606,
-0.02182697504758835,
0.11196599900722504,
0.07719838619232178,
-0.0461992509663105,
0.008422565646469593,
-0.03683803230524063,
0.057581886649131775,
-0.0597962960600853,
-0.14380291104316711,
-0.042881470173597336,
0.05969655141234398,
-0.0688849538564682,
-0.03123529627919197,
-0.061033301055431366,
0.07613973319530487,
0.07497656345367432,
0.03097078576683998,
0.019871650263667107,
-0.046646375209093094,
0.04515695199370384,
0.07786648720502853,
0.0578002855181694,
0.06951174885034561,
-0.01884888857603073,
0.003904201090335846
] | 0 | 0 | 0 | 0 |
|
Be More Diverse than the Most Diverse: Optimal Mixtures of Generative Models via Mixture-UCB Bandit Algorithms | https://openreview.net/forum?id=2Chkk5Ye2s | [
"Parham Rezaei",
"Farzan Farnia",
"Cheuk Ting Li"
] | Poster | The availability of multiple training algorithms and architectures for generative models requires a selection mechanism to form a single model over a group of well-trained generation models. The selection task is commonly addressed by identifying the model that maximizes an evaluation score based on the diversity and quality of the generated data. However, such a best-model identification approach overlooks the possibility that a mixture of available models can outperform each individual model. In this work, we numerically show that a mixture of generative models on benchmark image datasets can indeed achieve a better evaluation score (based on FID and KID scores), compared to the individual models. This observation motivates the development of efficient algorithms for selecting the optimal mixture of the models. To address this, we formulate a quadratic optimization problem to find an optimal mixture model achieving the maximum of kernel-based evaluation scores including kernel inception distance (KID) and Rényi kernel entropy (RKE). To identify the optimal mixture of the models using the fewest possible sample queries, we view the selection task as a multi-armed bandit (MAB) problem and propose the *Mixture Upper Confidence Bound (Mixture-UCB)* algorithm that provably converges to the optimal mixture of the involved models. More broadly, the proposed Mixture-UCB can be extended to optimize every convex quadratic function of the mixture weights in a general MAB setting. We prove a regret bound for the Mixture-UCB algorithm and perform several numerical experiments to show the success of Mixture-UCB in finding the optimal mixture of text and image generative models. The project code is available in the [Mixture-UCB Github repository](https://github.com/Rezaei-Parham/Mixture-UCB). | Multi-Armed Bandits, Evaluation of generative models, Kernel-based evaluation scores, Mixture-UCB, Diversity in data generation | null | 13,785 | 2412.17622 | [
-0.08362787216901779,
-0.06839866936206818,
-0.01129482313990593,
0.03290744125843048,
0.09021089971065521,
0.04356023669242859,
0.00763706024736166,
0.03336525708436966,
0.05764313414692879,
-0.04091421514749527,
-0.017821241170167923,
-0.1179572343826294,
0.08940371870994568,
0.002678374294191599,
-0.026570124551653862,
-0.011892134323716164,
0.11014501750469208,
0.1269494593143463,
-0.02336634323000908,
-0.04442908614873886,
-0.01444651186466217,
0.07094171643257141,
0.0754539966583252,
-0.046336639672517776,
0.05032670497894287,
-0.015319873578846455,
0.045060355216264725,
-0.03364111855626106,
0.016537901014089584,
0.017050810158252716,
0.013431018218398094,
0.037154220044612885,
-0.03026895597577095,
0.0017721030162647367,
-0.0688919872045517,
0.0973241850733757,
-0.05955442786216736,
0.04820069298148155,
0.042496129870414734,
-0.008982121013104916,
-0.04065867140889168,
0.05758744478225708,
-0.015155627392232418,
0.030342599377036095,
0.03926558792591095,
-0.011656437069177628,
-0.06605632603168488,
-0.05553086847066879,
0.10793714970350266,
-0.01810349151492119,
-0.040937814861536026,
-0.010698693804442883,
-0.028562720865011215,
0.058329589664936066,
-0.016883153468370438,
-0.013620207086205482,
0.053822778165340424,
-0.031622014939785004,
0.06767310947179794,
0.006001561414450407,
-0.06107949838042259,
-0.11408063024282455,
-0.0678786188364029,
-0.04614861682057381,
0.03890649974346161,
-0.08595934510231018,
-0.015300188213586807,
0.002686199964955449,
-0.037455081939697266,
-0.013775944709777832,
0.01881529949605465,
0.06619729846715927,
-0.04299871250987053,
0.03983772173523903,
0.026698604226112366,
0.068756602704525,
0.06654678285121918,
0.03647805377840996,
0.06233879551291466,
-0.07702207565307617,
0.0038745987694710493,
0.008475199341773987,
0.012716077268123627,
-0.04956435784697533,
0.14434392750263214,
0.011266768909990788,
0.053910914808511734,
0.04089223966002464,
0.05598720163106918,
-0.007075570523738861,
-0.07385995239019394,
0.062066975980997086,
-0.01465765479952097,
-0.019154952839016914,
0.07047596573829651,
-0.009897252544760704,
0.02192339301109314,
-0.08610761165618896,
0.04436264559626579,
0.06628341227769852,
-0.002965120831504464,
0.10188723355531693,
0.022390371188521385,
-0.07196922600269318,
0.041647277772426605,
-0.03706306964159012,
0.03635558485984802,
0.07022986561059952,
0.1037372350692749,
-0.08195719867944717,
-0.00048419839004054666,
0.05890609323978424,
0.02060706354677677,
0.05444158986210823,
-0.007050654850900173,
0.008396092802286148,
-0.0010757566196843982,
0.03261226415634155,
-0.015417502261698246,
0.12298999726772308,
-0.03446854650974274,
0.005658676382154226,
0.024553416296839714,
-0.007288007531315088,
0.020337603986263275,
-0.07919342815876007,
-0.06872329860925674,
1.4281222294804054e-33,
0.032776959240436554,
-0.005999845452606678,
0.07518837600946426,
0.04834791645407677,
0.013545525260269642,
-0.015170993283390999,
0.0010579259833320975,
-0.01345465611666441,
-0.04800907149910927,
0.015220733359456062,
-0.09084470570087433,
0.02236344665288925,
-0.11697624623775482,
0.13482674956321716,
0.10487643629312515,
0.04955355077981949,
-0.12495243549346924,
0.014875789172947407,
-0.012901675887405872,
-0.02931412309408188,
0.033858153969049454,
-0.025813419371843338,
0.03325682133436203,
-0.039223529398441315,
0.028376739472150803,
-0.0034078038297593594,
0.03415974602103233,
-0.04329951852560043,
-0.04385099560022354,
0.005136486142873764,
-0.1278928518295288,
-0.009566686116158962,
-0.011258826591074467,
0.020002884790301323,
0.04484999552369118,
0.015726039186120033,
-0.028372803702950478,
0.03154822066426277,
-0.05981217697262764,
0.016122428700327873,
-0.05968103185296059,
-0.0038935516495257616,
0.021408729255199432,
-0.022350864484906197,
-0.07103975117206573,
-0.06401112675666809,
-0.045077960938215256,
-0.0017638015560805798,
0.01406545378267765,
-0.008897511288523674,
-0.008713162504136562,
-0.0005200265441089869,
0.05148918554186821,
-0.05394173413515091,
-0.07056005299091339,
0.024394148960709572,
-0.038258299231529236,
0.05668039992451668,
0.0211916733533144,
-0.016372287645936012,
0.0501352995634079,
0.03280894085764885,
0.017136525362730026,
0.028172345831990242,
-0.030368085950613022,
0.003663210663944483,
0.021047992631793022,
0.05617349222302437,
0.030663540586829185,
0.016393475234508514,
0.008188940584659576,
-0.008785421960055828,
-0.014367303811013699,
-0.09910891205072403,
0.05235142260789871,
-0.03229649364948273,
0.07921808212995529,
-0.020109307020902634,
-0.01286307442933321,
-0.014233007095754147,
-0.015555690042674541,
0.01676778867840767,
-0.027881093323230743,
-0.12037736922502518,
0.006559242028743029,
0.03192678838968277,
0.0007638371898792684,
-0.07846740633249283,
-0.050522495061159134,
-0.002345195971429348,
-0.10150498151779175,
0.03376305103302002,
-0.08337105810642242,
-0.040111057460308075,
0.008321834728121758,
-1.7669467211614216e-33,
0.017660202458500862,
0.020216461271047592,
0.035259541124105453,
0.03596281632781029,
0.0019432158442214131,
-0.05524001270532608,
0.03533155471086502,
0.023394275456666946,
-0.05491393059492111,
-0.03547709435224533,
-0.02086304873228073,
0.017845073714852333,
0.04166390374302864,
-0.07679274678230286,
-0.02211037278175354,
0.04562901705503464,
-0.03645482659339905,
0.08054477721452713,
-0.007616363000124693,
0.04071889445185661,
0.05038675665855408,
0.09361021965742111,
-0.021646792069077492,
-0.00346101145260036,
-0.01648339442908764,
0.021763848140835762,
-0.07023385912179947,
0.09232399612665176,
-0.003032929264008999,
-0.0214315764605999,
0.06289246678352356,
-0.03823619335889816,
-0.031278155744075775,
0.018094319850206375,
-0.10479994863271713,
0.007263716775923967,
0.07940156757831573,
0.011079675517976284,
-0.056152548640966415,
0.22312800586223602,
-0.018899863585829735,
0.030152691528201103,
-0.0496392697095871,
0.03538528457283974,
-0.022885017096996307,
-0.033866036683321,
0.0042435661889612675,
-0.053112827241420746,
0.012701263651251793,
-0.008844300173223019,
0.0032008609268814325,
0.051266565918922424,
-0.12609495222568512,
0.08014225959777832,
-0.08879151940345764,
-0.0768977552652359,
-0.03543366491794586,
-0.01638507843017578,
0.06145428493618965,
0.01964532770216465,
-0.015241499058902264,
-0.029816478490829468,
-0.037100423127412796,
0.008610546588897705,
-0.005763270426541567,
-0.07658576220273972,
-0.03386763483285904,
-0.010122288949787617,
-0.03469204530119896,
0.09160741418600082,
-0.05202091857790947,
0.006881938315927982,
0.09013473987579346,
0.035410068929195404,
0.009867497719824314,
0.049768123775720596,
0.05791355296969414,
0.05175806209445,
0.03404553607106209,
-0.09984876960515976,
0.022080691531300545,
0.014344632625579834,
-0.014301193878054619,
0.003374201012775302,
0.03159251809120178,
0.04732766002416611,
0.06945789605379105,
0.01697762869298458,
0.009810471907258034,
-0.06354749947786331,
0.0023086562287062407,
-0.055344849824905396,
0.07529624551534653,
0.040305402129888535,
-0.04777384549379349,
-4.5292932071561154e-8,
0.015692774206399918,
0.02261703461408615,
0.031185230240225792,
-0.01569974608719349,
0.025626618415117264,
-0.0028253088239580393,
-0.08154790103435516,
0.09551767259836197,
-0.044113367795944214,
-0.037390001118183136,
0.0251591969281435,
0.04487498849630356,
-0.05879572778940201,
-0.020094817504286766,
-0.027591576799750328,
-0.0023372326977550983,
-0.04593287780880928,
0.004746885970234871,
0.01892991177737713,
-0.015215310268104076,
0.00624537467956543,
0.0004996058414690197,
0.06594695150852203,
-0.019149620085954666,
0.03213021904230118,
-0.07506618648767471,
-0.03331875801086426,
-0.027790797874331474,
0.01222566794604063,
0.014592430554330349,
-0.041064854711294174,
0.01482434943318367,
-0.032291363924741745,
-0.05224018916487694,
-0.014050869271159172,
0.07326140999794006,
-0.08440893143415451,
-0.015439242124557495,
-0.03398727625608444,
-0.032948773354291916,
0.029859736561775208,
0.04134569689631462,
-0.010355287231504917,
-0.015271843411028385,
0.018829345703125,
0.02606080286204815,
0.029989268630743027,
-0.04283180460333824,
0.027478620409965515,
0.05343644693493843,
0.04025505855679512,
-0.09642592072486877,
-0.007544832769781351,
0.045355234295129776,
0.10017179697751999,
-0.04024812579154968,
-0.025746595114469528,
-0.10209281742572784,
0.1183750182390213,
0.009382165968418121,
0.0020979824475944042,
-0.020108668133616447,
-0.06687045842409134,
0.0037814448587596416
] | https://github.com/rezaei-parham/mixture-ucb | 3 | 0 | 0 | 0 |
Arithmetic Transformers Can Length-Generalize in Both Operand Length and Count | https://openreview.net/forum?id=eIgGesYKLG | [
"Hanseul Cho",
"Jaeyoung Cha",
"Srinadh Bhojanapalli",
"Chulhee Yun"
] | Poster | Transformers often struggle with *length generalization*, meaning they fail to generalize to sequences longer than those encountered during training. While arithmetic tasks are commonly used to study length generalization, certain tasks are considered notoriously difficult, e.g., multi-operand addition (requiring generalization over both the number of operands and their lengths) and multiplication (requiring generalization over both operand lengths). In this work, we achieve approximately 2–3× length generalization on both tasks, which is the first such achievement in arithmetic Transformers. We design task-specific scratchpads enabling the model to focus on a fixed number of tokens per each next-token prediction step, and apply multi-level versions of *Position Coupling* (Cho et al., 2024; McLeish et al., 2024) to let Transformers know the right position to attend to. On the theory side, we prove that a 1-layer Transformer using our method can solve multi-operand addition, up to operand length and operand count that are exponential in embedding dimension. | Length Generalization, Transformers, Scratchpad, Position Coupling, Positional Encoding, Out-of-distribution Generalization, Arithmetic Tasks | We propose combining scratchpad with position coupling, and demonstrate that Transformers can achieve length generalization in both operand length and count for addition problems. | 13,776 | 2410.15787 | [
-0.12702308595180511,
-0.037197619676589966,
0.024250175803899765,
-0.04506266862154007,
-0.040260232985019684,
-0.0017864757683128119,
-0.03873872011899948,
0.04372580349445343,
0.017205867916345596,
-0.05112254619598389,
-0.007774851284921169,
-0.02323545515537262,
0.037514425814151764,
0.06266219913959503,
0.010690036229789257,
0.03402853384613991,
-0.06803910434246063,
0.14123545587062836,
-0.09065697342157364,
-0.06219661980867386,
0.0803324356675148,
-0.0027579155284911394,
-0.009297697804868221,
-0.0005403448012657464,
0.06396843492984772,
0.019431741908192635,
-0.04246445745229721,
-0.07613392174243927,
0.05014107748866081,
0.0006343580316752195,
-0.020035114139318466,
0.0018385782605037093,
0.014028478413820267,
0.05085255578160286,
-0.06722395867109299,
0.0298549085855484,
-0.07574190199375153,
0.018184836953878403,
0.03526844084262848,
0.015652325004339218,
0.06691541522741318,
-0.008390734903514385,
0.07800906151533127,
0.02283003367483616,
0.01028472650796175,
0.013214202597737312,
0.04539713263511658,
-0.07387404143810272,
-0.039658594876527786,
-0.06445436924695969,
0.058269694447517395,
-0.008529840968549252,
-0.06746253371238708,
0.10477845370769501,
-0.02938980795443058,
-0.035519737750291824,
0.023348765447735786,
-0.008593264035880566,
-0.012805381789803505,
0.011976893059909344,
-0.054790906608104706,
-0.04796067997813225,
0.0023536356166005135,
-0.08594117313623428,
-0.0577724389731884,
-0.0069588348269462585,
0.05314890295267105,
0.03505711257457733,
-0.07658564299345016,
0.10413481295108795,
-0.08615920692682266,
0.02238290011882782,
-0.02753087691962719,
0.07979830354452133,
0.07572795450687408,
0.035359300673007965,
0.12332824617624283,
-0.011268984526395798,
-0.002793289953842759,
-0.06229127198457718,
0.01922471821308136,
-0.006288677453994751,
0.037277866154909134,
0.02008202113211155,
0.06884250789880753,
-0.01463607233017683,
-0.03382575139403343,
0.012293723411858082,
0.006348119582980871,
-0.008224794641137123,
-0.025300636887550354,
-0.10644064843654633,
0.0026236812118440866,
0.015902096405625343,
0.03932397440075874,
-0.019753847271203995,
-0.01249589491635561,
-0.008519663475453854,
-0.06314191967248917,
0.052875250577926636,
-0.022931993007659912,
0.01849590241909027,
0.05746703967452049,
-0.027868933975696564,
-0.016701126471161842,
-0.004900881089270115,
-0.06073485314846039,
0.005077788606286049,
-0.004703280981630087,
-0.10811789333820343,
0.06246542930603027,
-0.025639066472649574,
0.015837160870432854,
-0.011988174170255661,
0.013079212047159672,
-0.012382734566926956,
0.0005371180013753474,
-0.013233691453933716,
0.011522618122398853,
0.03713227063417435,
0.04938812926411629,
0.03633565828204155,
-0.03004816174507141,
0.1239854246377945,
0.01267058402299881,
0.027355527505278587,
-0.004175394307821989,
2.155435665493687e-33,
-0.05517085641622543,
0.10586870461702347,
0.014132547192275524,
0.08966822922229767,
0.03143603354692459,
0.060054827481508255,
0.018169302493333817,
0.06786562502384186,
0.02801726758480072,
0.09182994067668915,
-0.00755663588643074,
0.1039159893989563,
-0.006548779085278511,
0.06387735903263092,
0.041537243872880936,
-0.015582341700792313,
0.04023467376828194,
0.0181714054197073,
-0.036257751286029816,
-0.029352139681577682,
0.04887751489877701,
-0.031283944845199585,
-0.03675464913249016,
-0.02778163179755211,
0.0022115525789558887,
-0.08073151856660843,
0.006969684734940529,
-0.10302843153476715,
-0.03796597197651863,
-0.023891961202025414,
-0.12923482060432434,
0.031230082735419273,
-0.024798652157187462,
0.009490955621004105,
0.03344608470797539,
-0.08201965689659119,
0.012931791134178638,
-0.05352729931473732,
0.02379790134727955,
-0.05444489046931267,
-0.06672869622707367,
-0.009757541120052338,
-0.01214708760380745,
-0.04387686029076576,
-0.06976065039634705,
-0.05431248992681503,
0.02628445252776146,
0.04960460588335991,
0.013573627918958664,
-0.04243231192231178,
0.02663728967308998,
0.00269895838573575,
-0.057884134352207184,
-0.0133667616173625,
0.08553618937730789,
-0.006139181554317474,
0.010456913150846958,
0.027357477694749832,
0.06002630293369293,
0.08450312912464142,
-0.04849430173635483,
0.08798076212406158,
0.005387822166085243,
0.006127649452537298,
-0.07143617421388626,
0.022329771891236305,
-0.015642991289496422,
0.01277885865420103,
0.06048945337533951,
-0.009915773756802082,
-0.052670784294605255,
0.0070446692407131195,
-0.0062216902151703835,
-0.022138910368084908,
0.049586012959480286,
-0.055377040058374405,
0.08613527566194534,
-0.09670557081699371,
-0.03926895558834076,
0.014146062545478344,
-0.027836721390485764,
0.02282225713133812,
0.0667252242565155,
-0.005661961622536182,
0.05331887677311897,
-0.038356419652700424,
-0.0009708292782306671,
-0.048478808254003525,
-0.06310815364122391,
-0.1104946881532669,
-0.028454475104808807,
0.013645973987877369,
-0.0332469567656517,
0.002286032773554325,
-0.013176975771784782,
-2.625529264369959e-33,
-0.020499665290117264,
0.028033215552568436,
-0.0568871833384037,
0.01909414306282997,
-0.003209146438166499,
-0.09527728706598282,
0.009426706470549107,
0.009551466442644596,
-0.013054068200290203,
-0.012814118526875973,
-0.031039679422974586,
0.02560078352689743,
0.0013878772733733058,
-0.0212278813123703,
0.033286288380622864,
0.02927473746240139,
-0.12513308227062225,
-0.006995147559791803,
0.03824641555547714,
-0.022046489641070366,
0.07877369970083237,
0.09099521487951279,
-0.1258864849805832,
0.002609358401969075,
-0.04870074242353439,
0.04685470089316368,
-0.08531150221824646,
0.05077841877937317,
0.04364006221294403,
0.042515385895967484,
-0.0019481555791571736,
-0.08068924397230148,
0.025171825662255287,
-0.019562123343348503,
-0.024910427629947662,
-0.007970877923071384,
0.04467417672276497,
0.07603902369737625,
0.03252620995044708,
0.03248101845383644,
0.030047815293073654,
-0.10907576233148575,
0.08200519531965256,
0.05402708798646927,
0.009623151272535324,
-0.03233402967453003,
-0.0016001294134184718,
-0.00034932486596517265,
-0.008009196259081364,
-0.006623516324907541,
-0.03571249172091484,
0.08022325485944748,
-0.12315119057893753,
0.001947091892361641,
-0.01678987592458725,
-0.018094200640916824,
-0.02555193193256855,
-0.029949095100164413,
0.07660999149084091,
-0.05127633735537529,
-0.03284749016165733,
-0.0501035638153553,
0.10020959377288818,
-0.0664098933339119,
0.021777018904685974,
0.007259703241288662,
-0.03946446627378464,
-0.03894145414233208,
-0.0078247906640172,
0.047818657010793686,
0.002744743600487709,
0.02334561012685299,
-0.029072938486933708,
-0.0596764013171196,
-0.07779859751462936,
-0.04991079121828079,
0.051964517682790756,
-0.049076147377491,
0.02744348533451557,
-0.003033288521692157,
-0.04847398400306702,
0.005718858912587166,
0.03150179609656334,
0.012738441117107868,
-0.04899530112743378,
0.05676279217004776,
0.13520729541778564,
0.11985160410404205,
0.04242321476340294,
-0.0350332148373127,
-0.05599084496498108,
0.08040372282266617,
0.046737682074308395,
0.05283554643392563,
-0.03433021903038025,
-4.936113029430089e-8,
-0.05105864629149437,
0.038797639310359955,
-0.0338033102452755,
-0.04963858798146248,
0.08152366429567337,
-0.006367373745888472,
0.031431250274181366,
0.09708084911108017,
-0.012172597460448742,
-0.014117171056568623,
0.08118090033531189,
-0.01006784662604332,
-0.020743353292346,
0.01700633578002453,
-0.018255390226840973,
0.08115077018737793,
-0.002396255498751998,
-0.04851464554667473,
-0.029075978323817253,
0.014510095119476318,
-0.05152429640293121,
0.08548428863286972,
-0.00776815228164196,
-0.050807204097509384,
-0.02730766497552395,
-0.11121966689825058,
-0.005511384457349777,
0.09498903155326843,
0.07212194055318832,
0.019600609317421913,
-0.00971129909157753,
0.03569084405899048,
0.00044993305345997214,
-0.01395617239177227,
-0.009450318291783333,
0.04193631932139397,
0.06468658149242401,
-0.020650234073400497,
0.0016712311189621687,
-0.03973069787025452,
-0.024639902636408806,
0.014416498132050037,
-0.06606201827526093,
0.08795172721147537,
0.06859177350997925,
-0.035269323736429214,
-0.008413915522396564,
-0.14002053439617157,
-0.014469094574451447,
0.02788238227367401,
0.07559159398078918,
0.018977271392941475,
-0.00412330124527216,
0.057691674679517746,
0.07101669907569885,
0.02181691862642765,
-0.05604157596826553,
-0.09580370783805847,
0.05719045177102089,
0.09551579505205154,
-0.07528860867023468,
-0.02168765664100647,
-0.05633716285228729,
0.03946700692176819
] | https://github.com/hanseuljo/position-coupling | 9 | 0 | 0 | 0 |
Stealthy Shield Defense: A Conditional Mutual Information-Based Post-Processing against Black-Box Model Inversion Attacks | https://openreview.net/forum?id=p0DjhjPXl3 | [
"Tianqu Zhuang",
"Hongyao Yu",
"Yixiang Qiu",
"Hao Fang",
"Bin Chen",
"Shu-Tao Xia"
] | Poster | Model inversion attacks (MIAs) aim to reconstruct the private training data by accessing a public model, raising concerns about privacy leakage. Black-box MIAs, where attackers can only query the model and obtain outputs, are closer to real-world scenarios. The latest black-box attacks have outperformed the state-of-the-art white-box attacks, and existing defenses cannot resist them effectively. To fill this gap, we propose Stealthy Shield Defense (SSD), a post-processing algorithm against black-box MIAs. Our idea is to modify the model's outputs to minimize the conditional mutual information (CMI). We mathematically prove that CMI is a special case of information bottlenecks (IB), and thus inherits the advantages of IB---making predictions less dependent on inputs and more dependent on ground truths. This theoretically guarantees our effectiveness, both in resisting MIAs and preserving utility. For minimizing CMI, we formulate a convex optimization problem and solve it via the water-filling method. Adaptive rate-distortion is introduced to constrain the modification to the outputs, and the water-filling is implemented on GPUs to address computation cost. Without the need to retrain the model, our algorithm is plug-and-play and easy to deploy. Experimental results indicate that SSD outperforms existing defenses, in terms of MIA resistance and model's utility, across various attack algorithms, training datasets, and model architectures. Our code is available at https://github.com/ZhuangQu/Stealthy-Shield-Defense. | model inversion attack, model inversion defense, conditional mutual information | null | 13,774 | null | [
-0.10822959989309311,
0.05695129558444023,
0.009606507606804371,
0.051570456475019455,
0.07191921770572662,
0.027909206226468086,
0.019896943122148514,
-0.08783215284347534,
0.05652729049324989,
-0.01923206076025963,
0.007413122337311506,
-0.04867669567465782,
0.09423425793647766,
-0.027009380981326103,
-0.0340845063328743,
0.008812502026557922,
0.050867192447185516,
0.08244422823190689,
-0.10043948143720627,
-0.022090010344982147,
-0.012970871292054653,
0.04983881860971451,
-0.06964756548404694,
0.0322769396007061,
-0.07722805440425873,
-0.005963951349258423,
0.05370284244418144,
-0.08107341825962067,
-0.05660530552268028,
-0.04776342958211899,
0.08212953060865402,
-0.06810685992240906,
0.04313204437494278,
-0.04084455221891403,
-0.02836747094988823,
-0.07088959217071533,
0.0004986487329006195,
-0.03435863181948662,
0.0446445569396019,
-0.03779448941349983,
-0.06665877997875214,
-0.0017802322981879115,
0.004639001097530127,
0.02960088662803173,
-0.02997133694589138,
-0.00010485681559657678,
0.09591833502054214,
0.012150721624493599,
0.003995508886873722,
-0.041144367307424545,
-0.009190633893013,
0.013387732207775116,
-0.04781927168369293,
-0.0024230133276432753,
-0.032155655324459076,
-0.08679396659135818,
-0.005093283019959927,
0.003272930858656764,
-0.04996195808053017,
0.09479591995477676,
-0.05405507981777191,
-0.06943727284669876,
0.01980508863925934,
0.02163637802004814,
0.03497762233018875,
-0.004118615295737982,
0.04007134214043617,
0.06941116601228714,
0.00879413541406393,
0.06773821264505386,
-0.014341862872242928,
0.07636567205190659,
-0.06895371526479721,
-0.04119868949055672,
0.04236050695180893,
0.04132750257849693,
0.03006649576127529,
-0.014892266131937504,
-0.007231913972645998,
-0.015765268355607986,
-0.004560618195682764,
0.022899949923157692,
-0.00044108537258580327,
0.0455474816262722,
0.07549883425235748,
0.012664541602134705,
-0.048774998635053635,
0.03962862864136696,
0.10904549807310104,
-0.010925877839326859,
-0.08154971152544022,
-0.06840287148952484,
-0.014943600632250309,
0.047754354774951935,
0.0693628117442131,
-0.03953735530376434,
-0.00719618983566761,
-0.0485168993473053,
-0.0541081577539444,
0.09954164177179337,
-0.07010411471128464,
0.05282079428434372,
0.03007253259420395,
-0.03293295204639435,
0.09157400578260422,
0.014682535082101822,
0.061777688562870026,
-0.0015858343103900552,
0.03206061199307442,
0.013355964794754982,
0.10363467782735825,
0.02386900596320629,
-0.037595901638269424,
-0.04213907569646835,
0.008709310553967953,
0.014512144029140472,
0.04424963891506195,
0.029667511582374573,
-0.025782769545912743,
0.0363718681037426,
0.0195598304271698,
0.004081779159605503,
0.012868359684944153,
0.03140679746866226,
0.010123138315975666,
-0.03390687331557274,
-0.07646507024765015,
-1.2013341962117577e-33,
0.0024057417176663876,
0.012327209115028381,
-0.015729259699583054,
-0.00726056145504117,
0.044353339821100235,
-0.03241656348109245,
0.0011883298866450787,
0.036322448402643204,
0.022320134565234184,
0.06252681463956833,
-0.01272570714354515,
0.0020036508794873953,
0.015408116392791271,
-0.0002241830516140908,
0.09097184985876083,
0.07910662889480591,
-0.03637038543820381,
0.0568496510386467,
0.019810618832707405,
-0.06150193512439728,
0.12008535861968994,
-0.017618976533412933,
-0.05168702080845833,
-0.08649815618991852,
0.051675017923116684,
-0.04172385111451149,
0.037361059337854385,
-0.02599123679101467,
0.012953638099133968,
0.03272583708167076,
-0.07499457150697708,
0.039230141788721085,
0.03946927934885025,
0.007029632572084665,
0.09139833599328995,
0.027157694101333618,
-0.014960209839046001,
-0.07731853425502777,
0.004492889624089003,
0.055945999920368195,
-0.017137154936790466,
-0.0045626419596374035,
-0.013606113381683826,
-0.04048759490251541,
-0.029834238812327385,
-0.10084298998117447,
-0.02256348542869091,
0.01478864997625351,
-0.041516534984111786,
-0.10082587599754333,
0.03221167251467705,
0.01679840125143528,
-0.08649751543998718,
-0.006989513523876667,
-0.07010985910892487,
-0.0525457039475441,
0.0377749465405941,
0.034080442041158676,
0.04851680248975754,
0.1316973865032196,
-0.05900682508945465,
0.04296427220106125,
0.0003849707718472928,
0.018980836495757103,
-0.030332380905747414,
0.042859721928834915,
-0.04958511143922806,
0.009680341929197311,
0.07130143791437149,
-0.06736879050731659,
-0.06571623682975769,
-0.007712092716246843,
-0.009856578893959522,
-0.046046797186136246,
0.013570619747042656,
-0.0013278297847136855,
0.10347893089056015,
-0.018236981704831123,
0.08524321019649506,
-0.04346543923020363,
0.07091712951660156,
0.02746453508734703,
-0.004249416757375002,
0.021643204614520073,
-0.04904377460479736,
-0.021615829318761826,
-0.014069119468331337,
0.0751330628991127,
-0.1043022945523262,
0.029002763330936432,
-0.08743540942668915,
-0.03786243498325348,
0.012059291824698448,
-0.04617065563797951,
0.012077840976417065,
-6.388519300663869e-34,
-0.08125677704811096,
-0.014034895226359367,
-0.10043897479772568,
-0.03398860991001129,
-0.03140968084335327,
-0.07514389604330063,
0.031605664640665054,
0.022388968616724014,
-0.016146570444107056,
-0.017309097573161125,
-0.10946950316429138,
0.057938698679208755,
0.05066407844424248,
0.020861197263002396,
-0.008409312926232815,
0.07428185641765594,
0.006008394528180361,
0.07929185032844543,
-0.03262168914079666,
0.013285558670759201,
0.0014586423058062792,
0.12193730473518372,
-0.03246837854385376,
0.010697807185351849,
-0.024554062634706497,
-0.0005390788428485394,
-0.04084855690598488,
0.14444933831691742,
0.0872163400053978,
0.07294489443302155,
-0.06602387130260468,
-0.008922162465751171,
-0.020649980753660202,
-0.10790528357028961,
-0.07135754823684692,
0.04115350916981697,
0.0301822442561388,
0.045760635286569595,
-0.009808219969272614,
0.05170100927352905,
0.01843496970832348,
-0.026872597634792328,
-0.10907497256994247,
0.023862628266215324,
-0.031132696196436882,
-0.0322149433195591,
-0.046645015478134155,
0.0060801454819738865,
0.04509995877742767,
-0.014014496468007565,
-0.011514395475387573,
0.0071813552640378475,
-0.014384605921804905,
0.042206548154354095,
-0.07933681458234787,
-0.0895056426525116,
0.003394551109522581,
0.028876662254333496,
0.08340656757354736,
0.08469563722610474,
0.03842337429523468,
0.047970615327358246,
-0.0661158487200737,
-0.022246861830353737,
0.017124181613326073,
-0.04541297256946564,
-0.01696368120610714,
-0.005468336865305901,
0.019680041819810867,
-0.0072334264405071735,
0.050197020173072815,
-0.018456438556313515,
0.0313466414809227,
0.062366798520088196,
0.03493545949459076,
0.026656590402126312,
0.019339503720402718,
-0.03320401906967163,
-0.030694473534822464,
0.15624384582042694,
0.027155084535479546,
-0.00787375122308731,
0.025365453213453293,
-0.011163098737597466,
0.07158127427101135,
0.023692110553383827,
0.07421191036701202,
-0.038647908717393875,
0.026956826448440552,
-0.008558530360460281,
-0.1035194918513298,
0.011037164367735386,
-0.025953462347388268,
0.04862292855978012,
-0.039175570011138916,
-4.6762110628151277e-8,
0.013400252908468246,
-0.0033076549880206585,
0.07120133936405182,
0.03236301615834236,
-0.0036157940048724413,
0.03691677004098892,
-0.02718368172645569,
-0.08155228197574615,
-0.009750346653163433,
-0.08870813250541687,
0.08231104165315628,
0.01725975051522255,
-0.038947366178035736,
-0.06549374014139175,
-0.05913369730114937,
0.05578513815999031,
-0.04037930816411972,
-0.14351804554462433,
-0.017588291317224503,
-0.01848440244793892,
-0.0321086049079895,
-0.04598413035273552,
-0.037010256201028824,
0.0008306705858558416,
0.055344510823488235,
-0.010597187094390392,
0.005081243347376585,
-0.0315670445561409,
0.06307586282491684,
0.09078202396631241,
-0.05115823075175285,
-0.0456780381500721,
0.01870173215866089,
0.07645951211452484,
-0.02625097520649433,
0.10637439787387848,
0.02335151471197605,
-0.03293757885694504,
-0.030042698606848717,
0.022610029205679893,
-0.04623189941048622,
-0.02294105663895607,
-0.033461328595876694,
-0.041480325162410736,
0.02593822404742241,
0.02203783579170704,
0.04615266993641853,
-0.10895302146673203,
0.00721517251804471,
-0.024332543835043907,
0.0877743661403656,
-0.04888789728283882,
-0.012998244725167751,
0.10966084152460098,
0.036137036979198456,
-0.00207010586746037,
0.009371504187583923,
-0.0516088642179966,
-0.005091435741633177,
0.11034563183784485,
0.019174987450242043,
0.03470923751592636,
-0.013137266039848328,
-0.016305072233080864
] | 0 | 0 | 0 | 0 |
|
NovelQA: Benchmarking Question Answering on Documents Exceeding 200K Tokens | https://openreview.net/forum?id=uMEsKEiB7J | [
"Cunxiang Wang",
"Ruoxi Ning",
"Boqi Pan",
"Tonghui Wu",
"Qipeng Guo",
"Cheng Deng",
"Guangsheng Bao",
"Xiangkun Hu",
"Zheng Zhang",
"Qian Wang",
"Yue Zhang"
] | Poster | Recent advancements in Large Language Models (LLMs) have pushed the boundaries of natural language processing, especially in long-context understanding. However, the evaluation of these models' long-context abilities remains a challenge due to the limitations of current benchmarks. To address this gap, we introduce NovelQA, a benchmark tailored for evaluating LLMs with complex, extended narratives. NovelQA, constructed from English novels, offers a unique blend of complexity, length, and narrative coherence, making it an ideal tool for assessing deep textual understanding in LLMs. This paper details the design and construction of NovelQA, focusing on its comprehensive manual annotation process and the variety of question types aimed at evaluating nuanced comprehension. Our evaluation of long-context LLMs on NovelQA reveals significant insights into their strengths and weaknesses. Notably, the models struggle with multi-hop reasoning, detail-oriented questions, and handling extremely long inputs, averaging over 200,000 tokens. Results highlight the need for substantial advancements in LLMs to enhance their long-context comprehension and contribute effectively to computational literary analysis. | Long-context, Large Language Models, Question Answering | We introduce NovelQA, the first question answering dataset on documents with an average length exceeding 200K tokens. | 13,767 | 2403.12766 | [
0.019419148564338684,
-0.004435016307979822,
0.024435266852378845,
0.025263328105211258,
0.07256744056940079,
0.01782277040183544,
-0.02790987305343151,
0.037386514246463776,
0.026149867102503777,
0.0023550833575427532,
-0.10406088083982468,
-0.02622414007782936,
0.014241756871342659,
0.009711623191833496,
0.03567085042595863,
0.04855678603053093,
0.045672234147787094,
0.03124670311808586,
-0.055967602878808975,
-0.022915920242667198,
0.058861684054136276,
0.022152911871671677,
0.07545645534992218,
-0.033989984542131424,
-0.009481696411967278,
-0.04432179406285286,
-0.06508326530456543,
-0.04021184518933296,
0.0025630625896155834,
0.0005806665867567062,
0.039804521948099136,
0.08101920038461685,
-0.04364911839365959,
0.06595152616500854,
0.04058437421917915,
0.05311178043484688,
-0.02436569333076477,
0.05067150294780731,
0.009605355560779572,
-0.03789753094315529,
0.04698776826262474,
-0.04105730354785919,
0.06380277872085571,
0.049006931483745575,
0.11107121407985687,
-0.05098598822951317,
-0.0758189782500267,
0.021434754133224487,
-0.008583789691329002,
0.03173083811998367,
-0.08906339854001999,
0.002314757788553834,
-0.014549028128385544,
0.02045244723558426,
-0.012468570843338966,
0.008112587966024876,
-0.06831970810890198,
0.07138119637966156,
-0.046024080365896225,
-0.09546969085931778,
-0.026658572256565094,
-0.15601901710033417,
-0.04272960498929024,
0.05793457105755806,
-0.0012469700304791331,
-0.06514090299606323,
-0.019936667755246162,
0.025197802111506462,
-0.06823773682117462,
0.04733504354953766,
-0.03120097704231739,
0.021053370088338852,
0.020689718425273895,
0.07961943000555038,
-0.0033180282916873693,
0.04093359038233757,
-0.005188198294490576,
-0.01028519868850708,
0.013006928376853466,
-0.04968608170747757,
0.022901885211467743,
0.03748425096273422,
0.0744820162653923,
0.04386626556515694,
0.10421781241893768,
0.06237746775150299,
0.002773754997178912,
0.01874794065952301,
-0.03128477558493614,
0.04177017882466316,
-0.025433577597141266,
-0.15602605044841766,
0.04007771611213684,
0.029165878891944885,
0.1096157655119896,
0.05746456980705261,
0.022191645577549934,
-0.0412452407181263,
-0.08519052714109421,
0.04557430371642113,
0.09986966848373413,
0.07573453336954117,
-0.028021303936839104,
-0.05692121013998985,
-0.042023058980703354,
-0.015560956671833992,
-0.011129111051559448,
0.02333027683198452,
-0.029845649376511574,
-0.08922219276428223,
0.025351542979478836,
0.05783175304532051,
0.0533844530582428,
0.008790218271315098,
0.062127094715833664,
-0.038074783980846405,
0.06096658110618591,
0.027333805337548256,
0.035499025136232376,
0.03334774449467659,
-0.027730792760849,
0.03413815051317215,
-0.010266735218465328,
-0.033876072615385056,
-0.03570004925131798,
-0.0109753692522645,
0.04958895593881607,
-1.0843052847217524e-33,
0.028200287371873856,
0.08561141788959503,
0.005304673220962286,
0.09458907693624496,
0.008695960976183414,
0.07263670116662979,
0.004362300969660282,
0.08885867148637772,
-0.07911714166402817,
0.03727656602859497,
0.010815584100782871,
0.018678199499845505,
-0.08284604549407959,
0.021528327837586403,
-0.045411303639411926,
-0.00655650720000267,
-0.11711715161800385,
-0.01011959183961153,
-0.04982699453830719,
-0.02601885236799717,
0.003257712349295616,
-0.03230075165629387,
-0.019049882888793945,
-0.001772041548974812,
0.03752338886260986,
-0.0015709211584180593,
0.0825292244553566,
-0.06389212608337402,
-0.08988682180643082,
0.014732947573065758,
-0.08845356851816177,
-0.06448299437761307,
-0.020639067515730858,
-0.017018379643559456,
-0.0017910492606461048,
-0.06226595491170883,
-0.001108868862502277,
-0.0445689857006073,
0.030031584203243256,
-0.01169283501803875,
-0.11283774673938751,
-0.03686372935771942,
0.020631492137908936,
0.029435312375426292,
-0.10816001892089844,
-0.06494024395942688,
-0.03180724382400513,
0.04824388399720192,
-0.039882466197013855,
0.007227674592286348,
0.06222953274846077,
-0.013695616275072098,
-0.04839930683374405,
-0.023211367428302765,
0.07787017524242401,
-0.017446910962462425,
0.000041309282096335664,
0.0105515718460083,
-0.0017533309292048216,
0.0981035828590393,
0.1027778759598732,
0.03700174763798714,
-0.0040833125822246075,
0.07435965538024902,
0.014401574619114399,
0.08055267482995987,
-0.07858022302389145,
0.0695308968424797,
0.03173140808939934,
0.01043691672384739,
-0.014821519143879414,
0.02022591046988964,
0.07025182992219925,
-0.022847367450594902,
-0.004230370745062828,
0.03827047348022461,
0.017148446291685104,
-0.06255558133125305,
0.02526102215051651,
0.007461585104465485,
0.032598111778497696,
-0.010427854023873806,
0.01359972357749939,
-0.033293865621089935,
-0.09806215763092041,
-0.05383855476975441,
0.062265753746032715,
-0.08825864642858505,
-0.0053714569658041,
-0.05134916678071022,
0.029904458671808243,
-0.00020980236877221614,
-0.03510798513889313,
-0.09457724541425705,
0.027026211842894554,
1.1325874294848316e-33,
0.032615333795547485,
-0.04735049605369568,
-0.10853457450866699,
0.03580239787697792,
0.00995106901973486,
-0.10984845459461212,
-0.0817732885479927,
0.05048288777470589,
0.023386994376778603,
-0.07125156372785568,
-0.03865551948547363,
-0.01704944483935833,
0.017722975462675095,
0.00516785541549325,
0.08328309655189514,
-0.028006166219711304,
-0.015393785201013088,
-0.04359596222639084,
0.03727760910987854,
0.09477971494197845,
-0.0004883147194050252,
0.054307449609041214,
-0.054535385221242905,
0.013110953383147717,
-0.020990386605262756,
0.05790279433131218,
-0.0365135557949543,
0.020210638642311096,
-0.04238416999578476,
-0.008901713415980339,
-0.03746585175395012,
-0.05489451438188553,
-0.02924356237053871,
-0.011978032998740673,
-0.057724129408597946,
0.05057555437088013,
0.1112460345029831,
-0.062079478055238724,
-0.03486168012022972,
0.010605177842080593,
0.0765766054391861,
-0.05371749773621559,
-0.05233049765229225,
-0.015087136998772621,
-0.05621519312262535,
0.007744616828858852,
-0.06477130949497223,
0.0073462240397930145,
0.04575526341795921,
0.028484897688031197,
-0.02436213754117489,
-0.009531981311738491,
-0.032401882112026215,
0.06850399076938629,
-0.051626548171043396,
-0.059665072709321976,
-0.08073509484529495,
-0.025391437113285065,
0.003558897180482745,
-0.05988098308444023,
-0.0365242138504982,
0.08279489725828171,
-0.011019805446267128,
-0.01762021705508232,
0.06662876158952713,
-0.08927834779024124,
-0.03691694512963295,
-0.021360967308282852,
-0.04534444212913513,
-0.0005379145150072873,
-0.0768878236413002,
-0.06597244739532471,
-0.02803151123225689,
-0.03064717911183834,
0.03641112893819809,
0.02279425412416458,
0.019134521484375,
-0.10084855556488037,
-0.024327127262949944,
-0.061174824833869934,
0.00043194773024879396,
-0.049842413514852524,
0.005569419823586941,
0.10582243651151657,
0.09709877520799637,
0.06636875122785568,
0.027911106124520302,
0.13276462256908417,
0.02098492905497551,
0.10705552995204926,
0.05089215189218521,
-0.004923945292830467,
-0.015001736581325531,
0.07513311505317688,
-0.08448866754770279,
-3.9962937847803914e-8,
-0.04650422930717468,
0.04832816496491432,
-0.10563994944095612,
0.03350977972149849,
-0.0047422111965715885,
-0.0006145001389086246,
-0.040156181901693344,
0.08488348871469498,
-0.009345054626464844,
0.024497708305716515,
0.010246671736240387,
0.030064299702644348,
-0.01583751104772091,
-0.03095095045864582,
-0.006880140397697687,
0.07614383101463318,
0.0375637412071228,
0.012251203879714012,
-0.015488854609429836,
0.003970467485487461,
0.12194062024354935,
0.06786051392555237,
-0.04895120486617088,
0.049817848950624466,
-0.05966067314147949,
0.041685618460178375,
-0.027949536219239235,
0.07808618992567062,
0.0756339430809021,
-0.04985049366950989,
-0.04809660464525223,
0.04929264262318611,
-0.05389440804719925,
-0.002301372354850173,
-0.030476680025458336,
0.07867531478404999,
0.03841860219836235,
0.051472458988428116,
0.009060663171112537,
-0.020024506375193596,
-0.008635153993964195,
0.005848821718245745,
-0.03238234668970108,
0.05582377314567566,
0.021345723420381546,
-0.05485381558537483,
-0.06255987286567688,
-0.04987480491399765,
0.023861194029450417,
-0.04019809141755104,
-0.016435490921139717,
-0.06531737744808197,
0.034520186483860016,
0.05583365261554718,
0.01230752281844616,
0.01643875427544117,
-0.030766978859901428,
0.06807202100753784,
-0.004392687231302261,
0.013550990261137486,
0.017456410452723503,
0.01574510894715786,
0.011755004525184631,
0.01863952726125717
] | https://github.com/novelqa/novelqa.github.io | 19 | 0 | 0 | 0 |
Look Before You Leap: Universal Emergent Mechanism for Retrieval in Language Models | https://openreview.net/forum?id=eIB1UZFcFg | [
"Alexandre Variengien",
"Eric Winsor"
] | Poster | When solving challenging problems, language models (LMs) are able to identify relevant information from long and complicated contexts. To study how LMs solve retrieval tasks in diverse situations, we introduce ORION, a collection of structured retrieval tasks spanning six domains, from text understanding to coding. Each task in ORION can be represented abstractly by a request (e.g. a question) that retrieves an attribute (e.g. the character name) from a context (e.g. a story). We apply causal analysis on 18 open-source language models with sizes ranging from 125 million to 70 billion parameters. We find that LMs internally decompose retrieval tasks in a modular way: middle layers at the last token position process the request, while late layers retrieve the correct entity from the context. After causally enforcing this decomposition, models are still able to solve the original task, preserving 70% of the original correct token probability in 98 of the 106 studied model-task pairs. We connect our macroscopic decomposition with a microscopic description by performing a fine-grained case study of a question-answering task on Pythia-2.8b. Building on our high-level understanding, we demonstrate a proof of concept application for scalable internal oversight of LMs to mitigate prompt-injection while requiring human supervision on only a single input. Our solution improves accuracy drastically (from 15.5% to 97.5% on Pythia-12b). This work presents evidence of a universal emergent modular processing of tasks across varied domains and models and is a pioneering effort in applying interpretability for scalable internal oversight of LMs. | Interpretability, LLM, Universality | We show that LM decompose retrieval internally by first compiling a representation of the query, and then looking for matching elements in the context. | 13,760 | null | [
-0.05071971192955971,
-0.01789005473256111,
0.06903406977653503,
0.04485827684402466,
0.047800805419683456,
-0.028177570551633835,
-0.022017089650034904,
0.029993832111358643,
0.08317022025585175,
-0.006222885102033615,
-0.03165459632873535,
-0.030041644349694252,
0.04697199538350105,
0.042237527668476105,
0.011572327464818954,
-0.019114816561341286,
0.07353025674819946,
0.03282012417912483,
-0.1460583657026291,
-0.0169464610517025,
0.04522712156176567,
0.038034871220588684,
0.041093043982982635,
0.021233873441815376,
-0.01590561494231224,
-0.029901321977376938,
0.006359250750392675,
-0.06811531633138657,
0.028874974697828293,
-0.04310670495033264,
0.032150499522686005,
0.06884175539016724,
-0.04585588723421097,
0.07322842627763748,
0.006050546187907457,
0.11083034425973892,
-0.09091437608003616,
-0.04301171749830246,
0.025735486298799515,
-0.10706044733524323,
0.019590163603425026,
0.010201063007116318,
0.018705984577536583,
0.051059823483228683,
0.11485710740089417,
-0.003133127000182867,
0.0013317043194547296,
0.013459541834890842,
0.00352477771230042,
-0.016551904380321503,
-0.14371740818023682,
-0.010918446816504002,
0.02605086751282215,
0.05041780695319176,
-0.0040040514431893826,
0.003616678761318326,
0.03940659016370773,
-0.025558138266205788,
-0.04564663767814636,
-0.10021866112947464,
-0.11388689279556274,
-0.05933914706110954,
-0.08969340473413467,
-0.016892462968826294,
0.03067520260810852,
-0.05467679351568222,
0.017141249030828476,
0.023580798879265785,
-0.00863422267138958,
0.02038823440670967,
-0.0023713880218565464,
0.008167346008121967,
-0.000666230742353946,
0.08744172751903534,
0.044564567506313324,
0.05843537673354149,
0.014033355750143528,
0.005071024876087904,
-0.015149790793657303,
-0.028551440685987473,
0.05578208714723587,
0.0009412362705916166,
0.07297445833683014,
0.025608154013752937,
0.11242755502462387,
-0.01579153724014759,
0.03483172878623009,
0.039574362337589264,
0.03353940695524216,
0.043878037482500076,
-0.024573536589741707,
-0.13309626281261444,
0.11342699825763702,
-0.013929017819464207,
0.06830756366252899,
0.03650029003620148,
0.06340862065553665,
-0.01876162551343441,
0.05694706737995148,
0.07619238644838333,
0.050344593822956085,
0.05326668918132782,
0.03960521146655083,
-0.060932308435440063,
-0.04856806993484497,
0.03507821261882782,
-0.0074266185984015465,
0.07599703222513199,
-0.0004361487808637321,
-0.08362242579460144,
0.0034906878136098385,
-0.011427344754338264,
0.014591841958463192,
-0.02216535061597824,
0.04418908804655075,
-0.07325772941112518,
0.056926798075437546,
-0.03218043968081474,
0.00786386989057064,
-0.002810055622830987,
-0.012987173162400723,
0.0587492436170578,
0.01130987610667944,
-0.044580064713954926,
-0.037480734288692474,
0.008676426485180855,
-0.03608186915516853,
3.282043823576916e-34,
0.0675203800201416,
0.052519794553518295,
-0.009709655307233334,
0.06688522547483444,
-0.008454904891550541,
0.014314466156065464,
-0.0018883482553064823,
0.0497128926217556,
-0.0003265048726461828,
-0.02640339732170105,
-0.006835396401584148,
0.053557369858026505,
-0.09928464144468307,
0.004170526750385761,
0.01952229253947735,
0.014080134220421314,
-0.09783507883548737,
0.023656735196709633,
-0.055775243788957596,
0.01977301947772503,
0.015229037031531334,
-0.028219979256391525,
-0.00232062884606421,
-0.017378730699419975,
0.003953731618821621,
0.019558807834982872,
0.07260739058256149,
-0.1331985890865326,
-0.11045904457569122,
0.0004164478159509599,
-0.09008505940437317,
-0.04196500405669212,
-0.015308442525565624,
0.008160684257745743,
0.01839568093419075,
-0.04138030484318733,
-0.016285190358757973,
-0.05599495396018028,
-0.011996498331427574,
-0.024878792464733124,
-0.0410226471722126,
-0.0061652883887290955,
-0.014931999146938324,
-0.0034298859536647797,
-0.10274925082921982,
-0.08234191685914993,
-0.020199086517095566,
0.005484696943312883,
-0.044421058148145676,
-0.043290987610816956,
0.11971701681613922,
0.01852498948574066,
0.012731978669762611,
-0.07705393433570862,
0.05694969743490219,
0.04657765477895737,
-0.031179746612906456,
-0.002147574443370104,
0.016991926357150078,
0.053730107843875885,
0.11682403832674026,
0.021520523354411125,
0.02088971622288227,
0.09158536046743393,
0.02906213328242302,
0.05968434363603592,
-0.07460188120603561,
0.08976729959249496,
0.06286866217851639,
0.03715845197439194,
-0.04328933358192444,
-0.0063500977121293545,
0.05429792031645775,
-0.05843965709209442,
0.04160400852560997,
0.040192611515522,
-0.007601158693432808,
-0.1321321278810501,
0.017320573329925537,
0.026816517114639282,
0.009616951458156109,
-0.09863507747650146,
-0.016228897497057915,
-0.02164616622030735,
-0.0778697058558464,
-0.021901695057749748,
0.04201370105147362,
-0.13493318855762482,
-0.024740539491176605,
-0.0969245582818985,
0.032837677747011185,
-0.04029446840286255,
-0.008013701066374779,
0.000029518214432755485,
0.018709953874349594,
-7.215847262835091e-35,
-0.006238814443349838,
-0.0916825607419014,
-0.022003065794706345,
-0.04031059145927429,
-0.031146133318543434,
-0.10192923247814178,
-0.031189177185297012,
-0.012797465547919273,
-0.03447619825601578,
-0.0010525749530643225,
-0.09753738343715668,
-0.0055871885269880295,
0.021774224936962128,
0.012043933384120464,
0.056663479655981064,
0.03869450092315674,
0.030141713097691536,
0.017522502690553665,
0.01880386285483837,
0.10266353189945221,
-0.027791492640972137,
0.08422020822763443,
-0.09349934756755829,
0.01533557754009962,
0.037359725683927536,
0.051751211285591125,
0.007071857340633869,
0.0001455532619729638,
0.022941892966628075,
-0.017019473016262054,
-0.057792894542217255,
-0.011169823817908764,
-0.006247745826840401,
-0.02029852382838726,
-0.0720885843038559,
0.06262718886137009,
0.05354362726211548,
-0.047126978635787964,
-0.023336563259363174,
0.01928100734949112,
0.09205247461795807,
-0.04264247044920921,
-0.0525076687335968,
0.014389445073902607,
0.023173285648226738,
-0.003293316112831235,
-0.1342829316854477,
0.045640476047992706,
0.021948521956801414,
0.01038600504398346,
0.0077911983244121075,
0.006437667179852724,
-0.01630406081676483,
0.02567816898226738,
-0.03637681528925896,
-0.06580032408237457,
-0.039818160235881805,
-0.02131626382470131,
0.06684523820877075,
0.008037949912250042,
-0.035081908106803894,
-0.0182536318898201,
-0.0006187618710100651,
0.04482564330101013,
0.06326666474342346,
-0.08310048282146454,
0.0007802758482284844,
-0.037601813673973083,
-0.07294099777936935,
-0.05619559437036514,
0.0006003109738230705,
-0.06318075954914093,
0.010464530438184738,
-0.01161517295986414,
0.08155795931816101,
0.06908464431762695,
0.006061098538339138,
-0.02431861124932766,
-0.04795265197753906,
-0.0410236120223999,
-0.026163747534155846,
-0.03481940180063248,
0.024000104516744614,
0.06734379380941391,
0.06072328984737396,
-0.023276375606656075,
0.018267832696437836,
0.10181225091218948,
0.04403143748641014,
-0.01152738370001316,
-0.030843785032629967,
-0.03763878345489502,
0.03292336314916611,
0.08203331381082535,
-0.012700172141194344,
-4.6600195702239944e-8,
-0.01608896628022194,
0.08230340480804443,
-0.07218579202890396,
0.03307533264160156,
0.027542654424905777,
0.016644958406686783,
-0.10957226157188416,
0.08861272782087326,
-0.03165535256266594,
0.01833057403564453,
0.03246108815073967,
-0.01141231507062912,
-0.06495659798383713,
-0.012197873555123806,
0.023457201197743416,
0.07707761973142624,
0.018977288156747818,
0.013192309997975826,
-0.029404563829302788,
-0.011358394287526608,
0.10572067648172379,
0.04329199716448784,
-0.015035111457109451,
0.07428567111492157,
-0.014183739200234413,
-0.002927228109911084,
-0.05685751512646675,
0.11538916826248169,
0.03169398754835129,
-0.04112677276134491,
-0.006932303309440613,
0.03995467349886894,
-0.05837724730372429,
0.000486541073769331,
-0.01952959969639778,
0.07789535075426102,
0.04324813932180405,
-0.06717290729284286,
-0.011892054229974747,
-0.0033255547750741243,
0.060461658984422684,
0.050034426152706146,
-0.01688370481133461,
0.052410468459129333,
0.05637521669268608,
-0.0030350498855113983,
-0.05233045294880867,
-0.04315730556845665,
0.057124219834804535,
0.018931051716208458,
-0.05239866301417351,
-0.026954559609293938,
-0.06226664036512375,
0.0855586975812912,
-0.004286708310246468,
-0.004222911782562733,
0.021040599793195724,
0.027865329757332802,
0.017130715772509575,
-0.05480866879224777,
0.11010324209928513,
0.04858420044183731,
-0.02209564484655857,
0.023192444816231728
] | 0 | 0 | 0 | 0 |
|
A Multi-Power Law for Loss Curve Prediction Across Learning Rate Schedules | https://openreview.net/forum?id=KnoS9XxIlK | [
"Kairong Luo",
"Haodong Wen",
"Shengding Hu",
"Zhenbo Sun",
"Maosong Sun",
"Zhiyuan Liu",
"Kaifeng Lyu",
"Wenguang Chen"
] | Poster | Training large models is both resource-intensive and time-consuming, making it crucial to understand the quantitative relationship between model performance and hyperparameters. In this paper, we derive an empirical law that predicts pretraining loss for large language models for every intermediate training step across various learning rate schedules, including constant, cosine, and step decay schedules. Our proposed law takes a multi-power form, combining a power law based on the sum of learning rates and additional power laws to account for a loss reduction effect as learning rate decays. We validate this law extensively on Llama-2 models of varying sizes and demonstrate that, after fitting on a few learning rate schedules, it accurately predicts the loss curves for unseen schedules of different shapes and horizons. Moreover, by minimizing the predicted final pretraining loss across learning rate schedules, we are able to find a schedule that outperforms the widely-used cosine learning rate schedule. Interestingly, this automatically discovered schedule bears some resemblance to the recently proposed Warmup-Stable-Decay (WSD) schedule (Hu et al, 2024) but achieves a slightly lower final loss. We believe these results could offer valuable insights for understanding the dynamics of pretraining and for designing learning rate schedules to improve efficiency. | Large language model, Learning rate scheduler, Scaling Law, Hyperparameter optimization | Loss curve prediction and optimized learning rate schedule | 13,754 | 2503.12811 | [
-0.08972218632698059,
-0.015865035355091095,
0.021371586248278618,
0.08443307131528854,
0.07736677676439285,
-0.0010595070198178291,
-0.11925346404314041,
-0.03856164962053299,
0.0856919065117836,
0.015246013179421425,
-0.04739729315042496,
0.018870508298277855,
-0.009792204946279526,
0.021394284442067146,
0.03902682289481163,
0.008148670196533203,
-0.037165869027376175,
0.07285435497760773,
-0.0863799899816513,
-0.036180317401885986,
0.07490065693855286,
0.012328855693340302,
0.06761819124221802,
0.028422469273209572,
0.0873381569981575,
-0.01634414680302143,
-0.01631232351064682,
-0.03323959559202194,
-0.00023793878790456802,
-0.009981479495763779,
0.007276938296854496,
-0.04230107367038727,
-0.018462151288986206,
-0.018587952479720116,
-0.03962080925703049,
0.006593402475118637,
0.01997440680861473,
0.0053735822439193726,
0.04117392376065254,
0.06579753756523132,
0.005365799181163311,
0.0227296631783247,
0.05902814865112305,
0.036812376230955124,
0.031217986717820168,
-0.010109043680131435,
-0.0002868439187295735,
-0.04370466247200966,
-0.07538174837827682,
0.05203121155500412,
-0.03190450742840767,
-0.12078417092561722,
-0.028496986255049706,
0.09332670271396637,
-0.03452453762292862,
0.060109205543994904,
0.026201652362942696,
0.042719267308712006,
-0.021413352340459824,
0.01506037451326847,
-0.06628284603357315,
-0.11215701699256897,
-0.08485611528158188,
-0.03598179668188095,
-0.039510540664196014,
0.01361721009016037,
0.056037548929452896,
0.07412310689687729,
-0.029077788814902306,
0.06637469679117203,
-0.0515366792678833,
0.044882871210575104,
-0.03226746246218681,
0.028307916596531868,
-0.0032085380516946316,
0.012807228602468967,
0.08107151091098785,
0.01569199748337269,
0.048985909670591354,
-0.09844347089529037,
-0.036031316965818405,
-0.030014468356966972,
0.007513625547289848,
-0.08209299296140671,
0.115473672747612,
-0.034701015800237656,
0.08680238574743271,
0.008276781998574734,
0.06923913955688477,
-0.0023898696526885033,
-0.03907248005270958,
-0.012797437608242035,
0.028921518474817276,
0.005077732726931572,
-0.012273191474378109,
0.056407321244478226,
0.001078197150491178,
-0.011427627876400948,
-0.03239941596984863,
0.06016836687922478,
-0.025258412584662437,
0.00011393570457585156,
0.030398854985833168,
-0.0026787607930600643,
0.036692433059215546,
-0.051557738333940506,
0.033226873725652695,
0.06372260302305222,
0.041307512670755386,
0.0019599739462137222,
0.07541251927614212,
0.07276029139757156,
-0.020905552431941032,
0.05402055382728577,
0.047269031405448914,
0.1195564866065979,
0.005063900724053383,
0.010634676553308964,
-0.02306746318936348,
0.09922533482313156,
-0.040886323899030685,
0.018383389338850975,
0.01321263425052166,
-0.03888187184929848,
0.00008625130431028083,
-0.07721534371376038,
0.009583418257534504,
2.4424269174461868e-33,
0.05918740853667259,
0.026227159425616264,
-0.028607677668333054,
-0.053004924207925797,
0.040994513779878616,
-0.0695260614156723,
-0.02692314423620701,
0.05335632711648941,
0.0019182214746251702,
-0.027027716860175133,
-0.05006707087159157,
-0.01423158310353756,
-0.040336884558200836,
0.06052963063120842,
0.02100854367017746,
-0.009742476977407932,
0.014884913340210915,
0.06762780249118805,
0.05930263549089432,
-0.056033939123153687,
0.08325416594743729,
-0.03744674474000931,
-0.006371155846863985,
-0.12219957262277603,
-0.04807569086551666,
0.07842906564474106,
0.06764606386423111,
-0.009540703147649765,
-0.07276025414466858,
0.023837776854634285,
-0.013336376287043095,
0.006054973229765892,
-0.043581437319517136,
0.01661459170281887,
0.03162478283047676,
-0.018266450613737106,
-0.029498910531401634,
0.003966685850173235,
0.07477939128875732,
-0.03374604135751724,
-0.020320001989603043,
0.013164881616830826,
0.05487871915102005,
-0.07281135767698288,
-0.03961695730686188,
-0.12682217359542847,
0.06286046653985977,
-0.03245309367775917,
-0.07111453264951706,
-0.04595506936311722,
0.012852423824369907,
-0.06992587447166443,
-0.06100538745522499,
0.00856042094528675,
-0.04413607344031334,
0.04603852704167366,
0.009497075341641903,
0.02040751650929451,
0.03194176405668259,
0.08444097638130188,
0.03523418679833412,
0.04864920675754547,
0.020359793677926064,
0.06815926730632782,
-0.03767158091068268,
0.007331362925469875,
-0.03309793025255203,
-0.0014085571747273207,
0.047816235572099686,
-0.025143837556242943,
0.024728914722800255,
-0.02999085746705532,
-0.06510568410158157,
-0.08663596212863922,
0.11347094178199768,
-0.05294593423604965,
0.11254021525382996,
-0.06570341438055038,
0.008314055390655994,
0.025207923725247383,
0.03427769988775253,
-0.023764293640851974,
0.04221801459789276,
-0.14845150709152222,
-0.039068542420864105,
-0.028561783954501152,
0.028065582737326622,
-0.08954509347677231,
-0.03547009080648422,
-0.011812390759587288,
-0.05624111741781235,
-0.05823386460542679,
-0.013062451966106892,
-0.006396578159183264,
0.028344104066491127,
-2.308595048895146e-33,
-0.06123507767915726,
-0.012776682153344154,
-0.021121423691511154,
0.1375255286693573,
-0.02738550491631031,
-0.03295309096574783,
-0.020857831463217735,
0.05262705683708191,
-0.06027236953377724,
-0.06271300464868546,
0.03364574536681175,
-0.049148108810186386,
-0.04958168417215347,
-0.021360833197832108,
0.07738850265741348,
-0.026656022295355797,
0.02146218903362751,
-0.0030339788645505905,
0.07400734722614288,
-0.010994549840688705,
0.0631740465760231,
0.0410805344581604,
-0.13482335209846497,
-0.011442869901657104,
-0.028426190838217735,
-0.0016740477876737714,
-0.0508912168443203,
0.08190339058637619,
0.028746023774147034,
0.051907263696193695,
-0.06882312148809433,
-0.03310607373714447,
-0.03750210627913475,
-0.002323630265891552,
-0.0921674370765686,
0.06281815469264984,
-0.01454460620880127,
0.02813822031021118,
-0.0387384295463562,
0.12109900265932083,
0.12762397527694702,
0.010569194331765175,
-0.03928673267364502,
-0.016216399148106575,
-0.04522500932216644,
0.005422385409474373,
-0.07114321738481522,
0.0481082908809185,
-0.010407296940684319,
-0.00005278418757370673,
0.029128141701221466,
-0.03176293894648552,
-0.061598990112543106,
0.07458838075399399,
-0.0591554194688797,
-0.05930483713746071,
-0.03287672996520996,
-0.03079477697610855,
-0.045837633311748505,
0.0092343520373106,
-0.05818905308842659,
0.0023423496168106794,
0.0034000524319708347,
0.022233974188566208,
0.0068131787702441216,
-0.03922390565276146,
0.005901752971112728,
0.027377899736166,
0.03594832122325897,
-0.003591877408325672,
0.057854197919368744,
0.027632052078843117,
-0.01282409392297268,
-0.010459748096764088,
-0.07192163914442062,
0.009820587001740932,
0.011040262877941132,
-0.0442376434803009,
-0.07250988483428955,
-0.0017492210026830435,
-0.05884382873773575,
0.03218517079949379,
-0.018994199112057686,
0.03618345409631729,
0.018201353028416634,
0.0833982303738594,
0.02496502175927162,
0.10030828416347504,
0.051481056958436966,
0.010712606832385063,
0.012386471964418888,
0.02963673509657383,
0.022678693756461143,
0.07158250361680984,
-0.012702804058790207,
-4.635200312463894e-8,
-0.05074213445186615,
0.062108028680086136,
-0.002838104497641325,
0.06210268288850784,
0.09982718527317047,
-0.04871254786849022,
-0.021166466176509857,
0.042076509445905685,
0.020871171727776527,
0.02547142654657364,
0.1599971204996109,
0.008296907879412174,
0.03115951083600521,
-0.04461383447051048,
-0.03286394849419594,
0.07709341496229172,
0.023307519033551216,
0.03879622742533684,
-0.010533136315643787,
-0.044417597353458405,
-0.010814782232046127,
-0.029846066609025,
0.09224963933229446,
-0.014528514817357063,
0.032714150846004486,
-0.01869959384202957,
0.012087956070899963,
0.12213069200515747,
0.054197046905756,
0.005043397657573223,
-0.08062998205423355,
0.01308764610439539,
-0.024305615574121475,
-0.04906636103987694,
-0.0042702932842075825,
0.07077872008085251,
0.00969792902469635,
-0.07787342369556427,
0.05457494407892227,
0.11417565494775772,
0.0022451423574239016,
-0.014116674661636353,
-0.05122934281826019,
0.023123249411582947,
0.041484225541353226,
-0.017689457163214684,
0.0009327879524789751,
-0.1077975407242775,
-0.038908861577510834,
-0.034493520855903625,
0.018065566197037697,
-0.027528179809451103,
-0.01644439809024334,
-0.005125471856445074,
0.07318706810474396,
0.007553333416581154,
-0.05882230028510094,
-0.03034994564950466,
-0.04698162153363228,
0.045918505638837814,
0.03513581305742264,
-0.042454350739717484,
-0.09789793938398361,
0.005196535959839821
] | https://github.com/thu-yao-01-luo/multipowerlaw | 7 | 0 | 0 | 0 |
LLaVA-Mini: Efficient Image and Video Large Multimodal Models with One Vision Token | https://openreview.net/forum?id=UQJ7CDW8nb | [
"Shaolei Zhang",
"Qingkai Fang",
"Zhe Yang",
"Yang Feng"
] | Poster | The advent of real-time large multimodal models (LMMs) like GPT-4o has sparked considerable interest in efficient LMMs. LMM frameworks typically encode visual inputs into vision tokens (continuous representations) and integrate them and textual instructions into the context of large language models (LLMs), where large-scale parameters and numerous context tokens (predominantly vision tokens) result in substantial computational overhead. Previous efforts towards efficient LMMs always focus on replacing the LLM backbone with smaller models, while neglecting the crucial issue of token quantity. In this paper, we introduce LLaVA-Mini, an efficient LMM with minimal vision tokens. To achieve a high compression ratio of vision tokens while preserving visual information, we first analyze how LMMs understand vision tokens and find that most vision tokens only play a crucial role in the early layers of LLM backbone, where they mainly fuse visual information into text tokens. Building on this finding, LLaVA-Mini introduces modality pre-fusion to fuse visual information into text tokens in advance, thereby facilitating the extreme compression of vision tokens fed to LLM backbone into one token. LLaVA-Mini is a unified large multimodal model that can support the understanding of images, high-resolution images, and videos in an efficient manner. Experiments across 11 image-based and 7 video-based benchmarks demonstrate that LLaVA-Mini outperforms LLaVA-v1.5 with just 1 vision token instead of 576. Efficiency analyses reveal that LLaVA-Mini can reduce FLOPs by 77%, deliver low-latency responses within 40 milliseconds, and process over 10,000 frames of video on the GPU hardware with 24GB of memory. | Large Multimodal Models, Large Language Models | null | 13,752 | null | [
0.02409641444683075,
-0.05561100319027901,
0.029878973960876465,
-0.01477209571748972,
0.1622084528207779,
-0.02341500110924244,
-0.051242146641016006,
0.030534885823726654,
0.07229170948266983,
-0.056438133120536804,
-0.0388859324157238,
-0.042065251618623734,
0.04792212322354317,
0.057855263352394104,
0.026164423674345016,
0.020468756556510925,
0.06982416659593582,
0.13957644999027252,
-0.14430420100688934,
-0.03598221391439438,
0.023747334256768227,
-0.11471844464540482,
0.004117058590054512,
-0.05541059002280235,
0.09948156028985977,
0.05960644409060478,
-0.02643548510968685,
-0.08362115919589996,
0.06728509813547134,
-0.044189583510160446,
-0.009393704123795033,
0.010356446728110313,
-0.001815253053791821,
0.05383765324950218,
-0.10049252212047577,
0.0858592689037323,
-0.008953158743679523,
0.008376761339604855,
-0.03998996317386627,
-0.07627377659082413,
-0.007870685309171677,
0.019843867048621178,
-0.011380008421838284,
0.0075155990198254585,
0.06311012804508209,
0.0028510421980172396,
0.008134624920785427,
-0.03578045964241028,
-0.005444337613880634,
-0.027486206963658333,
-0.07448608428239822,
-0.049647144973278046,
-0.061599887907505035,
0.012898534536361694,
-0.06593401730060577,
-0.005377673543989658,
-0.0032387387473136187,
-0.04037627950310707,
0.039366401731967926,
0.0173301063477993,
0.01908012293279171,
-0.035532087087631226,
0.013310338370501995,
0.04314848780632019,
0.03286369889974594,
-0.06913116574287415,
0.09619875252246857,
0.002548136981204152,
0.013409141451120377,
0.00985761173069477,
-0.04112911969423294,
0.03238346055150032,
-0.047364089637994766,
0.02750176563858986,
-0.04138213023543358,
0.001689553027972579,
0.06647597253322601,
0.01142781600356102,
0.0790722668170929,
-0.030934417620301247,
0.09303206950426102,
0.026346705853939056,
0.053447164595127106,
0.00561561668291688,
0.10241170972585678,
0.004828034900128841,
-0.025686856359243393,
0.049506284296512604,
0.01002614013850689,
0.007994831539690495,
-0.08676256984472275,
-0.05560632422566414,
0.058287326246500015,
-0.004365266766399145,
0.04128512740135193,
-0.06378201395273209,
-0.028286578133702278,
-0.08255140483379364,
-0.019244439899921417,
0.08150995522737503,
0.027459964156150818,
0.007484911009669304,
0.00276523781940341,
-0.013858729973435402,
0.043786849826574326,
-0.043580494821071625,
0.07977057993412018,
0.04583115875720978,
0.017651140689849854,
0.0007476687314920127,
0.08325491845607758,
0.04173307493329048,
-0.001808498171158135,
-0.030930791050195694,
-0.009554915130138397,
0.00754837691783905,
-0.007791043724864721,
-0.004838017746806145,
0.06496596336364746,
0.07915222644805908,
0.00873017217963934,
-0.03375370800495148,
-0.03947117552161217,
-0.06486465036869049,
-0.00129956123419106,
-0.016293849796056747,
-0.03078419528901577,
2.7435274076864156e-33,
0.05805821344256401,
0.0544290691614151,
0.02219589240849018,
0.029874471947550774,
0.07446000725030899,
0.02529118023812771,
-0.03154178336262703,
0.04670483618974686,
-0.06352287530899048,
-0.011756183579564095,
-0.009900011122226715,
-0.04215546324849129,
-0.051848869770765305,
0.12909632921218872,
-0.017474716529250145,
-0.0579802431166172,
-0.014003021642565727,
0.004705457482486963,
-0.013312791474163532,
-0.004351580049842596,
-0.0012580007314682007,
-0.004210212733596563,
-0.027798375114798546,
0.013790265657007694,
-0.010380231775343418,
0.04945533350110054,
0.11034148186445236,
-0.11244947463274002,
-0.03574589267373085,
0.03059711493551731,
-0.1567196249961853,
-0.028061851859092712,
0.007128576282411814,
0.029392201453447342,
0.05523725226521492,
-0.030471671372652054,
-0.00975054781883955,
-0.07162678241729736,
0.040651895105838776,
-0.0819949060678482,
-0.051557037979364395,
0.0198846273124218,
-0.053420595824718475,
-0.12035982310771942,
-0.11706288158893585,
-0.018288111314177513,
0.041229840368032455,
0.0983792394399643,
-0.08212623000144958,
-0.004622089676558971,
0.05869927257299423,
0.040155891329050064,
-0.07928074896335602,
-0.044564757496118546,
-0.06446534395217896,
-0.019288744777441025,
0.0117338290438056,
-0.007302204612642527,
0.021983254700899124,
0.020782509818673134,
-0.05336303636431694,
0.04535210505127907,
-0.007456823252141476,
0.12281103432178497,
0.028286561369895935,
-0.010499018244445324,
-0.01105682086199522,
0.008980494923889637,
0.03284395486116409,
0.015900708734989166,
-0.06072499603033066,
0.0322861447930336,
0.030455656349658966,
-0.07398897409439087,
0.0002685489598661661,
-0.005138896871358156,
0.10655669867992401,
-0.02557617798447609,
0.006172499153763056,
0.09648839384317398,
-0.035940952599048615,
0.023158902302384377,
0.041461072862148285,
-0.07186802476644516,
0.008213544264435768,
-0.0326358936727047,
0.04003535583615303,
-0.0646517425775528,
-0.05307229608297348,
-0.09343216568231583,
-0.05756275728344917,
-0.04657311737537384,
0.008219437673687935,
-0.05796770378947258,
-0.009388773702085018,
-3.179376468400334e-33,
0.03037494607269764,
0.0483001209795475,
0.01163594238460064,
0.04190768674015999,
-0.017232459038496017,
-0.09877488762140274,
0.038965675979852676,
0.020438961684703827,
-0.04867438226938248,
-0.045399729162454605,
-0.032803699374198914,
-0.026549678295850754,
-0.04399755969643593,
-0.02616203762590885,
0.03330482169985771,
-0.02181425876915455,
0.0026631781365722418,
-0.06457819789648056,
-0.003840640652924776,
0.0036624944768846035,
0.003144684946164489,
0.05047522485256195,
-0.05404890328645706,
0.05763595551252365,
-0.0780755802989006,
0.07398784905672073,
-0.05547937750816345,
0.011809558607637882,
-0.02680486999452114,
-0.03259728476405144,
0.012154700234532356,
-0.0057600270956754684,
-0.021269844844937325,
-0.04494225233793259,
-0.07242587953805923,
0.02488456852734089,
0.02626907266676426,
0.046383075416088104,
-0.02020977810025215,
0.127499520778656,
0.10025670379400253,
-0.05775420740246773,
-0.02179049327969551,
0.010073215700685978,
-0.007766903378069401,
-0.06061595305800438,
-0.06355533748865128,
-0.06111496314406395,
0.05636158958077431,
0.014342041686177254,
-0.02445869892835617,
0.008977845311164856,
-0.11591539531946182,
0.017627211287617683,
-0.04662386327981949,
-0.11923123151063919,
-0.08862355351448059,
0.031283240765333176,
0.03671792522072792,
-0.021652471274137497,
0.016949182376265526,
-0.08886579424142838,
-0.059331219643354416,
-0.01575542986392975,
0.0069846101105213165,
-0.04126103222370148,
-0.026605114340782166,
-0.06291662901639938,
-0.009403001517057419,
0.006400029640644789,
0.039263851940631866,
-0.030721250921487808,
-0.025486329570412636,
0.0739300474524498,
-0.008028693497180939,
0.01552603580057621,
0.020761096850037575,
-0.014709669165313244,
0.05314970761537552,
-0.029430758208036423,
0.018335504457354546,
-0.033509183675050735,
0.045168038457632065,
0.07859162241220474,
0.04886786639690399,
0.06632430851459503,
0.02150188945233822,
0.045456595718860626,
0.0491531677544117,
0.05150682479143143,
-0.0852130651473999,
0.08366082608699799,
0.046644169837236404,
0.06693733483552933,
0.02509891241788864,
-4.0263710587851165e-8,
-0.04471689835190773,
-0.0068974681198596954,
-0.07148414105176926,
-0.013262542895972729,
-0.04514606297016144,
-0.054864346981048584,
-0.03154419735074043,
0.06884731352329254,
0.06033995747566223,
-0.01709878444671631,
0.04372655972838402,
-0.03604923561215401,
-0.07775067538022995,
-0.07793090492486954,
0.020160699263215065,
0.039484020322561264,
0.016804231330752373,
-0.012194260954856873,
-0.006613490637391806,
0.013095108792185783,
0.007587000261992216,
0.05117735639214516,
0.008787770755589008,
-0.03198622167110443,
0.006558964028954506,
-0.03227871283888817,
0.0038302522152662277,
0.12584388256072998,
0.06350013613700867,
-0.0046254429034888744,
-0.04997576028108597,
0.08468904346227646,
-0.02667258307337761,
-0.060775239020586014,
-0.06651730090379715,
0.0012163089122623205,
0.052267178893089294,
-0.007347913924604654,
0.0028446619398891926,
-0.04359830543398857,
0.04745468497276306,
0.02078402228653431,
-0.017744772136211395,
0.03501619026064873,
-0.03029962256550789,
0.023470982909202576,
0.0428941547870636,
-0.10766518115997314,
0.000025750687200343236,
-0.0052435388788580894,
-0.030258990824222565,
0.00872426200658083,
-0.027687106281518936,
0.1341727077960968,
0.013296568766236305,
-0.014715940691530704,
0.054329682141542435,
0.028526799753308296,
0.08508045971393585,
0.04395625740289688,
0.014315973967313766,
0.01784597337245941,
-0.00762637984007597,
-0.05028143525123596
] | 0 | 0 | 0 | 0 |
|
URLOST: Unsupervised Representation Learning without Stationarity or Topology | https://openreview.net/forum?id=MBBRHDuiwM | [
"Zeyu Yun",
"Juexiao Zhang",
"Yann LeCun",
"Yubei Chen"
] | Poster | Unsupervised representation learning has seen tremendous progress. However, it is constrained by its reliance on domain specific stationarity and topology, a limitation not found in biological intelligence systems. For instance, unlike computer vision, human vision can process visual signals sampled from highly irregular and non-stationary sensors. We introduce a novel framework that learns from high-dimensional data without prior knowledge of stationarity and topology. Our model, abbreviated as URLOST, combines a learnable self-organizing layer, spectral clustering, and a masked autoencoder (MAE). We evaluate its effectiveness on three diverse data modalities including simulated biological vision data, neural recordings from the primary visual cortex, and gene expressions. Compared to state-of-the-art unsupervised learning methods like SimCLR and MAE, our model excels at learning meaningful representations across diverse modalities without knowing their stationarity or topology. It also outperforms other methods that are not dependent on these factors, setting a new benchmark in the field. We position this work as a step toward unsupervised learning methods capable of generalizing across diverse high-dimensional data modalities. | Unsupervised learning, Self-Supervised Learning, NeuroAI, Multi-Modality, Human Vision, Biologically-inspired Models | null | 13,751 | 2310.04496 | [
-0.019166886806488037,
-0.18490727245807648,
0.0076830568723380566,
0.0662587434053421,
0.03867965191602707,
0.04836741462349892,
-0.024773892015218735,
-0.05852290242910385,
0.024766329675912857,
-0.05208044499158859,
-0.03690110892057419,
-0.08159451931715012,
-0.023689690977334976,
0.08143774420022964,
-0.06042571738362312,
-0.036139652132987976,
0.020226934924721718,
0.020357996225357056,
-0.019146963953971863,
-0.057679396122694016,
0.06100143864750862,
-0.009747074916958809,
0.04322326183319092,
0.08063795417547226,
0.10394099354743958,
0.03588619455695152,
0.02196550741791725,
-0.006477443967014551,
0.02063796855509281,
-0.05534069612622261,
0.04506997391581535,
-0.011910226196050644,
-0.020946882665157318,
0.0027454823721200228,
0.011740654706954956,
0.02627323381602764,
-0.03331359848380089,
0.08680669218301773,
-0.009237698279321194,
-0.03789598494768143,
0.03983268514275551,
0.005653346888720989,
0.018585138022899628,
0.03051133267581463,
0.045407600700855255,
0.021821612492203712,
0.008899529464542866,
-0.09493862837553024,
-0.01683015376329422,
-0.06016584113240242,
-0.047780442982912064,
0.004970982205122709,
-0.04977000504732132,
0.06744998693466187,
0.050565991550683975,
0.051136236637830734,
0.025231847539544106,
-0.03085073083639145,
0.015502948313951492,
-0.03496026247739792,
0.07563309371471405,
-0.04562261700630188,
0.03305434063076973,
-0.041897207498550415,
-0.0344230942428112,
0.026787271723151207,
0.0018239488126710057,
0.06898172944784164,
0.037079986184835434,
-0.04062625393271446,
0.03301043063402176,
0.05810730159282684,
-0.07114527374505997,
0.016742749139666557,
0.09941445291042328,
0.05851856246590614,
-0.012453592382371426,
-0.03601546585559845,
0.11219971626996994,
-0.04066484421491623,
0.04903730750083923,
0.052973922342061996,
0.006757538765668869,
0.05831257253885269,
0.07419013977050781,
-0.005687362048774958,
0.001766991219483316,
0.08300742506980896,
0.00356096844188869,
-0.03751537948846817,
-0.046931903809309006,
-0.009955824352800846,
-0.04551100730895996,
-0.08472094684839249,
-0.011397548951208591,
-0.042068302631378174,
0.02800811268389225,
-0.05274193361401558,
0.011743050068616867,
0.05976550281047821,
-0.030087970197200775,
0.009355786256492138,
-0.02346733771264553,
-0.003897450165823102,
0.05443374440073967,
-0.016271516680717468,
0.06737484037876129,
0.09211453050374985,
0.07969457656145096,
-0.06187346205115318,
0.042759887874126434,
0.018431879580020905,
-0.0752314031124115,
0.07125590741634369,
-0.04205973073840141,
0.050129275768995285,
0.061164118349552155,
0.016057392582297325,
0.09503072500228882,
0.09269629418849945,
-0.027658604085445404,
-0.01484241895377636,
-0.007612796500325203,
0.03521093726158142,
0.04850497841835022,
-0.052175089716911316,
-0.06914294511079788,
1.84581706564282e-33,
-0.026099957525730133,
0.01129672210663557,
0.029922323301434517,
-0.03542325645685196,
0.04807499423623085,
-0.04191228002309799,
-0.03738236799836159,
0.0015622273785993457,
0.061779700219631195,
0.06390559673309326,
-0.04546914994716644,
0.04318595305085182,
-0.0303350780159235,
0.14723612368106842,
0.04226703569293022,
-0.01695798523724079,
-0.06171535328030586,
0.01825554482638836,
-0.022990819066762924,
-0.13423889875411987,
0.0022660703398287296,
0.0534258708357811,
-0.07508090883493423,
-0.06903901696205139,
-0.044813062995672226,
-0.0066580260172486305,
0.01718590036034584,
-0.05991439148783684,
0.00887293927371502,
0.031490519642829895,
-0.07343512028455734,
0.055638547986745834,
0.013287610374391079,
0.04339621216058731,
0.11035440862178802,
-0.003949963953346014,
0.050137218087911606,
-0.05350036174058914,
-0.014559678733348846,
0.0007451728451997042,
0.007263259496539831,
0.07755620032548904,
-0.0016949160490185022,
-0.07041113078594208,
-0.003586459206417203,
-0.0451679565012455,
0.04955067113041878,
0.022803349420428276,
-0.08719509094953537,
-0.03791511803865433,
0.038579750806093216,
-0.021634474396705627,
-0.05330776795744896,
-0.06916134804487228,
-0.0453840009868145,
0.04670511931180954,
0.008257914334535599,
-0.010880611836910248,
0.12997649610042572,
0.029425527900457382,
-0.015537193976342678,
-0.03440557420253754,
0.03363829106092453,
0.037823017686605453,
0.02209234982728958,
-0.043030086904764175,
0.010385946370661259,
0.01205505896359682,
-0.0033995818812400103,
0.037084903568029404,
-0.018551334738731384,
0.07879400998353958,
-0.09301798790693283,
-0.08474084734916687,
0.04137498885393143,
0.01752644218504429,
0.039240479469299316,
-0.0007104846881702542,
-0.14986178278923035,
0.0192050039768219,
0.034210171550512314,
-0.0491635724902153,
-0.04404662922024727,
-0.006026941351592541,
-0.061304427683353424,
-0.006903296802192926,
0.08811485767364502,
-0.012709167785942554,
-0.07226353138685226,
0.007093279156833887,
-0.0008584872703067958,
-0.022899089381098747,
0.05128256976604462,
-0.03282508626580238,
-0.04021020233631134,
-1.2248691543186545e-33,
-0.01688329689204693,
0.05863003432750702,
-0.007773331366479397,
-0.004147028550505638,
-0.028725242242217064,
-0.058243025094270706,
0.04449452459812164,
0.039448242634534836,
-0.016831615939736366,
-0.04726994037628174,
0.00047099494258873165,
-0.020730696618556976,
0.04217570647597313,
0.02188313938677311,
-0.0078007872216403484,
0.008660780265927315,
-0.009963151067495346,
0.05892196297645569,
-0.011205258779227734,
0.024310791864991188,
-0.12333264201879501,
0.022131413221359253,
-0.06435234099626541,
0.016705237329006195,
-0.02673090621829033,
0.10962963104248047,
-0.07681910693645477,
0.09042970836162567,
-0.009034310467541218,
0.08629389852285385,
-0.062120694667100906,
0.020581811666488647,
-0.05337340012192726,
-0.0557895302772522,
-0.07284510135650635,
0.134846031665802,
-0.019532231613993645,
-0.0072130365297198296,
-0.04353122040629387,
0.015171263366937637,
0.01951891928911209,
-0.03228886052966118,
-0.0839194729924202,
0.01226405706256628,
0.020937960594892502,
-0.08959130942821503,
-0.013013116084039211,
0.031051304191350937,
-0.01837035082280636,
0.07647336274385452,
-0.03514128550887108,
0.01316019520163536,
-0.06329645216464996,
0.0013755526160821319,
-0.03506316989660263,
0.029659125953912735,
-0.004086563363671303,
-0.007206568494439125,
0.015885604545474052,
0.016426142305135727,
0.01269844826310873,
-0.06319560110569,
0.010592701844871044,
-0.02708406187593937,
-0.04668277129530907,
-0.059851378202438354,
-0.023830866441130638,
-0.012085717171430588,
-0.015624403953552246,
0.045118048787117004,
-0.023000668734312057,
-0.01042049378156662,
0.02930314652621746,
-0.003715794300660491,
0.012121986597776413,
-0.009812360629439354,
0.039788100868463516,
0.028418980538845062,
-0.03740272670984268,
-0.03100779838860035,
0.009415902197360992,
-0.06139638274908066,
-0.030792567878961563,
0.02430890128016472,
0.1096421480178833,
0.12065214663743973,
0.02783897891640663,
0.014425893314182758,
-0.0027214090805500746,
-0.035977646708488464,
-0.010927584022283554,
0.0013405701611191034,
-0.059373416006565094,
0.024474358186125755,
-0.0425744503736496,
-4.2751825191089665e-8,
-0.1208166852593422,
0.0009104158380068839,
0.03797589987516403,
-0.06911743432283401,
0.028957989066839218,
-0.05896474048495293,
0.0025497237220406532,
0.0794452354311943,
-0.03267340734601021,
0.012283381074666977,
0.09183105826377869,
-0.10568699240684509,
-0.0775720402598381,
-0.008517561480402946,
0.028020374476909637,
0.06211911886930466,
0.023966405540704727,
0.0829688012599945,
-0.021192407235503197,
0.03586506471037865,
-0.07627497613430023,
-0.003436360042542219,
-0.0127498097717762,
-0.06500393897294998,
0.07083427160978317,
-0.06138591468334198,
-0.029233569279313087,
0.0836278423666954,
0.046388376504182816,
0.014286531135439873,
0.06792061775922775,
0.0470096655189991,
0.07349002361297607,
-0.02446424588561058,
-0.04797319322824478,
0.09648623317480087,
0.026028135791420937,
0.050102077424526215,
-0.08958739787340164,
-0.06483663618564606,
-0.023319195955991745,
0.0808812603354454,
-0.04680030047893524,
0.021837886422872543,
0.04175550863146782,
0.007558916695415974,
0.1135949045419693,
-0.02207481674849987,
0.012394418008625507,
0.020643429830670357,
0.007210317999124527,
-0.04611974582076073,
0.016104785725474358,
0.08082973212003708,
0.06408518552780151,
0.024365853518247604,
-0.032846592366695404,
-0.0928962379693985,
0.0147563386708498,
0.05286075174808502,
-0.0020452532917261124,
0.014199988916516304,
-0.09589125216007233,
0.006098589859902859
] | 0 | 0 | 0 | 0 |
|
One-for-All Few-Shot Anomaly Detection via Instance-Induced Prompt Learning | https://openreview.net/forum?id=Zzs3JwknAY | [
"Wenxi Lv",
"Qinliang Su",
"Wenchao Xu"
] | Poster | Anomaly detection methods under the 'one-for-all' paradigm aim to develop a unified model capable of detecting anomalies across multiple classes. However, these approaches typically require a large number of normal samples for model training, which may not always be feasible in practice. Few-shot anomaly detection methods can address scenarios with limited data but often require a tailored model for each class, struggling within the 'one-for-one' paradigm. In this paper, we first proposed the one-for-all few-shot anomaly detection method with the assistance of vision-language model. Different from previous CLIP-based methods learning fix prompts for each class, our method learn a class-shared prompt generator to adaptively generate suitable prompt for each instance. The prompt generator is trained by aligning the prompts with the visual space and utilizing guidance from general textual descriptions of normality and abnormality. Furthermore, we address the mismatch problem of the memory bank within one-for-all paradigm. Extensive experimental results on MVTec and VisA demonstrate the superiority of our method in few-shot anomaly detection task under the one-for-all paradigm. | Anomaly detection, few-shot, vision-language model | null | 13,750 | null | [
-0.012677863240242004,
-0.061983995139598846,
0.02052246406674385,
0.00110633356962353,
0.0992264375090599,
-0.005684396717697382,
0.06854160875082016,
-0.018709246069192886,
0.021579109132289886,
-0.03504207357764244,
0.03370637819170952,
-0.07355929166078568,
0.07951544970273972,
0.026984933763742447,
-0.03421443700790405,
-0.044592130929231644,
0.022168613970279694,
0.023901958018541336,
-0.06942374259233475,
-0.0080936960875988,
-0.049723897129297256,
0.021116433665156364,
0.01805928349494934,
0.04823174327611923,
-0.040876004844903946,
-0.009292949922382832,
0.012564966455101967,
-0.05973827838897705,
0.013143258169293404,
-0.05229896306991577,
0.022312013432383537,
-0.00646075326949358,
-0.02111840434372425,
0.08426899462938309,
0.007760453503578901,
-0.0422944538295269,
0.005456928629428148,
0.07369682937860489,
-0.020039720460772514,
-0.03570857271552086,
-0.027265893295407295,
-0.020675528794527054,
0.029138576239347458,
0.002723917830735445,
-0.03498402610421181,
-0.05993863195180893,
-0.09068050235509872,
-0.06691772490739822,
0.07878778129816055,
-0.06973899155855179,
-0.029849234968423843,
-0.10223790258169174,
-0.017556054517626762,
0.010075309313833714,
-0.04293329641222954,
-0.04485195130109787,
0.02710704505443573,
0.04200231283903122,
0.029298823326826096,
-0.028992578387260437,
-0.012330606579780579,
-0.06929336488246918,
-0.015644343569874763,
0.016132397577166557,
0.01671394146978855,
0.01569283939898014,
0.030384114012122154,
0.041499629616737366,
0.031362272799015045,
0.02074357494711876,
-0.04927761107683182,
0.11963747441768646,
-0.03031061589717865,
0.07166214287281036,
-0.0105508454144001,
0.05638348311185837,
-0.015443531796336174,
-0.03252958133816719,
0.03048158437013626,
-0.010105693712830544,
0.02531115710735321,
0.025412561371922493,
0.08174653351306915,
-0.025240851566195488,
0.18811769783496857,
0.05126633495092392,
-0.0143338767811656,
0.03611027076840401,
0.035031095147132874,
0.036856863647699356,
-0.1156420037150383,
-0.03921092301607132,
0.06947331875562668,
0.02066200040280819,
0.04255211353302002,
0.05133352056145668,
-0.04374178498983383,
-0.06678368151187897,
0.04263903573155403,
0.04157457500696182,
-0.026487264782190323,
0.00379150896333158,
0.059561602771282196,
-0.006071362644433975,
0.04383152723312378,
0.042806219309568405,
0.042672496289014816,
-0.039548154920339584,
0.08940541744232178,
-0.12314524501562119,
0.019555745646357536,
0.0270556453615427,
0.013717424124479294,
-0.05544547736644745,
0.010280543938279152,
0.004740628879517317,
0.015877699479460716,
-0.048955049365758896,
-0.05794228985905647,
0.03844494745135307,
0.017241917550563812,
-0.006439976394176483,
-0.0331704281270504,
0.007958580739796162,
0.03130514547228813,
-0.02065214142203331,
-0.04866165295243263,
8.70334456145387e-34,
0.0899881049990654,
0.057720113545656204,
-0.027485569939017296,
0.02340618148446083,
0.07003790885210037,
-0.007300949655473232,
-0.05826806277036667,
0.11146312206983566,
-0.03487250208854675,
-0.009561843238770962,
0.04409689083695412,
-0.05284350737929344,
-0.00031590042635798454,
0.004474729299545288,
0.038396984338760376,
0.017167486250400543,
0.007539760787039995,
0.04128732904791832,
-0.052815794944763184,
0.06186110898852348,
-0.025169242173433304,
-0.0496518574655056,
-0.005476366728544235,
-0.03277616202831268,
0.03557529300451279,
0.06791326403617859,
0.07704326510429382,
0.023847002536058426,
0.0516161248087883,
0.04117671772837639,
-0.04971623420715332,
0.022616935893893242,
-0.03997235372662544,
0.009407484903931618,
0.020787814632058144,
-0.05502869188785553,
0.015550242736935616,
0.01149070169776678,
0.043773021548986435,
-0.06118183210492134,
-0.012279866263270378,
-0.00027664375375024974,
-0.009779632091522217,
-0.059652723371982574,
-0.026858162134885788,
-0.049522243440151215,
0.004686209373176098,
0.04780273139476776,
-0.1272009313106537,
-0.03722406178712845,
-0.0009470190270803869,
-0.003232992487028241,
-0.043818339705467224,
-0.06695279479026794,
-0.07744826376438141,
-0.023715713992714882,
-0.010684971697628498,
-0.001944183255545795,
0.010739712044596672,
-0.005577516742050648,
0.053989000618457794,
0.054083451628685,
-0.03990524262189865,
0.07745583355426788,
-0.016863515600562096,
-0.01064788457006216,
0.04749103635549545,
0.02575353905558586,
0.037070389837026596,
0.07103057950735092,
-0.018259450793266296,
-0.02179792895913124,
-0.10438257455825806,
-0.06149851903319359,
0.0340336337685585,
0.01095748133957386,
0.07781131565570831,
-0.05133551359176636,
0.0000032654809274390573,
0.043088242411613464,
-0.015494247898459435,
-0.10626892000436783,
0.05980852618813515,
-0.03252139687538147,
-0.10579497367143631,
0.008056772872805595,
0.06620890647172928,
-0.004130870569497347,
-0.025973087176680565,
0.015120445750653744,
-0.007075005676597357,
-0.013524153269827366,
-0.07392950356006622,
-0.019677186384797096,
-0.01486737746745348,
-1.2234413878598862e-33,
0.039905279874801636,
0.03661230579018593,
-0.11468958854675293,
-0.008886554278433323,
0.001142301713116467,
-0.03247038275003433,
0.0359027199447155,
0.027554631233215332,
-0.04716292396187782,
-0.07339046150445938,
-0.06182607635855675,
-0.005606014281511307,
-0.025687575340270996,
-0.021650811657309532,
-0.015927797183394432,
-0.01798276975750923,
-0.05393996089696884,
0.03867029398679733,
-0.04703411087393761,
0.07346765697002411,
0.011172614991664886,
0.05491291359066963,
-0.03838306665420532,
-0.034776948392391205,
-0.03024659864604473,
0.06036682799458504,
-0.07479550689458847,
0.08742287755012512,
-0.008969319984316826,
0.01278630830347538,
-0.049097251147031784,
0.07444099336862564,
0.036649949848651886,
-0.013653920963406563,
-0.09722360968589783,
0.07854653894901276,
0.0582745224237442,
-0.02028702013194561,
0.033128201961517334,
0.08081626892089844,
0.06033255532383919,
-0.04061700031161308,
-0.026793377473950386,
0.043995726853609085,
-0.044143542647361755,
-0.0223750751465559,
0.02854328788816929,
0.05215708166360855,
-0.02833840250968933,
0.03362042084336281,
-0.12524160742759705,
0.013571776449680328,
-0.03743003308773041,
0.05903368815779686,
-0.09826204925775528,
-0.018742870539426804,
-0.005890138912945986,
-0.012365909293293953,
-0.049966126680374146,
0.05846294388175011,
-0.029756678268313408,
0.0084285419434309,
-0.01355842687189579,
-0.03065786510705948,
0.01884075626730919,
-0.02857564017176628,
-0.035255249589681625,
0.011135406792163849,
-0.04437265545129776,
-0.03487010300159454,
0.07711382955312729,
0.03425667807459831,
-0.056390177458524704,
-0.044721975922584534,
-0.00820411927998066,
0.030079854652285576,
0.004151950590312481,
-0.09880837798118591,
0.0007475269376300275,
0.014628085307776928,
0.027967868372797966,
-0.04378378391265869,
0.04341236501932144,
0.09798166900873184,
0.06029420346021652,
0.11517494171857834,
0.0843561515212059,
0.04867953062057495,
0.0023460451047867537,
-0.013431794941425323,
-0.09912178665399551,
0.04230254516005516,
0.03452298045158386,
0.03302653506398201,
0.01021155621856451,
-4.3228528312511116e-8,
-0.10534631460905075,
0.0178889948874712,
-0.0008222666219808161,
0.015401464886963367,
0.05352761596441269,
0.0012616306776180863,
-0.16165561974048615,
0.06245950236916542,
0.04856996238231659,
-0.06521037966012955,
-0.027647415176033974,
-0.024892225861549377,
-0.010466327890753746,
-0.07836251705884933,
-0.01025047805160284,
-0.000528045289684087,
-0.055980704724788666,
0.06595367193222046,
-0.06284761428833008,
0.0769038125872612,
0.043914590030908585,
-0.0165996216237545,
0.0652129128575325,
-0.0614473931491375,
-0.02195129171013832,
-0.07436729222536087,
0.006995883770287037,
0.06592321395874023,
0.11384382098913193,
-0.08148161321878433,
0.0022312444634735584,
-0.006938735023140907,
0.07335107773542404,
-0.0010460130870342255,
0.0497715026140213,
0.10659033060073853,
-0.01036887988448143,
-0.019972968846559525,
0.01898394338786602,
-0.03556694835424423,
0.04895978793501854,
-0.018182944506406784,
-0.008983450941741467,
0.04384541139006615,
-0.07158701866865158,
0.006134470924735069,
0.028525633737444878,
-0.17707037925720215,
0.04589194059371948,
-0.03343190625309944,
-0.012509950436651707,
-0.045609697699546814,
-0.0032426558900624514,
0.16002386808395386,
0.010825793258845806,
0.00694234436377883,
0.045842401683330536,
-0.07282066345214844,
0.05449461564421654,
0.09127229452133179,
0.06797497719526291,
0.057506389915943146,
-0.05882306024432182,
0.00899199303239584
] | 0 | 0 | 0 | 0 |
|
K-HALU: Multiple Answer Korean Hallucination Benchmark for Large Language Models | https://openreview.net/forum?id=VnLhUogHYE | [
"Jaehyung Seo",
"Heuiseok Lim"
] | Poster | Recent researchers and companies have been developing large language models (LLMs) specifically designed for particular purposes and have achieved significant advancements in various natural language processing tasks. However, LLMs are still prone to generating hallucinations—results that are unfaithful or inconsistent with the given input. As a result, the need for datasets to evaluate and demonstrate the hallucination detection capabilities of LLMs is increasingly recognized. Nonetheless, the Korean NLP community lacks publicly available benchmark datasets demonstrating the faithfulness of knowledge-based information. Furthermore, the few existing datasets that evaluate hallucination are limited in their access to the entire dataset, restricting detailed analysis beyond simple scoring, and are based on translated English knowledge. To address these challenges, we introduce K-HALU, a Korean benchmark designed to evaluate LLMs' hallucination detection in Korean. This benchmark contains seven domains, considering the faithfulness of statements based on knowledge documents compiled from Korean news, magazines, and books. For more strict evaluation, 40% of the dataset is structured as multiple-answer questions, requiring models to select all possible correct answers from the given options. Our empirical results show that open-source LLMs still struggle with hallucination detection in Korean knowledge, emphasizing the need for a more detailed analysis of their limitations. | Hallucination, Benchmark dataset, Multiple answer, Korean, Large language model | Multiple-answer Korean hallucination benchmark for large language models | 13,748 | null | [
-0.002492539118975401,
0.006633593235164881,
0.05863382667303085,
0.026024851948022842,
0.036106839776039124,
0.01145009882748127,
-0.004602481611073017,
-0.04783112555742264,
0.014590028673410416,
-0.008887981064617634,
-0.030420448631048203,
-0.07432287931442261,
0.05587184429168701,
0.006864598952233791,
0.025713644921779633,
0.02027255669236183,
0.025649098679423332,
0.05243905261158943,
-0.09101098775863647,
-0.10692759603261948,
0.060705699026584625,
0.020834269002079964,
0.07898043841123581,
-0.017353732138872147,
0.07791808992624283,
-0.01238919235765934,
-0.00044210831401869655,
-0.015416420064866543,
-0.02629898302257061,
-0.05432455986738205,
-0.014135058037936687,
0.08384773880243301,
-0.002534556668251753,
0.09808214753866196,
0.008109626360237598,
0.020297177135944366,
-0.058531757444143295,
0.026759125292301178,
-0.0010099492501467466,
-0.04063126817345619,
-0.060169924050569534,
-0.012910610064864159,
0.08253023028373718,
0.005999840330332518,
0.10704144090414047,
0.005321518983691931,
-0.10288812965154648,
0.002539793960750103,
-0.05385732278227806,
0.030231870710849762,
-0.09951899945735931,
0.012885292991995811,
0.07833652198314667,
0.07462641596794128,
-0.05876742675900459,
-0.006652748212218285,
-0.05949664115905762,
0.051553767174482346,
0.03025668114423752,
-0.04982379451394081,
-0.08817525953054428,
-0.11648167669773102,
-0.09947879612445831,
-0.0004615762154571712,
-0.025530286133289337,
-0.03340480104088783,
-0.01073813159018755,
0.03221583738923073,
-0.041973765939474106,
0.04305243119597435,
0.016149085015058517,
0.05395307391881943,
0.010762455873191357,
0.121391661465168,
-0.016696393489837646,
0.028631150722503662,
0.022560207173228264,
-0.049170929938554764,
0.04323086142539978,
-0.0939614325761795,
-0.020202886313199997,
0.04604676365852356,
0.06924086809158325,
0.015301750972867012,
0.05389861389994621,
-0.056479427963495255,
-0.010593226179480553,
0.05914616957306862,
-0.07196890562772751,
0.030134497210383415,
-0.053502123802900314,
-0.08398199081420898,
0.05257389321923256,
-0.05583806708455086,
0.05921480432152748,
0.07614880800247192,
-0.04011744260787964,
-0.04784010723233223,
-0.013056720606982708,
0.035371918231248856,
0.01252609584480524,
0.13334332406520844,
0.01996723562479019,
-0.1164262667298317,
-0.02778041549026966,
-0.06261130422353745,
0.054656896740198135,
0.04412474483251572,
0.030063029378652573,
-0.047339554876089096,
-0.015986669808626175,
0.07974828034639359,
0.012690290808677673,
-0.025522548705339432,
0.10130847245454788,
-0.07371848076581955,
0.05020758509635925,
-0.020131973549723625,
0.01221020519733429,
0.06251225620508194,
-0.02366453967988491,
0.06451496481895447,
0.06298553198575974,
-0.03087221458554268,
-0.020789187401533127,
-0.01906638965010643,
0.012327708303928375,
2.2830381486693388e-33,
0.08859500288963318,
0.03647778928279877,
0.024274280294775963,
-0.01592688448727131,
-0.009604071266949177,
-0.059556227177381516,
-0.057401057332754135,
0.04076385125517845,
-0.03533007949590683,
-0.01775861345231533,
-0.06974785029888153,
-0.005598476622253656,
-0.07022219151258469,
-0.03247753158211708,
0.027393797412514687,
0.05850929394364357,
-0.0577581450343132,
0.05257677659392357,
-0.08522767573595047,
0.053892333060503006,
0.09598635137081146,
0.026246441528201103,
0.0692417249083519,
-0.03214633837342262,
0.0008539539994671941,
-0.016957448795437813,
0.040489647537469864,
-0.09415531158447266,
-0.06040080636739731,
0.0431668758392334,
-0.04978223890066147,
-0.0468825101852417,
-0.011918216943740845,
0.038065940141677856,
0.01582247018814087,
-0.025079837068915367,
-0.05097880959510803,
-0.05021129921078682,
0.009246163070201874,
-0.0358491986989975,
-0.0521838404238224,
0.0020662909373641014,
-0.0541238896548748,
0.09131983667612076,
-0.07716001570224762,
0.034179169684648514,
-0.0381268709897995,
-0.07558201998472214,
0.004030920099467039,
-0.006475464440882206,
0.05864701047539711,
0.024206403642892838,
-0.050012312829494476,
-0.007170029915869236,
0.05415433645248413,
0.055751681327819824,
0.0437178760766983,
-0.06653781980276108,
0.028839653357863426,
0.024509362876415253,
0.050388094037771225,
0.09170084446668625,
-0.01548941619694233,
0.07351574301719666,
0.0016564968973398209,
-0.026955798268318176,
-0.07265224307775497,
0.02108539640903473,
0.000011233369150431827,
-0.05015581473708153,
-0.01898253709077835,
-0.05637010559439659,
0.0017883830005303025,
-0.0033860947005450726,
-0.006862003821879625,
0.02664472721517086,
0.045425496995449066,
-0.044749680906534195,
0.041693273931741714,
-0.002434422727674246,
0.05849699303507805,
-0.05502257123589516,
-0.018504787236452103,
-0.07717760652303696,
-0.03348192200064659,
-0.05804945528507233,
0.04463563859462738,
-0.08152125775814056,
-0.04324779286980629,
-0.0203926432877779,
-0.010499182157218456,
0.0144401416182518,
-0.0052013099193573,
-0.03515607863664627,
-0.003410893026739359,
-2.1291549178871624e-33,
-0.03224511072039604,
-0.05663064494729042,
0.03087463788688183,
0.08062317222356796,
-0.0666406899690628,
0.01059539895504713,
-0.02052861638367176,
0.07533606141805649,
0.020545650273561478,
-0.042119842022657394,
-0.014835508540272713,
-0.05206377059221268,
0.07337523251771927,
-0.02327173762023449,
0.06251666694879532,
-0.04009346663951874,
0.061241596937179565,
0.07385246455669403,
-0.019198814406991005,
0.16571688652038574,
-0.05954559147357941,
0.10330750793218613,
-0.13222476840019226,
0.04139956086874008,
-0.021098405122756958,
0.044523704797029495,
-0.028979016467928886,
0.025565819814801216,
-0.0023810805287212133,
0.005903662648051977,
0.04906940832734108,
0.02985217049717903,
-0.06334912776947021,
-0.013247894123196602,
-0.07426155358552933,
-0.030294472351670265,
0.005854310933500528,
-0.03118489310145378,
-0.06214640289545059,
0.02694670297205448,
0.05710536241531372,
-0.001314534805715084,
-0.06505873054265976,
-0.004518650006502867,
-0.034284524619579315,
-0.05995221063494682,
-0.10461414605379105,
0.019129928201436996,
-0.0034774939995259047,
-0.03979882970452309,
0.0009624756057746708,
-0.05316673219203949,
-0.05744162201881409,
0.013941106386482716,
0.02698100171983242,
-0.0592188686132431,
-0.10613347589969635,
-0.04694129526615143,
0.03435221314430237,
0.030359139665961266,
-0.08567901700735092,
0.02342463657259941,
0.05219913646578789,
-0.024805646389722824,
0.041972797363996506,
-0.06515037268400192,
0.04648871347308159,
0.04270622134208679,
0.01620238833129406,
-0.020264752209186554,
0.02994958683848381,
-0.01929582841694355,
0.026450365781784058,
0.04628553241491318,
-0.004730814136564732,
0.030567029491066933,
0.013613013550639153,
-0.044496357440948486,
-0.04076078161597252,
-0.07800467312335968,
0.007207594811916351,
0.008223488926887512,
-0.027949634939432144,
0.0942036435008049,
0.0749899297952652,
0.015288973227143288,
0.025873877108097076,
0.06871843338012695,
0.02969776652753353,
0.03569946810603142,
0.017493318766355515,
-0.01415672805160284,
-0.0827893316745758,
0.10339600592851639,
-0.05432048439979553,
-3.9549671981831125e-8,
-0.0403544157743454,
-0.009952886961400509,
-0.06546671688556671,
0.043603915721178055,
-0.006298784166574478,
-0.043835125863552094,
-0.027601974084973335,
0.05860082060098648,
0.011266948655247688,
-0.08828864246606827,
0.0959293469786644,
0.04701579362154007,
-0.03580761328339577,
-0.0634012222290039,
-0.03199921175837517,
0.04003914073109627,
0.023457063362002373,
0.0860016793012619,
0.01111983135342598,
0.015324262902140617,
0.14634594321250916,
0.019664069637656212,
0.04574185982346535,
0.002505648648366332,
0.017726533114910126,
0.03196756914258003,
-0.07071594148874283,
0.05130472406744957,
-0.01920822262763977,
-0.03617941960692406,
0.04371987283229828,
-0.012051299214363098,
-0.022163625806570053,
0.01817142404615879,
0.07711086422204971,
0.05523741990327835,
-0.005880314856767654,
-0.03772352635860443,
0.023580720648169518,
-0.008471093140542507,
0.05518746003508568,
0.01973583549261093,
-0.03980853781104088,
0.06230873614549637,
-0.02201809734106064,
-0.039020102471113205,
0.008672477677464485,
-0.06619933247566223,
0.021808495745062828,
-0.052919670939445496,
-0.02558724954724312,
-0.04838551580905914,
-0.011732425540685654,
0.045797448605298996,
0.09490827471017838,
0.0894276350736618,
0.03432286158204079,
-0.005402096547186375,
0.028129616752266884,
-0.022109558805823326,
0.08969710022211075,
-0.01837218552827835,
-0.04481775686144829,
0.06756296008825302
] | 0 | 0 | 0 | 0 |
|
Charting the Design Space of Neural Graph Representations for Subgraph Matching | https://openreview.net/forum?id=5pd78GmXC6 | [
"Vaibhav Raj",
"Indradyumna Roy",
"Ashwin Ramachandran",
"Soumen Chakrabarti",
"Abir De"
] | Poster | Subgraph matching is vital in knowledge graph (KG) question answering, molecule design, scene graph, code and circuit search, etc.
Neural methods have shown promising results for subgraph matching.
Our study of recent systems suggests refactoring them into a unified design space for graph matching networks.
Existing methods occupy only a few isolated patches in this space, which remains largely uncharted.
We undertake the first comprehensive exploration of this space, featuring such axes as attention-based vs. soft permutation-based interaction between query and corpus graphs, aligning nodes vs. edges, and the form of the final scoring network that integrates neural representations of the graphs.
Our extensive experiments reveal that judicious and hitherto-unexplored combinations of choices in this space lead to large performance benefits.
Beyond better performance, our study uncovers valuable insights and establishes general design principles for neural graph representation and interaction, which may be of wider interest. | Graph Retrieval, Graph Neural Networks, Subgraph Matching | We propose a unified framework for graph matching networks and experiment with various alternatives for each design axis to obtain state-of-the-art results on the subgraph isomorphism task. | 13,746 | null | [
-0.069887675344944,
-0.041696805506944656,
0.022267550230026245,
0.004487580619752407,
0.036477018147706985,
0.026789456605911255,
-0.006514003500342369,
0.035570867359638214,
-0.0015341369435191154,
-0.03011915273964405,
-0.08094403147697449,
-0.03432461991906166,
0.03231647238135338,
0.09826818108558655,
-0.0436360128223896,
0.08923395723104477,
0.07194413244724274,
0.043239448219537735,
0.019374847412109375,
-0.01444562990218401,
0.0038360164035111666,
0.024557536467909813,
0.035817693918943405,
-0.05719322711229324,
0.03570952266454697,
0.04760507494211197,
-0.04791500046849251,
-0.0023014701437205076,
0.007849938236176968,
0.0019821261521428823,
0.11967387050390244,
0.03352852538228035,
-0.006046018097549677,
0.08508144319057465,
-0.04160056263208389,
0.025655999779701233,
-0.06936704367399216,
0.0032428812701255083,
0.04585910588502884,
-0.04924419894814491,
-0.04947289451956749,
0.015538307838141918,
0.02789686806499958,
0.11005216091871262,
0.08561047166585922,
0.027035638689994812,
-0.06091323122382164,
0.023929176852107048,
0.015582498162984848,
-0.016737787052989006,
-0.06925210356712341,
-0.03769787400960922,
-0.03800176829099655,
-0.017745250836014748,
0.04550408571958542,
0.11085192114114761,
-0.032556433230638504,
0.008905565366148949,
0.005916542839258909,
0.006337085738778114,
0.047548577189445496,
-0.05024262145161629,
0.01283629797399044,
0.01149496715515852,
-0.0012452489463612437,
-0.011777508072555065,
0.011006230488419533,
0.012562134303152561,
-0.022340774536132812,
0.01622827537357807,
0.04240330308675766,
0.08223048597574234,
-0.05894741043448448,
0.048541370779275894,
0.040322195738554,
0.13874408602714539,
0.0067449649795889854,
-0.08404194563627243,
-0.03853445127606392,
-0.11298084259033203,
-0.05125724524259567,
-0.017355741932988167,
0.062148597091436386,
0.053458087146282196,
0.13636870682239532,
-0.036557260900735855,
0.01760750450193882,
-0.02327953651547432,
0.0014352195430547,
-0.03049366921186447,
-0.1021028459072113,
-0.05689936876296997,
0.04334604740142822,
-0.09532829374074936,
0.024685753509402275,
0.10853661596775055,
0.08994036912918091,
-0.012236963026225567,
0.007694427389651537,
0.04495967924594879,
-0.03291093185544014,
0.06878245621919632,
0.015552470460534096,
-0.05755352973937988,
-0.02063731476664543,
0.016400761902332306,
-0.017862865701317787,
0.04355461150407791,
0.03685295954346657,
-0.11252885311841965,
-0.039448078721761703,
0.09381764382123947,
0.010813645087182522,
0.05362715572118759,
0.05536291375756264,
-0.06212238594889641,
0.0599023774266243,
-0.0373237207531929,
0.024748854339122772,
0.11027264595031738,
-0.09801596403121948,
0.08333165943622589,
-0.08713606745004654,
-0.03007827326655388,
-0.06513311713933945,
0.02527068369090557,
-0.04446238651871681,
3.093521667094136e-33,
0.07194377481937408,
0.024689000099897385,
0.062069620937108994,
-0.0029893219470977783,
0.06884609162807465,
-0.05154958739876747,
-0.040172550827264786,
0.07307019829750061,
-0.059999365359544754,
0.036895137280225754,
-0.09980108588933945,
0.01622968167066574,
0.0065778582356870174,
0.075571708381176,
-0.040253929793834686,
0.031140530481934547,
-0.023988554254174232,
-0.04386959597468376,
-0.06809768825769424,
-0.11962603777647018,
-0.05833030119538307,
0.006218995898962021,
0.06138158217072487,
0.04730014130473137,
0.023520134389400482,
0.031152671203017235,
-0.005709222052246332,
0.006465406622737646,
0.021385101601481438,
-0.053412795066833496,
-0.12600751221179962,
0.0634133443236351,
0.05298909544944763,
0.007319075055420399,
-0.011644585058093071,
-0.057284221053123474,
0.040959980338811874,
-0.07973074913024902,
0.043511997908353806,
-0.033820293843746185,
-0.06600760668516159,
0.059274882078170776,
-0.02398587204515934,
0.002952281851321459,
-0.04569721221923828,
0.030618540942668915,
-0.03857368230819702,
0.04001535847783089,
0.059762854129076004,
0.016286760568618774,
0.010650851763784885,
0.013982926495373249,
-0.02224232815206051,
-0.0065168053843081,
0.050640832632780075,
-0.07603537291288376,
0.0643085166811943,
0.050193123519420624,
0.046028006821870804,
0.07475581765174866,
-0.01432702038437128,
0.007922991178929806,
0.031222715973854065,
0.028968749567866325,
0.04183892533183098,
0.06849575787782669,
-0.05001860111951828,
0.004329643677920103,
0.09255170077085495,
-0.0026759454049170017,
-0.01792723685503006,
0.07352477312088013,
-0.05001741647720337,
-0.06050310283899307,
-0.01938553713262081,
-0.0014881088864058256,
-0.05431542918086052,
-0.09798740595579147,
-0.04470212757587433,
-0.016860362142324448,
-0.08304347842931747,
-0.05464458838105202,
-0.03336882218718529,
-0.029980890452861786,
-0.04726791754364967,
-0.045791808515787125,
0.07792424410581589,
-0.07247606664896011,
0.05170558765530586,
-0.025188250467181206,
-0.0041954112239181995,
-0.06803632527589798,
-0.023541059345006943,
-0.05423949658870697,
0.06819837540388107,
-1.6671698423086642e-33,
-0.0039223735220730305,
0.02165856584906578,
0.0030289357528090477,
-0.011881436221301556,
0.03270117938518524,
-0.0616290345788002,
-0.0356709323823452,
-0.05825657397508621,
-0.004590159747749567,
0.10539492219686508,
-0.0063077537342906,
-0.08221271634101868,
0.0324193611741066,
-0.08443378657102585,
-0.03708622232079506,
0.024765808135271072,
-0.03506702557206154,
0.03597729280591011,
-0.028790229931473732,
0.031943824142217636,
0.026645679026842117,
0.13027231395244598,
-0.09225865453481674,
0.023036658763885498,
-0.01065546553581953,
0.03833227604627609,
-0.0010848718229681253,
-0.04845729470252991,
-0.027220869436860085,
0.01762908324599266,
-0.02139212377369404,
-0.06465798616409302,
-0.03339395299553871,
-0.047051697969436646,
0.04723474010825157,
0.06645340472459793,
0.019188635051250458,
-0.04514336213469505,
-0.001484560314565897,
-0.04772375896573067,
-0.024864880368113518,
0.017942164093255997,
-0.08419927209615707,
0.04368489980697632,
-0.01463024877011776,
0.025611665099859238,
-0.08749082684516907,
-0.006441330537199974,
-0.06640999019145966,
0.013745587319135666,
-0.01465397048741579,
0.010769490152597427,
0.030242666602134705,
-0.09224254637956619,
-0.04275855794548988,
-0.0890859067440033,
-0.013562338426709175,
0.08063920587301254,
0.0745973065495491,
-0.013365311548113823,
0.011357411742210388,
-0.05687946081161499,
0.042778417468070984,
-0.06246441602706909,
-0.021312622353434563,
-0.0351920910179615,
-0.006040062289685011,
0.054656658321619034,
0.03515341505408287,
0.017955616116523743,
-0.011603718623518944,
-0.0756419450044632,
0.02705693058669567,
0.03274981305003166,
0.04792848601937294,
-0.019431648775935173,
0.07135382294654846,
-0.007493196055293083,
-0.037219297140836716,
0.0066211228258907795,
-0.03551863878965378,
0.003982281778007746,
0.026579635217785835,
0.07790640741586685,
0.07444040477275848,
0.07869692146778107,
-0.0514770969748497,
0.08150607347488403,
0.055309928953647614,
0.01741190254688263,
0.0057897064834833145,
0.024545392021536827,
-0.00959240272641182,
0.06376544386148453,
-0.0005751829012297094,
-4.310977885779721e-8,
-0.13212504982948303,
0.058100923895835876,
0.01321262028068304,
0.05674191936850548,
0.01832982711493969,
-0.03373884782195091,
0.038066521286964417,
0.10925329476594925,
-0.06991889327764511,
0.0659569501876831,
0.13096998631954193,
0.03465539589524269,
0.026248455047607422,
-0.041871871799230576,
-0.0014317752793431282,
-0.005851835012435913,
0.0072150034829974174,
-0.004054817836731672,
0.019738830626010895,
-0.00021054217359051108,
-0.025438016280531883,
0.008043389767408371,
-0.015186951495707035,
0.08607392758131027,
0.021025871858000755,
-0.055798761546611786,
-0.06196797639131546,
0.01023721881210804,
0.008057460188865662,
0.0264949519187212,
-0.00980339478701353,
0.013276893645524979,
0.09926924854516983,
-0.02295519970357418,
0.06575385481119156,
0.0324401780962944,
0.0065774754621088505,
0.0045070103369653225,
-0.08322574943304062,
0.014195208437740803,
0.011142532341182232,
0.017125362530350685,
-0.042353205382823944,
0.027776699513196945,
0.0006723369006067514,
0.01764747127890587,
0.036857396364212036,
0.020205900073051453,
0.03684139624238014,
-0.047145020216703415,
-0.009653475135564804,
-0.1082693487405777,
-0.0378396138548851,
-0.08848603814840317,
0.020613808184862137,
0.018501365557312965,
-0.0388515405356884,
-0.006755239795893431,
0.05292707309126854,
-0.02066263183951378,
0.009093046188354492,
0.0058497716672718525,
-0.023275122046470642,
-0.028153613209724426
] | 0 | 0 | 0 | 0 |
|
Convergence and Implicit Bias of Gradient Descent on Continual Linear Classification | https://openreview.net/forum?id=DTqx3iqjkz | [
"Hyunji Jung",
"Hanseul Cho",
"Chulhee Yun"
] | Poster | We study continual learning on multiple linear classification tasks by sequentially running gradient descent (GD) for a fixed budget of iterations per each given task. When all tasks are jointly linearly separable and are presented in a cyclic/random order, we show the directional convergence of the trained linear classifier to the joint (offline) max-margin solution. This is surprising because GD training on a single task is implicitly biased towards the individual max-margin solution for the task, and the direction of the joint max-margin solution can be largely different from these individual solutions. Additionally, when tasks are given in a cyclic order, we present a non-asymptotic analysis on cycle-averaged forgetting, revealing that (1) alignment between tasks is indeed closely tied to catastrophic forgetting and backward knowledge transfer and (2) the amount of forgetting vanishes to zero as the cycle repeats. Lastly, we analyze the case where the tasks are no longer jointly separable and show that the model trained in a cyclic order converges to the unique minimum of the joint loss function. | Continual Learning, Sequential Learning, Gradient Descent, Linear Classification, Convergence, Implicit Bias | null | 13,745 | null | [
-0.028666147962212563,
-0.11025361716747284,
-0.02059934101998806,
0.049956828355789185,
0.06549162417650223,
0.022868869826197624,
-0.052683521062135696,
-0.07669506222009659,
0.04113851487636566,
-0.02152608148753643,
0.002337900921702385,
0.14105699956417084,
0.04461200162768364,
-0.005317698232829571,
-0.09126950800418854,
-0.018630845472216606,
-0.03980391100049019,
0.05187651887536049,
-0.042705051600933075,
-0.02569987252354622,
-0.01631186343729496,
0.07920614629983902,
0.01949079893529415,
0.09330546855926514,
-0.0031673815101385117,
-0.02163192443549633,
-0.04472661763429642,
-0.0556078664958477,
-0.0022291296627372503,
-0.04983619600534439,
0.009932681918144226,
0.015796665102243423,
-0.024200037121772766,
0.00023421179503202438,
-0.05833963677287102,
0.03610547259449959,
-0.09829622507095337,
0.05604176223278046,
0.024814262986183167,
0.004246698692440987,
-0.007879413664340973,
0.03678029030561447,
-0.009855655953288078,
0.020755495876073837,
0.07096675038337708,
0.005081542767584324,
0.007618704345077276,
-0.1290474385023117,
0.020078200846910477,
0.011205928400158882,
-0.04337520897388458,
-0.008168117143213749,
-0.033320166170597076,
0.07951300591230392,
0.0102470638230443,
-0.02143111452460289,
0.016299689188599586,
0.05425029620528221,
-0.058168623596429825,
0.033034056425094604,
0.020592134445905685,
-0.10058803111314774,
-0.03772906959056854,
-0.03540061041712761,
-0.03489748761057854,
-0.03121153824031353,
0.06441809982061386,
0.037925660610198975,
-0.011113734915852547,
0.0938125029206276,
-0.02836110070347786,
0.07284853607416153,
-0.044333409518003464,
0.001080651069059968,
0.06473302841186523,
0.06162188947200775,
0.037550993263721466,
-0.036743007600307465,
0.02978535369038582,
-0.040479306131601334,
-0.02022886648774147,
0.04113796725869179,
0.07649821788072586,
-0.09327440708875656,
0.08911901712417603,
-0.04125547036528587,
0.032290585339069366,
0.034077584743499756,
0.08687541633844376,
-0.06181851401925087,
-0.05167083069682121,
-0.006378798745572567,
-0.017462680116295815,
-0.06264723092317581,
0.0227519404143095,
0.00590096041560173,
0.0692262277007103,
0.007699944544583559,
-0.02701634354889393,
0.07695908099412918,
-0.009089081548154354,
-0.016787506639957428,
-0.0583333857357502,
0.028618641197681427,
-0.01138706598430872,
-0.06656134873628616,
-0.009352111257612705,
0.02741878665983677,
0.08714944124221802,
-0.10183542221784592,
0.0033775395713746548,
-0.015183274634182453,
-0.014060364104807377,
0.07920031994581223,
-0.00566971255466342,
-0.06437069177627563,
0.027565879747271538,
-0.016240853816270828,
0.0017509673489257693,
0.09251155704259872,
-0.009433929808437824,
0.017649052664637566,
0.02136004902422428,
0.06736224889755249,
-0.03175283968448639,
0.016840996220707893,
0.025814736261963844,
1.7557240561889304e-33,
-0.02177301049232483,
0.016143370419740677,
0.09007734060287476,
-0.05838121846318245,
0.02928984723985195,
-0.024973582476377487,
0.0018416675738990307,
0.055923257023096085,
-0.015568510629236698,
0.0010793646797537804,
0.032159797847270966,
-0.007053795736283064,
-0.030419984832406044,
0.10408540070056915,
0.009763198904693127,
-0.04883026331663132,
-0.03511478006839752,
0.03430977091193199,
0.03011370077729225,
-0.06463293731212616,
0.020273689180612564,
0.0017072021728381515,
-0.003534663002938032,
-0.1020364984869957,
-0.008371623232960701,
-0.03457812964916229,
0.07637152820825577,
0.024698171764612198,
-0.07074172794818878,
0.01734890043735504,
-0.04888378083705902,
0.008422776125371456,
-0.009623190388083458,
0.022725727409124374,
-0.003667246550321579,
-0.017758561298251152,
-0.015112870372831821,
0.03281979262828827,
0.037890903651714325,
-0.019643224775791168,
-0.03341001644730568,
0.021696820855140686,
-0.004236254375427961,
-0.020449651405215263,
-0.07391076534986496,
-0.1438785046339035,
0.07885362952947617,
0.03409529849886894,
-0.030280670151114464,
-0.03402111679315567,
-0.07408182322978973,
-0.004564008209854364,
-0.051035743206739426,
-0.001322546973824501,
-0.04480525106191635,
0.03742878884077072,
-0.06048513203859329,
0.032311197370290756,
-0.05364736169576645,
0.013276452198624611,
0.07381444424390793,
0.014246892184019089,
-0.04668417572975159,
0.05963108688592911,
0.03425062075257301,
0.07091665267944336,
-0.005598515272140503,
0.08326224237680435,
0.03821414336562157,
-0.013513920828700066,
0.03497763350605965,
-0.03904713690280914,
0.026171423494815826,
-0.08043722063302994,
0.11161831021308899,
-0.0025137795601040125,
0.0910082757472992,
-0.13811363279819489,
-0.013338333927094936,
-0.0017211745725944638,
-0.006782620679587126,
-0.044412240386009216,
0.02487982250750065,
-0.06103266775608063,
-0.09037855267524719,
-0.03685889393091202,
0.09704729169607162,
-0.09823867678642273,
-0.047906335443258286,
0.08444107323884964,
-0.06905190646648407,
0.01023897249251604,
0.03471795469522476,
0.014391436241567135,
0.017581507563591003,
-3.099722767136708e-33,
0.002729596570134163,
0.024774599820375443,
0.04189739376306534,
0.008751723915338516,
0.05350685492157936,
0.0006777557427994907,
-0.07035967707633972,
-0.0014384342357516289,
-0.13570131361484528,
-0.04898761212825775,
0.018979348242282867,
-0.01624017022550106,
-0.024398790672421455,
0.0783277228474617,
-0.02189977653324604,
0.017935166135430336,
-0.04135330021381378,
0.032286688685417175,
-0.0072762127965688705,
-0.005387242883443832,
0.02556106448173523,
0.039329033344984055,
-0.020984476432204247,
0.032655008137226105,
0.010519451461732388,
0.03611382469534874,
0.002162767108529806,
0.12466053664684296,
-0.02450813353061676,
-0.019720589742064476,
-0.037826355546712875,
-0.027865298092365265,
-0.008061950094997883,
-0.054131072014570236,
-0.0476510189473629,
0.07336442172527313,
-0.047606270760297775,
0.0032349503599107265,
-0.021250920370221138,
0.16612131893634796,
0.05076972022652626,
-0.052411261945962906,
-0.04609202593564987,
-0.01431583147495985,
0.12756291031837463,
-0.11518415808677673,
-0.08225621283054352,
0.0722157284617424,
0.027579225599765778,
0.06371837109327316,
0.013178455643355846,
0.06062064319849014,
-0.15971867740154266,
0.014193695038557053,
-0.05271970480680466,
-0.01713573932647705,
0.016492068767547607,
0.013084346428513527,
0.015937743708491325,
-0.028545616194605827,
-0.0038980459794402122,
0.03463307023048401,
-0.022216198965907097,
0.015757614746689796,
0.06575672328472137,
-0.032087985426187515,
0.04642034322023392,
0.007010126020759344,
0.023651646450161934,
0.07410093396902084,
0.006032001227140427,
0.0722154900431633,
-0.054948654025793076,
-0.058824680745601654,
0.00882942695170641,
0.005892823915928602,
-0.059314608573913574,
-0.08103471249341965,
-0.045555517077445984,
-0.06379459798336029,
-0.001115109771490097,
0.013647657819092274,
0.02131597325205803,
0.07787837833166122,
-0.004451260436326265,
0.015826324000954628,
0.028253206983208656,
0.07258839905261993,
0.039662495255470276,
-0.05607260391116142,
0.0165072213858366,
-0.03230489045381546,
0.025804225355386734,
-0.018261272460222244,
-0.04049740731716156,
-4.581733392683418e-8,
-0.05272611975669861,
0.01871621422469616,
-0.008057672530412674,
0.005947728641331196,
0.07517626881599426,
-0.0627017617225647,
0.05165065452456474,
0.08805792778730392,
-0.01874426193535328,
-0.024157581850886345,
0.08800562471151352,
0.038723621517419815,
-0.004659335594624281,
-0.01669125072658062,
-0.02792121283710003,
0.04360739886760712,
-0.02581815980374813,
-0.01661113277077675,
-0.0037226174026727676,
-0.007612354122102261,
0.04900052770972252,
-0.047301534563302994,
-0.006030898075550795,
-0.010327504947781563,
-0.015794673934578896,
-0.0987178236246109,
0.03836003690958023,
0.17774556577205658,
0.07549326866865158,
-0.01711067371070385,
-0.010651796124875546,
0.019604604691267014,
-0.00556517019867897,
-0.05584053322672844,
-0.012218773365020752,
0.036723434925079346,
0.04875628277659416,
-0.023923497647047043,
-0.02616472914814949,
-0.014561455696821213,
-0.044559627771377563,
0.0426640585064888,
0.01901129074394703,
0.059372249990701675,
0.039443161338567734,
-0.021169422194361687,
-0.0748034417629242,
-0.07194160670042038,
0.0209468062967062,
-0.09176348149776459,
-0.011987527832388878,
0.04244082048535347,
0.027010733261704445,
0.11495955288410187,
0.0680142194032669,
-0.04079704359173775,
-0.013794795610010624,
-0.036830250173807144,
-0.062414366751909256,
0.04103309288620949,
0.011801579035818577,
0.022654512897133827,
-0.07916375249624252,
-0.0077037205919623375
] | 0 | 0 | 0 | 0 |
|
The Unreasonable Ineffectiveness of the Deeper Layers | https://openreview.net/forum?id=ngmEcEer8a | [
"Andrey Gromov",
"Kushal Tirumala",
"Hassan Shapourian",
"Paolo Glorioso",
"Dan Roberts"
] | Poster | How is knowledge stored in an LLM’s weights? We study this via layer pruning: if removing a certain layer does not affect model performance in common question-answering benchmarks, then the weights in that layer are not necessary for storing the knowledge needed to answer those questions. To find these unnecessary parameters, we identify the optimal block of layers to prune by considering similarity across layers; then, to “heal” the damage, we perform a small amount of finetuning. Surprisingly, with this method we find minimal degradation of performance until after a large fraction (up to half) of the layers are removed for some common open-weight models. From a scientific perspective, the robustness of these LLMs to the deletion of layers implies either that current pretraining methods are not properly leveraging the parameters in the deeper layers of the network or that the shallow layers play a critical role in storing knowledge. For our study, we use parameter-efficient finetuning (PEFT) methods, specifically quantization and Low Rank Adapters (QLoRA), such that each of our experiments can be performed on a single 40GB A100 GPU. | NLP, Pruning, Science of Deep Learning, Efficient Inference | We use model pruning as tool to understand how and where knowledge is located in open-weight LLMs: we find that we can remove up to half the layers of Llama-2 70B with essentially no impact on performance on QA benchmarks. | 13,737 | 2403.17887 | [
-0.06773434579372406,
-0.04332989454269409,
0.05669306218624115,
0.06660756468772888,
0.04941438511013985,
-0.00973748229444027,
-0.04740341380238533,
0.016833556815981865,
0.03573419898748398,
-0.037793174386024475,
-0.01273720245808363,
0.01894397661089897,
0.03957858681678772,
0.030612464994192123,
-0.047374825924634933,
0.03957260400056839,
0.10338092595338821,
0.12438502907752991,
-0.15181384980678558,
0.011579391546547413,
-0.01832459308207035,
-0.03792108595371246,
-0.017192043364048004,
0.04152145981788635,
0.04096302390098572,
-0.07055991888046265,
-0.044952020049095154,
-0.01545387040823698,
0.035499729216098785,
-0.06127464026212692,
0.06413765996694565,
0.022022202610969543,
-0.02552133984863758,
0.05579651892185211,
-0.06800640374422073,
0.08692007511854172,
-0.06891920417547226,
-0.035555750131607056,
-0.08491996675729752,
-0.08000313490629196,
-0.010222801007330418,
0.020810170099139214,
-0.06235333904623985,
0.05141604691743851,
0.04814552515745163,
0.049807559698820114,
0.015767250210046768,
0.0032647415064275265,
-0.06419757008552551,
-0.05235118046402931,
-0.022262532263994217,
0.023904766887426376,
-0.06944770365953445,
0.08171188831329346,
0.006688481196761131,
-0.006935188081115484,
0.034229494631290436,
-0.052696917206048965,
-0.08274074643850327,
0.03730271756649017,
-0.04530851170420647,
-0.0871695801615715,
-0.03757501393556595,
-0.004872098099440336,
0.05398067459464073,
-0.04587914049625397,
0.06532856822013855,
0.04504678025841713,
0.04926180839538574,
0.05106862261891365,
0.01929830200970173,
0.05041027069091797,
-0.09827980399131775,
0.02471543289721012,
0.04326172173023224,
0.0971883162856102,
0.04136304557323456,
-0.016985943540930748,
-0.030861418694257736,
-0.05167364329099655,
0.07834064960479736,
-0.002276673447340727,
0.011730075813829899,
-0.016280023381114006,
0.095035620033741,
-0.05891045928001404,
-0.021468807011842728,
-0.0008962706197053194,
0.013328121043741703,
-0.014920666813850403,
-0.014440595172345638,
0.035660114139318466,
-0.017644058912992477,
0.025519324466586113,
0.12364412844181061,
0.0058064847253263,
-0.014284898526966572,
-0.06768649071455002,
-0.06689810007810593,
0.09220114350318909,
0.017064914107322693,
0.05670857056975365,
0.029049407690763474,
0.021112769842147827,
0.014702805317938328,
0.009397462010383606,
0.019962573423981667,
0.091147281229496,
-0.025038717314600945,
-0.11735192686319351,
0.05193323642015457,
0.016767146065831184,
-0.0020075333304703236,
0.019287629052996635,
0.04276587814092636,
-0.02718404121696949,
0.05197198688983917,
-0.0013646819861605763,
-0.048426058143377304,
0.0685611367225647,
-0.10915970057249069,
-0.02197134494781494,
-0.013536220416426659,
-0.015633098781108856,
-0.0006568850367330015,
0.05026973783969879,
-0.09847242385149002,
3.9304776856414356e-33,
0.09796470403671265,
0.03693346306681633,
0.0057364352978765965,
-0.01972089149057865,
0.020596912130713463,
0.008447745814919472,
0.07009575515985489,
0.029731430113315582,
-0.026573117822408676,
0.026568030938506126,
-0.07627293467521667,
0.027405109256505966,
-0.06825848668813705,
0.06057644262909889,
0.1098138689994812,
-0.029439223930239677,
-0.03711667284369469,
0.09938555955886841,
-0.024299494922161102,
-0.07282952964305878,
0.09887194633483887,
-0.042194440960884094,
-0.0031403745524585247,
-0.05882103741168976,
-0.007433854043483734,
-0.001605067402124405,
0.053703296929597855,
-0.027643028646707535,
-0.08819904923439026,
0.02393929474055767,
-0.08089718967676163,
-0.02723039500415325,
-0.008399983868002892,
0.039252061396837234,
-0.021723855286836624,
0.006649603135883808,
-0.014375670813024044,
-0.021268615499138832,
0.039894670248031616,
-0.06181856617331505,
-0.0037840416189283133,
0.0507042296230793,
0.04015548154711723,
-0.009813676588237286,
-0.056208912283182144,
-0.07033522427082062,
-0.012340368703007698,
0.019871598109602928,
-0.09467735886573792,
-0.049268826842308044,
0.09885615110397339,
-0.016323579475283623,
-0.04802476614713669,
-0.045757781714200974,
-0.04123584181070328,
-0.03553221374750137,
0.053366176784038544,
0.014721608720719814,
0.03815899044275284,
0.10637940466403961,
0.010437767952680588,
0.05147881060838699,
-0.0995430052280426,
0.09424038976430893,
0.05750405788421631,
0.002927782479673624,
-0.016703350469470024,
0.034255243837833405,
-0.020809441804885864,
-0.0028592669405043125,
-0.05227773264050484,
0.02875181846320629,
-0.014024491421878338,
-0.07922881841659546,
0.08163157850503922,
-0.03862021490931511,
-0.015915006399154663,
-0.09831458330154419,
0.03327549993991852,
-0.00810991134494543,
0.028842706233263016,
0.10843402147293091,
-0.02961704507470131,
-0.021593855693936348,
-0.07784250378608704,
-0.038863539695739746,
0.05517183616757393,
-0.046917084604501724,
0.017216654494404793,
0.07017754018306732,
-0.011940422467887402,
-0.07274886965751648,
-0.02035791240632534,
-0.009432279504835606,
-0.028494562953710556,
-1.911683136526262e-33,
-0.017644735053181648,
-0.019034182652831078,
-0.008048299700021744,
0.14306680858135223,
0.025333508849143982,
-0.11046630889177322,
-0.05193300545215607,
0.07007986307144165,
-0.09826045483350754,
-0.06835155934095383,
-0.029491662979125977,
-0.05719246342778206,
0.004296933300793171,
-0.005187555216252804,
0.06236638128757477,
-0.02921757847070694,
-0.0837060734629631,
-0.06538171321153641,
0.0008156454423442483,
-0.009402808733284473,
-0.017896614968776703,
0.02381369099020958,
-0.07229915261268616,
0.07082676142454147,
-0.01133702788501978,
0.03978316858410835,
-0.06443136185407639,
0.006263032555580139,
0.023991268128156662,
0.022714711725711823,
-0.04509003832936287,
-0.05360069125890732,
-0.050266969949007034,
-0.01996433176100254,
-0.01073585543781519,
0.08112514764070511,
0.047677475959062576,
0.04615970328450203,
0.00048090191558003426,
0.03686913847923279,
-0.0017222727183252573,
0.00369128305464983,
-0.12455973774194717,
-0.03516567125916481,
0.030753061175346375,
-0.024996526539325714,
-0.09280810505151749,
-0.05447116121649742,
0.027141587808728218,
0.001177509082481265,
0.02237938717007637,
0.0026229482609778643,
0.008812476880848408,
0.004066027235239744,
0.028948325663805008,
-0.011941866017878056,
-0.07319885492324829,
0.01945062354207039,
0.026226894930005074,
-0.010814271867275238,
-0.004572402220219374,
-0.056521009653806686,
-0.03134211525321007,
-0.01678428053855896,
0.0075456686317920685,
-0.03668854758143425,
0.03893183171749115,
0.06180591508746147,
-0.048558156937360764,
0.03813447430729866,
-0.031161392107605934,
-0.0028832126408815384,
0.004314833786338568,
-0.046579040586948395,
-0.0006621430511586368,
-0.010152622126042843,
-0.007965506054461002,
0.011548575945198536,
0.008522043004631996,
-0.000562658766284585,
-0.017123347148299217,
-0.005689759738743305,
0.029875636100769043,
0.0003363995929248631,
0.08861131221055984,
0.04635315015912056,
0.06987636536359787,
0.06921453028917313,
0.030193014070391655,
-0.0010170303285121918,
-0.03466305509209633,
-0.01223189476877451,
0.03537134826183319,
0.02014704793691635,
-0.04803701490163803,
-4.8982780498363354e-8,
-0.029127411544322968,
0.04228922724723816,
-0.0020730197429656982,
0.049351051449775696,
0.033548880368471146,
-0.033698149025440216,
0.016483820974826813,
0.18182873725891113,
0.003456246806308627,
0.02116881124675274,
0.08552828431129456,
-0.00409719068557024,
-0.04374513775110245,
-0.00809035450220108,
-0.007678676396608353,
0.09542416781187057,
0.005385787226259708,
-0.004717029631137848,
-0.014959726482629776,
-0.038167234510183334,
0.01878724992275238,
0.07963234186172485,
0.005210746079683304,
0.03751261159777641,
0.04753319174051285,
-0.059206120669841766,
0.02244950458407402,
0.10854640603065491,
0.0903657078742981,
0.0911191776394844,
0.0021056837867945433,
-0.0593724325299263,
0.030048275366425514,
0.0114297429099679,
0.0661950632929802,
0.11429419368505478,
-0.009346673265099525,
0.01501061674207449,
-0.00910855270922184,
-0.013850639574229717,
-0.009019961580634117,
0.05047007277607918,
0.009669894352555275,
0.04291411489248276,
0.09780963510274887,
-0.048893604427576065,
-0.08084004372358322,
0.01861828751862049,
-0.02853926457464695,
-0.0008394913165830076,
-0.021588224917650223,
0.0006017440464347601,
-0.05903889611363411,
0.06819412857294083,
0.033266596496105194,
0.004227395169436932,
0.010702135972678661,
-0.03380908817052841,
0.003971036523580551,
-0.04508117213845253,
0.04291132837533951,
-0.09867726266384125,
-0.061080072075128555,
0.03161818906664848
] | 0 | 0 | 0 | 0 |
|
Distilling Dataset into Neural Field | https://openreview.net/forum?id=nCrJD7qPJN | [
"Donghyeok Shin",
"HeeSun Bae",
"Gyuwon Sim",
"Wanmo Kang",
"Il-chul Moon"
] | Poster | Utilizing a large-scale dataset is essential for training high-performance deep learning models, but it also comes with substantial computation and storage costs. To overcome these challenges, dataset distillation has emerged as a promising solution by compressing the large-scale dataset into a smaller synthetic dataset that retains the essential information needed for training. This paper proposes a novel parameterization framework for dataset distillation, coined Distilling Dataset into Neural Field (DDiF), which leverages the neural field to store the necessary information of the large-scale dataset. Due to the unique nature of the neural field, which takes coordinates as input and output quantity, DDiF effectively preserves the information and easily generates various shapes of data. We theoretically confirm that DDiF exhibits greater expressiveness than some previous literature when the utilized budget for a single synthetic instance is the same. Through extensive experiments, we demonstrate that DDiF achieves superior performance on several benchmark datasets, extending beyond the image domain to include video, audio, and 3D voxel. We release the code at \url{https://github.com/aailab-kaist/DDiF}. | Dataset distillation, Dataset condensation, Neural field | This paper proposes an utilization framework of neural field for dataset distillation. | 13,708 | 2503.04835 | [
-0.08915054798126221,
-0.05931934341788292,
0.07023239135742188,
0.013514011166989803,
0.01568031497299671,
-0.030214928090572357,
-0.032784536480903625,
-0.06550934165716171,
-0.022813698276877403,
-0.08673232048749924,
-0.05779401212930679,
-0.021383751183748245,
-0.01062872726470232,
-0.03544795140624046,
-0.070171058177948,
0.034280698746442795,
0.0414133183658123,
0.14448675513267517,
-0.1823599487543106,
0.006866689305752516,
-0.011700405739247799,
-0.03224123641848564,
-0.024769069626927376,
0.014598075300455093,
0.10181720554828644,
0.04216587543487549,
0.06583324819803238,
-0.04454013705253601,
0.05371765047311783,
-0.07138119637966156,
0.07249563187360764,
-0.03037935495376587,
-0.07996203005313873,
0.050995390862226486,
-0.008550780825316906,
0.07007017731666565,
-0.014545026235282421,
0.02863720804452896,
-0.022787922993302345,
-0.03239830210804939,
0.053509775549173355,
0.08157454431056976,
0.0321866013109684,
0.059026509523391724,
0.04694905877113342,
0.03528434410691261,
0.03281482309103012,
-0.054368700832128525,
0.08134205639362335,
0.02484467811882496,
-0.05869355425238609,
0.01557756308466196,
-0.03448032960295677,
0.15391188859939575,
-0.008965504355728626,
0.012797638773918152,
0.05070574954152107,
0.07396595180034637,
-0.07866378873586655,
0.044043801724910736,
-0.04130600020289421,
-0.005546221975237131,
0.04082607850432396,
-0.01804177463054657,
-0.01568339206278324,
-0.04835384711623192,
0.028554828837513924,
0.03635241091251373,
0.038703273981809616,
-0.01516670174896717,
-0.013455095700919628,
0.1298648566007614,
0.0032886317931115627,
0.02461659163236618,
0.023450706154108047,
0.05217908322811127,
0.030651919543743134,
-0.013573713600635529,
0.0738227516412735,
-0.0827142670750618,
0.05434166267514229,
-0.0027399668470025063,
0.031915679574012756,
-0.003511896589770913,
0.0793667659163475,
-0.0061963340267539024,
0.013856444507837296,
0.016226226463913918,
0.003893783316016197,
-0.005749586969614029,
-0.08487796783447266,
-0.008402973413467407,
0.02456500567495823,
-0.03980349749326706,
-0.021451342850923538,
0.04458798095583916,
0.0038907688576728106,
-0.06146484613418579,
-0.050952259451150894,
0.0754101499915123,
-0.02948608435690403,
0.04736475646495819,
0.03528115153312683,
-0.028735684230923653,
-0.001609815051779151,
-0.09126521646976471,
0.09072860330343246,
0.05577662214636803,
-0.012871635146439075,
-0.05003415793180466,
0.05603532865643501,
0.07110356539487839,
0.013884267769753933,
-0.010275067761540413,
0.032585203647613525,
0.00022979453206062317,
-0.07671285420656204,
-0.029468834400177002,
-0.030443303287029266,
-0.0007619461393915117,
-0.0729447603225708,
0.005266634281724691,
-0.09721572697162628,
0.03404349833726883,
-0.07768338918685913,
-0.07785762846469879,
-0.04037162661552429,
4.722144478245445e-33,
0.039746981114149094,
0.01363519299775362,
0.010320909321308136,
-0.008943680673837662,
0.04091940447688103,
-0.02265036106109619,
-0.0017723432974889874,
0.0567292645573616,
-0.061289798468351364,
0.03178899735212326,
-0.07301086187362671,
-0.046861350536346436,
-0.09103850275278091,
0.07065065205097198,
0.009668483398854733,
-0.034549564123153687,
-0.032330822199583054,
0.0803055465221405,
-0.0013776670675724745,
-0.006544496398419142,
0.06739792972803116,
-0.0006086506182327867,
-0.04044346138834953,
-0.011180072091519833,
-0.0010014589643105865,
-0.014622115530073643,
0.030913686379790306,
0.048027992248535156,
-0.02185455895960331,
0.007480286527425051,
-0.1038740947842598,
-0.042566973716020584,
0.042818304151296616,
0.04969551041722298,
0.011852274648845196,
-0.06066957116127014,
-0.028984008356928825,
-0.03481312468647957,
0.024506546556949615,
0.021197328343987465,
0.05101379379630089,
0.05082884430885315,
-0.03408312797546387,
-0.0004843480128329247,
-0.09320160746574402,
-0.030907711014151573,
0.06237516179680824,
0.010612331330776215,
-0.06677433848381042,
-0.033259715884923935,
0.035138022154569626,
0.023061858490109444,
0.03000139258801937,
-0.09383663535118103,
-0.030903825536370277,
-0.08270014077425003,
0.044601261615753174,
-0.05275183171033859,
0.052166301757097244,
0.07050090283155441,
-0.03634538874030113,
0.06224551796913147,
-0.05496271327137947,
0.08290483802556992,
-0.03159213811159134,
-0.000644623942207545,
0.02725031040608883,
0.058532409369945526,
0.055399831384420395,
-0.013726441189646721,
-0.07584123313426971,
0.013197544030845165,
-0.06798357516527176,
-0.03653760254383087,
0.020612437278032303,
-0.033654678612947464,
0.02828041836619377,
-0.04368693754076958,
0.0007330499938689172,
0.01195082999765873,
-0.004196865018457174,
0.050182685256004333,
-0.03563333675265312,
0.02744842879474163,
-0.02487456612288952,
0.03577784076333046,
0.020408298820257187,
-0.06445751339197159,
-0.031991299241781235,
-0.02845362201333046,
-0.09276128560304642,
-0.03631038963794708,
-0.05003322288393974,
-0.07770829647779465,
0.015223681926727295,
-2.8822504347628308e-33,
-0.002857829676941037,
0.03424236550927162,
-0.0717642605304718,
0.08635789155960083,
-0.002387027023360133,
0.018818821758031845,
-0.04712672531604767,
0.03663753718137741,
0.029063574969768524,
-0.011446458287537098,
-0.018906019628047943,
-0.0752105712890625,
-0.003124272683635354,
-0.06553366780281067,
-0.020526433363556862,
0.001609821105375886,
-0.05123081058263779,
-0.03829261660575867,
-0.05267079547047615,
-0.02650236152112484,
-0.02346714772284031,
0.10060535371303558,
-0.03432087600231171,
0.016350949183106422,
-0.11200539767742157,
0.03440133482217789,
-0.027153030037879944,
0.10200078785419464,
0.03545871376991272,
0.005362865049391985,
-0.038831956684589386,
-0.00685526616871357,
-0.019600236788392067,
-0.08104041963815689,
-0.09708724170923233,
0.06144518405199051,
0.10509517043828964,
-0.01840943470597267,
-0.06483922153711319,
0.023299941793084145,
0.03953409194946289,
-0.02733190916478634,
-0.0698581337928772,
0.03312479332089424,
-0.012578573077917099,
0.010745896957814693,
-0.0717458426952362,
-0.009766209870576859,
0.050199780613183975,
0.06943058967590332,
0.017560584470629692,
0.028055815026164055,
-0.04559389129281044,
0.013669604435563087,
0.019137423485517502,
-0.09308179467916489,
-0.045018285512924194,
0.003364258212968707,
0.025676386430859566,
0.029305048286914825,
-0.0697409063577652,
-0.08131309598684311,
-0.008449329994618893,
-0.03398025780916214,
-0.047286052256822586,
-0.0015760521637275815,
-0.01769605465233326,
0.01829882338643074,
0.005448896437883377,
0.03066830337047577,
0.039066240191459656,
-0.045239925384521484,
0.0574536956846714,
0.0632624700665474,
-0.11268430948257446,
-0.027068575844168663,
0.04634031280875206,
-0.019282616674900055,
0.059005290269851685,
0.05660314857959747,
0.05976209044456482,
0.016018159687519073,
0.015918590128421783,
0.04312038794159889,
0.12790994346141815,
0.09225771576166153,
-0.00459458539262414,
0.05216105654835701,
-0.014415315352380276,
0.0846557468175888,
-0.01931304857134819,
0.015628252178430557,
-0.050566330552101135,
0.08910057693719864,
-0.007276149466633797,
-4.3104666502813416e-8,
-0.04541657492518425,
0.1302347481250763,
0.04272596910595894,
0.037752680480480194,
-0.07071720063686371,
-0.0529274120926857,
0.08817072957754135,
0.15147103369235992,
0.03999608755111694,
-0.0040153260342776775,
0.11725801974534988,
-0.001961491536349058,
-0.007537566591054201,
-0.04548296332359314,
0.01052122376859188,
0.09698069840669632,
0.0033494376111775637,
-0.008470830507576466,
-0.00439448282122612,
0.035228483378887177,
0.050252869725227356,
0.024858562275767326,
-0.009270301088690758,
-0.04329397529363632,
0.1316428780555725,
-0.08534765988588333,
-0.009425142779946327,
0.0759676843881607,
0.026400480419397354,
0.023020677268505096,
-0.03433888778090477,
0.009415117092430592,
0.030556723475456238,
-0.0725541040301323,
-0.010188756510615349,
0.04901769384741783,
-0.020351797342300415,
0.004758468829095364,
-0.020680859684944153,
-0.0801580399274826,
-0.010662611573934555,
0.03445703163743019,
-0.0426616333425045,
0.007485248148441315,
-0.005153911653906107,
-0.005902031436562538,
0.05311445891857147,
0.032233089208602905,
0.031855568289756775,
0.05301712080836296,
-0.018178818747401237,
0.0005737738101743162,
0.03890363499522209,
0.0369952917098999,
0.03784320876002312,
-0.0018383687129244208,
-0.052777498960494995,
-0.011598673649132252,
-0.024372737854719162,
0.0540790818631649,
-0.008231491781771183,
-0.02205567993223667,
-0.006581202615052462,
0.014545277692377567
] | https://github.com/aailab-kaist/ddif | 13 | 0 | 0 | 0 |
Relax and Merge: A Simple Yet Effective Framework for Solving Fair k-Means and k-sparse Wasserstein Barycenter Problems | https://openreview.net/forum?id=n8h1z588eu | [
"Shihong Song",
"Guanlin Mo",
"Hu Ding"
] | Poster | The fairness of clustering algorithms has gained widespread attention across various areas, including machine learning, In this paper, we study fair $k$-means clustering in Euclidean space.
Given a dataset comprising several groups, the fairness constraint requires that each cluster should contain a proportion of points from each group within specified lower and upper bounds.
Due to these fairness constraints, determining the optimal locations of $k$ centers is a quite challenging task.
We propose a novel ``Relax and Merge'' framework that returns a $(1+4\rho + O(\epsilon))$-approximate solution, where $\rho$ is the approximate ratio of an off-the-shelf vanilla $k$-means algorithm and $O(\epsilon)$ can be an arbitrarily small positive number. If equipped with a PTAS of $k$-means, our solution can achieve an approximation ratio of $(5+O(\epsilon))$ with only a slight violation of the fairness constraints, which improves the current state-of-the-art approximation guarantee. Furthermore, using our framework, we can also obtain a $(1+4\rho +O(\epsilon))$-approximate solution for the $k$-sparse Wasserstein Barycenter problem, which is a fundamental optimization problem in the field of optimal transport, and a $(2+6\rho)$-approximate solution for the strictly fair $k$-means clustering with no violation, both of which are better than the current state-of-the-art methods. In addition, the empirical results demonstrate that our proposed algorithm can significantly outperform baseline approaches in terms of clustering cost. | clustering, k-means, fairness, approxiamte algorithm, optimal transport | An improved algorithm for fair k-means problem. | 13,707 | null | [
0.0011310054687783122,
-0.027339771389961243,
0.02648085355758667,
-0.017848564311861992,
0.03725672513246536,
-0.04147963598370552,
0.03340733051300049,
-0.02622545324265957,
0.01168962474912405,
0.05643872171640396,
-0.004460005555301905,
-0.03795577958226204,
0.05301998555660248,
0.019820010289549828,
-0.08657597750425339,
0.004000856541097164,
0.11428377777338028,
0.04944445937871933,
-0.06167084723711014,
0.004258257336914539,
-0.04260993003845215,
-0.03095867671072483,
0.017439326271414757,
0.06408146023750305,
0.03965859115123749,
-0.047810882329940796,
0.07954965531826019,
-0.034494757652282715,
-0.02777957357466221,
-0.0017578923143446445,
-0.03996521607041359,
-0.010414404794573784,
0.06432792544364929,
-0.05015227571129799,
-0.02699257992208004,
0.06745047867298126,
0.01546595711261034,
0.02127091959118843,
-0.042903680354356766,
0.030385835096240044,
-0.03928859531879425,
0.065956249833107,
0.014347698539495468,
0.043197184801101685,
-0.03399055078625679,
0.06081195920705795,
0.02102431282401085,
-0.0006984646897763014,
0.03728707134723663,
-0.0801025852560997,
0.02010205201804638,
-0.002465846948325634,
-0.06379736214876175,
0.046141888946294785,
-0.0390331894159317,
-0.05692465975880623,
-0.09401419013738632,
0.008007216267287731,
0.027852538973093033,
0.028143443167209625,
-0.020162824541330338,
-0.16297900676727295,
0.015236899256706238,
0.006914944853633642,
0.006919035688042641,
0.015788882970809937,
0.07139956206083298,
0.08867762982845306,
-0.05578107014298439,
0.016874292865395546,
0.04952150955796242,
0.08073168247938156,
-0.016408735886216164,
0.025742482393980026,
0.06081000715494156,
0.0752110630273819,
0.032216429710388184,
-0.026354843750596046,
0.06639336794614792,
0.029564600437879562,
0.008416596800088882,
0.012410098686814308,
-0.026319170370697975,
-0.021792326122522354,
0.06694851815700531,
-0.02520889788866043,
-0.013587492518126965,
0.01884608156979084,
0.10430335253477097,
-0.05006616935133934,
0.03460313752293587,
0.04657899588346481,
-0.041450053453445435,
-0.09776293486356735,
-0.034222960472106934,
-0.05366326868534088,
0.07997722178697586,
0.03638767823576927,
0.06763067841529846,
0.12628503143787384,
-0.042655885219573975,
0.0466940812766552,
-0.0005710613331757486,
-0.11239549517631531,
0.03575646132230759,
-0.02354666404426098,
0.019424818456172943,
0.03247633948922157,
0.04754658415913582,
-0.0667356327176094,
0.007404796779155731,
-0.04795337840914726,
-0.017224861308932304,
0.05707044154405594,
-0.036083199083805084,
-0.0354078970849514,
0.12989738583564758,
-0.08247002959251404,
0.022350862622261047,
0.0003645300166681409,
-0.05819759517908096,
-0.001811246038414538,
0.05887159705162048,
0.05732553452253342,
0.034986577928066254,
0.036544520407915115,
-0.05562446266412735,
5.1005131509011e-34,
0.023647066205739975,
0.0766414925456047,
0.04583755135536194,
-0.07506602257490158,
-0.007815930061042309,
0.006411159876734018,
0.008706895634531975,
0.03905513137578964,
-0.016798362135887146,
-0.003751252545043826,
-0.0012724518310278654,
-0.01451209094375372,
-0.01575222983956337,
-0.05741569399833679,
0.0773128792643547,
-0.07381312549114227,
0.03146682679653168,
0.05021440610289574,
-0.07055120915174484,
-0.08499405533075333,
0.05649678409099579,
-0.036353953182697296,
-0.030084317550063133,
0.019642716273665428,
-0.09982185810804367,
-0.03956400975584984,
0.013332045637071133,
-0.05557185038924217,
0.10345110297203064,
0.0029762820340692997,
-0.10589109361171722,
0.024423880502581596,
0.03731314837932587,
0.0961587205529213,
-0.03069399483501911,
0.04578569158911705,
-0.02941887453198433,
-0.03497139737010002,
0.01196304988116026,
-0.06240338087081909,
-0.05203266441822052,
0.06814990937709808,
0.026183465495705605,
-0.08482832461595535,
-0.008186684921383858,
-0.044756222516298294,
0.049308691173791885,
-0.00664740614593029,
0.02408546395599842,
-0.03286192938685417,
-0.007754265796393156,
-0.01837676204741001,
0.05498220771551132,
0.011055324226617813,
-0.002742272336035967,
0.017479104921221733,
-0.02271461859345436,
-0.014883885160088539,
0.07330499589443207,
0.08866377174854279,
0.0203546192497015,
-0.03747138753533363,
-0.03972001373767853,
0.028563518077135086,
0.0056554777547717094,
0.004495210479944944,
0.017607953399419785,
0.04759705811738968,
0.07413782179355621,
0.034029506146907806,
0.0018931328086182475,
0.02519962377846241,
-0.03937326371669769,
-0.03534258157014847,
-0.040274422615766525,
0.04557231441140175,
0.08979582786560059,
0.070086270570755,
-0.029492055997252464,
-0.08248356729745865,
0.023306047543883324,
-0.025493074208498,
-0.047305285930633545,
-0.04355883225798607,
-0.04327171668410301,
-0.029215628281235695,
0.0704503059387207,
0.020517684519290924,
-0.110981784760952,
-0.031319405883550644,
-0.07859757542610168,
0.010928280651569366,
0.01802017353475094,
0.028931643813848495,
0.009090402163565159,
-5.184565588759701e-34,
0.029995637014508247,
-0.007598259020596743,
-0.020688725635409355,
0.06798846274614334,
0.031189726665616035,
-0.025668535381555557,
-0.03368736058473587,
-0.04350842908024788,
0.014645734801888466,
-0.06242746859788895,
-0.04414515942335129,
-0.06806532293558121,
0.1523180454969406,
0.014180952683091164,
-0.012506754137575626,
0.00986372958868742,
-0.0071586258709430695,
0.09336152672767639,
0.029802188277244568,
0.04674394428730011,
0.019812731072306633,
-0.04179894179105759,
-0.007059558294713497,
0.03304662927985191,
0.02957633137702942,
0.04349558427929878,
-0.056399986147880554,
0.021183840930461884,
-0.06965400278568268,
0.059444308280944824,
-0.049072083085775375,
-0.08104733377695084,
-0.11098325252532959,
-0.04352426901459694,
0.002985776402056217,
-0.04687480255961418,
0.016503676772117615,
0.051045503467321396,
-0.007355148438364267,
0.022178705781698227,
0.05042993277311325,
-0.03714451566338539,
-0.11896669119596481,
-0.009257051162421703,
0.011659245006740093,
-0.0027341421227902174,
-0.01826479472219944,
-0.06875043362379074,
-0.027942242100834846,
-0.06920572370290756,
-0.05410308390855789,
-0.0448027141392231,
-0.011819650419056416,
0.10415906459093094,
-0.016237277537584305,
-0.006505935452878475,
0.023897685110569,
0.07065281271934509,
0.026464303955435753,
-0.036630548536777496,
-0.043782658874988556,
-0.10197745263576508,
-0.12052254378795624,
0.07274165004491806,
-0.018255790695548058,
-0.05156382545828819,
0.01356884092092514,
-0.08667748421430588,
-0.09139915555715561,
0.07868772745132446,
-0.049982115626335144,
-0.0159797091037035,
0.042817503213882446,
0.005734647624194622,
-0.06399506330490112,
0.03338688611984253,
0.09978518635034561,
-0.010929563082754612,
-0.02112087607383728,
0.015611412934958935,
-0.001295791007578373,
-0.009887331165373325,
0.0760624036192894,
-0.013484659604728222,
0.10461466014385223,
-0.0012800818076357245,
0.06464240700006485,
-0.03663603961467743,
0.015992548316717148,
0.04824372008442879,
-0.054242584854364395,
0.0141240693628788,
0.06434733420610428,
0.0042336611077189445,
-0.08217048645019531,
-4.743479919966376e-8,
-0.0222760196775198,
-0.013779893517494202,
-0.03319259360432625,
-0.0443229004740715,
0.07462404668331146,
-0.027838023379445076,
-0.03259774670004845,
0.04216845706105232,
-0.09085118025541306,
-0.021471716463565826,
0.0949372798204422,
-0.06562169641256332,
-0.008507220074534416,
-0.00874034408479929,
-0.020395394414663315,
0.0637768879532814,
-0.04525066539645195,
-0.0019945199601352215,
0.0005321924109011889,
0.06374429166316986,
-0.056212760508060455,
-0.011790287680923939,
0.05275104567408562,
-0.048685554414987564,
-0.008850346319377422,
0.018487248569726944,
0.04307674989104271,
0.044834475964307785,
0.04451250284910202,
-0.0619155615568161,
-0.06057896837592125,
-0.030525853857398033,
0.08072555810213089,
-0.019601495936512947,
-0.014554589986801147,
0.11104186624288559,
-0.11250588297843933,
0.08145041018724442,
-0.09538505226373672,
0.013470331206917763,
0.040087271481752396,
0.04854400083422661,
-0.002699705772101879,
-0.05530444532632828,
0.10659191757440567,
0.02883361279964447,
0.026227621361613274,
0.0876605436205864,
0.009632289409637451,
-0.033860158175230026,
0.004869419150054455,
0.042545583099126816,
-0.011300487443804741,
0.021984774619340897,
0.057807039469480515,
-0.04029159992933273,
0.002803903305903077,
-0.05277406796813011,
0.055784329771995544,
-0.03664825111627579,
0.07419027388095856,
-0.016019171103835106,
-0.038448065519332886,
-0.022069571539759636
] | 0 | 0 | 0 | 0 |
|
Neural Dueling Bandits: Preference-Based Optimization with Human Feedback | https://openreview.net/forum?id=VELhv9BBfn | [
"Arun Verma",
"Zhongxiang Dai",
"Xiaoqiang Lin",
"Patrick Jaillet",
"Bryan Kian Hsiang Low"
] | Poster | Contextual dueling bandit is used to model the bandit problems, where a learner's goal is to find the best arm for a given context using observed noisy human preference feedback over the selected arms for the past contexts. However, existing algorithms assume the reward function is linear, which can be complex and non-linear in many real-life applications like online recommendations or ranking web search results. To overcome this challenge, we use a neural network to estimate the reward function using preference feedback for the previously selected arms. We propose upper confidence bound- and Thompson sampling-based algorithms with sub-linear regret guarantees that efficiently select arms in each round. We also extend our theoretical results to contextual bandit problems with binary feedback, which is in itself a non-trivial contribution. Experimental results on the problem instances derived from synthetic datasets corroborate our theoretical results. | Contextual Dueling Bandits, Preferences Learning, Human Feedback, Neural Bandits, Thompson Sampling | We study contextual dueling bandits problem and propose upper confidence bound- and Thompson sampling-based algorithms that use a neural network to estimate the reward function using human preference feedback and have sub-linear regret guarantees. | 13,697 | null | [
-0.13467876613140106,
-0.04005115479230881,
0.01608957163989544,
0.039872586727142334,
0.00038742137257941067,
0.030830323696136475,
0.08773380517959595,
0.036456525325775146,
-0.004121554549783468,
-0.03759130463004112,
-0.064114049077034,
-0.011233669705688953,
0.09462884068489075,
-0.02642212249338627,
-0.009480905719101429,
0.049233853816986084,
0.1170332282781601,
0.0596216544508934,
0.02892005257308483,
0.038315217941999435,
-0.053978342562913895,
0.01134116854518652,
0.0828942134976387,
-0.07679585367441177,
0.029997797682881355,
-0.05704294890165329,
-0.030893990769982338,
-0.022964801639318466,
0.02367912046611309,
0.030043000355362892,
0.03342802822589874,
-0.021042710170149803,
-0.020414594560861588,
-0.1023872047662735,
-0.05629705265164375,
-0.0008113127551041543,
-0.1538519412279129,
-0.08549443632364273,
0.02845630794763565,
0.007944479584693909,
-0.010134436190128326,
0.04120957851409912,
-0.01427770871669054,
0.05002640187740326,
0.03570076823234558,
0.0046053603291511536,
0.029338201507925987,
0.00961900595575571,
0.03870915621519089,
0.024719733744859695,
-0.08235558122396469,
-0.02362414263188839,
0.005983473267406225,
-0.022788427770137787,
-0.018827686086297035,
0.014986804686486721,
0.024448147043585777,
0.026036158204078674,
0.002009420655667782,
0.016986556351184845,
0.024118170142173767,
-0.15774114429950714,
-0.03436028212308884,
-0.03145522251725197,
-0.005441823974251747,
-0.00815068930387497,
-0.008573533035814762,
0.06006942689418793,
-0.004187350161373615,
0.03920671343803406,
0.07802030444145203,
0.07858831435441971,
-0.03138779476284981,
0.025598473846912384,
0.08526382595300674,
0.012508382089436054,
-0.007785416673868895,
-0.04110678657889366,
-0.006244052201509476,
0.010292341001331806,
-0.08055684715509415,
-0.03560372814536095,
0.01562293991446495,
-0.02817566692829132,
0.08735723793506622,
-0.014833089895546436,
0.04035057872533798,
0.01851954497396946,
0.09836515784263611,
-0.047923628240823746,
-0.05697327479720116,
0.012530477717518806,
-0.04492610692977905,
-0.014463775791227818,
-0.06054646894335747,
0.0548664927482605,
-0.0013654540525749326,
-0.012260392308235168,
-0.030262360349297523,
0.10745878517627716,
0.04484451562166214,
0.06570491939783096,
-0.06071840599179268,
-0.049943502992391586,
0.001529464847408235,
0.01495016273111105,
0.020572176203131676,
0.07302441447973251,
0.07316339761018753,
-0.1184157133102417,
0.003698447486385703,
0.035952381789684296,
0.034446053206920624,
0.04310530796647072,
0.0018323136027902365,
-0.032668374478816986,
0.03289603441953659,
-0.016917895525693893,
-0.00554300844669342,
0.09270932525396347,
-0.011095551773905754,
0.010621731169521809,
0.034280553460121155,
-0.011760461144149303,
-0.02397545427083969,
0.023273890838027,
-0.020193593576550484,
-5.437911712833271e-34,
0.031481605023145676,
-0.048956286162137985,
0.05284242704510689,
-0.049930963665246964,
0.024666940793395042,
0.018004560843110085,
0.0627216324210167,
-0.04877834767103195,
-0.07799650728702545,
-0.014175666496157646,
0.036973416805267334,
-0.006567683536559343,
-0.06251353025436401,
0.08587872982025146,
0.07891978323459625,
-0.031364068388938904,
-0.005035729613155127,
0.008828488178551197,
0.05721428245306015,
-0.05964934080839157,
0.04594279080629349,
0.033829301595687866,
0.0006544450297951698,
-0.03642437607049942,
0.017348159104585648,
-0.03232414647936821,
-0.042474765330553055,
0.019564436748623848,
0.01358350832015276,
-0.013164197094738483,
-0.07258821278810501,
0.008013906888663769,
-0.03888052701950073,
0.034714099019765854,
0.010326913557946682,
0.0359700508415699,
-0.038270629942417145,
0.03805055096745491,
-0.10101807862520218,
0.06580901145935059,
0.027784820646047592,
0.041261009871959686,
0.06743227690458298,
0.026776697486639023,
-0.08313017338514328,
-0.11698400974273682,
-0.029013948515057564,
-0.006380399689078331,
-0.046910569071769714,
-0.05498098209500313,
-0.03872663155198097,
0.05318371579051018,
0.017010141164064407,
0.028649073094129562,
-0.016601664945483208,
0.022324811667203903,
-0.008246190845966339,
0.0676242932677269,
-0.020979400724172592,
0.0013799479929730296,
0.0699022188782692,
-0.029744772240519524,
0.04936786741018295,
0.007614087779074907,
-0.09207544475793839,
0.1041388139128685,
-0.017656464129686356,
-0.01452675275504589,
0.008894582279026508,
-0.018545005470514297,
0.011776632629334927,
0.037004999816417694,
-0.002474481938406825,
-0.04069505259394646,
0.09413252025842667,
-0.0092176403850317,
0.021921245381236076,
-0.014737741090357304,
0.03165345638990402,
-0.06362095475196838,
0.0581289641559124,
-0.050573498010635376,
-0.026127668097615242,
-0.026182761415839195,
-0.07334572076797485,
-0.005620077718049288,
0.024145567789673805,
-0.14675232768058777,
0.006318018306046724,
0.011814611032605171,
-0.10921969264745712,
0.023514563217759132,
-0.08229809999465942,
0.03874022141098976,
0.028001822531223297,
-2.111743918006065e-34,
0.03032328188419342,
-0.038169797509908676,
0.09494133293628693,
0.01526612974703312,
-0.005789532791823149,
0.04649946093559265,
-0.017698058858513832,
-0.06383422762155533,
-0.011455130763351917,
0.0020914117339998484,
-0.03811313211917877,
0.03768201172351837,
0.03526897728443146,
-0.029298117384314537,
0.023065567016601562,
-0.047355346381664276,
0.03779587894678116,
0.027694378048181534,
0.0271977037191391,
-0.022334294393658638,
0.04822337254881859,
0.08505868166685104,
-0.028961649164557457,
0.009243984706699848,
-0.017385683953762054,
-0.0320698618888855,
-0.019845694303512573,
0.060908034443855286,
-0.01177490409463644,
-0.023730799555778503,
-0.03009577840566635,
-0.10538586974143982,
-0.03990361839532852,
-0.016053635627031326,
-0.01358689647167921,
0.03821251168847084,
0.06854405999183655,
0.04846889153122902,
-0.06487828493118286,
0.14465895295143127,
0.06202370300889015,
-0.023227060213685036,
-0.007243753876537085,
0.020012596622109413,
0.0347520112991333,
0.008110376074910164,
-0.10092028975486755,
-0.02747766301035881,
0.03747745230793953,
-0.02416493557393551,
0.04339034855365753,
0.0785006433725357,
-0.09361539036035538,
0.054749101400375366,
-0.07072155922651291,
-0.11293425410985947,
-0.0005470507894642651,
0.0015123820630833507,
0.16818638145923615,
0.04588593542575836,
-0.042646851390600204,
0.015976544469594955,
-0.07645910233259201,
0.015931883826851845,
-0.018008265644311905,
-0.016736606135964394,
-0.04164814203977585,
0.020793091505765915,
0.02345021441578865,
0.005709612742066383,
-0.023952782154083252,
0.018147416412830353,
0.048403672873973846,
-0.007600446697324514,
-0.020714830607175827,
0.09761705249547958,
0.039485156536102295,
-0.02082231640815735,
0.015607092529535294,
-0.03656838834285736,
0.03163900971412659,
0.003256011288613081,
0.09016747027635574,
-0.002397683449089527,
0.051432643085718155,
0.09037140011787415,
0.03197145462036133,
0.06497596204280853,
-0.031288519501686096,
-0.08259282261133194,
0.032546721398830414,
-0.08468905091285706,
0.0548742413520813,
-0.027064479887485504,
-0.07239138334989548,
-3.9064076418071636e-8,
-0.07537464052438736,
0.0507669672369957,
0.04702197015285492,
0.10715855658054352,
0.029733942821621895,
0.05298621580004692,
0.00820990838110447,
-0.0064123040065169334,
-0.06194229796528816,
-0.017780208960175514,
0.05920444428920746,
0.06404916942119598,
0.0068574994802474976,
-0.06818446516990662,
-0.08486144989728928,
-0.0028997529298067093,
-0.001963761867955327,
-0.03934324160218239,
0.025636611506342888,
-0.02679118514060974,
0.04590335115790367,
0.02083655074238777,
0.0960971787571907,
-0.07411547005176544,
0.02670125663280487,
-0.032388608902692795,
-0.01629316434264183,
0.09876471012830734,
0.04667338356375694,
0.07678224891424179,
-0.04477285221219063,
-0.03234119340777397,
-0.013584649190306664,
-0.03738107904791832,
0.04828066751360893,
0.00183265027590096,
-0.04574395343661308,
-0.051505282521247864,
-0.013861636631190777,
-0.08876770734786987,
-0.024296263232827187,
0.01570569910109043,
0.0008168403874151409,
-0.028222456574440002,
0.006175287067890167,
0.03130783885717392,
0.03804599121212959,
-0.09872081130743027,
0.07016822695732117,
-0.06415953487157822,
0.08121462911367416,
-0.0772218182682991,
0.00027943376335315406,
0.03001510538160801,
0.07631572335958481,
0.003798654768615961,
-0.02134130522608757,
-0.09129590541124344,
-0.06877744942903519,
0.06884267181158066,
0.029126785695552826,
-0.01615137979388237,
-0.10013124346733093,
0.01802254095673561
] | 0 | 0 | 0 | 0 |
|
SPORTU: A Comprehensive Sports Understanding Benchmark for Multimodal Large Language Models | https://openreview.net/forum?id=x1yOHtFfDh | [
"Haotian Xia",
"Zhengbang Yang",
"Junbo Zou",
"Rhys Tracy",
"Yuqing Wang",
"Chi Lu",
"Christopher Lai",
"Yanjun He",
"Xun Shao",
"Zhuoqing Xie",
"Yuan-fang Wang",
"Weining Shen",
"Hanjie Chen"
] | Poster | Multimodal Large Language Models (MLLMs) are advancing the ability to reason about complex sports scenarios by integrating textual and visual information. To comprehensively evaluate their capabilities, we introduce SPORTU, a benchmark designed to assess MLLMs across multi-level sports reasoning tasks. SPORTU comprises two key components: SPORTU-text, featuring 900 multiple-choice questions with human-annotated explanations for rule comprehension and strategy understanding. This component focuses on testing models' ability to reason about sports solely through question-answering (QA), without requiring visual inputs; SPORTU-video, consisting of 1,701 slow-motion video clips across 7 different sports and 12,048 QA pairs, designed to assess multi-level reasoning, from simple sports recognition to complex tasks like foul detection and rule application. We evaluated four prevalent LLMs mainly utilizing few-shot learning paradigms supplemented by chain-of-thought (CoT) prompting on the SPORTU-text part. GPT-4o achieves the highest accuracy of 71\%, but still falls short of human-level performance, highlighting room for improvement in rule comprehension and reasoning. The evaluation for the SPORTU-video part includes 6 proprietary and 8 open-source MLLMs. Experiments show that models fall short on hard tasks that require deep reasoning and rule-based understanding. GPT-4o performs the best with only 57.8\% accuracy on the hard task, showing large room for improvement. We hope that SPORTU will serve as a critical step toward evaluating models' capabilities in sports understanding and reasoning. The dataset is available at [https://github.com/chili-lab/SPORTU](https://github.com/chili-lab/SPORTU). | Multimodal Large Language Models, Sports Understanding, Benchmark | null | 13,686 | 2410.08474 | [
0.054796505719423294,
-0.02953081764280796,
0.004595778416842222,
-0.08235884457826614,
0.15136651694774628,
0.045686274766922,
0.028189269825816154,
0.05338246747851372,
0.07318402081727982,
0.026711009442806244,
-0.07999908179044724,
-0.1078893318772316,
-0.024154430255293846,
0.11937965452671051,
0.07175850123167038,
-0.04539870470762253,
0.033170927315950394,
0.03517545759677887,
-0.11202975362539291,
-0.012197231873869896,
0.05714518204331398,
-0.009231164120137691,
0.08156603574752808,
-0.012043624185025692,
-0.03293876349925995,
0.00870422087609768,
-0.01360609196126461,
0.0010803942568600178,
-0.015914149582386017,
0.011955810710787773,
0.0018989980453625321,
-0.009096702560782433,
0.028434116393327713,
0.09609127044677734,
-0.07228366285562515,
0.023140819743275642,
-0.0502987764775753,
0.02260124497115612,
-0.05826865881681442,
-0.043387413024902344,
-0.08082998543977737,
-0.06508713215589523,
0.021224986761808395,
-0.009665949270129204,
0.10553950816392899,
0.06428968161344528,
-0.02908521145582199,
-0.005062638316303492,
0.03674490377306938,
0.02332363650202751,
-0.11168334633111954,
-0.033561740070581436,
0.0441252738237381,
-0.00002705404767766595,
0.03904232382774353,
0.026473430916666985,
0.011760696768760681,
-0.0010680524865165353,
0.013104473240673542,
-0.0415530800819397,
-0.011325906962156296,
-0.1224910244345665,
-0.026828326284885406,
0.04574119299650192,
-0.044668372720479965,
-0.04985221475362778,
0.006714161019772291,
0.016869422048330307,
-0.027846429497003555,
-0.013362156227231026,
-0.046151306480169296,
0.07037036865949631,
0.014339444227516651,
0.04448755457997322,
-0.014739803038537502,
0.10166031122207642,
-0.024784961715340614,
-0.053843092173337936,
0.009237093850970268,
-0.031061489135026932,
0.021869411692023277,
-0.05256784334778786,
0.027711942791938782,
0.03779027983546257,
0.14292247593402863,
-0.0013544209068641067,
0.005182237364351749,
0.055450599640607834,
-0.03165691718459129,
0.04775303229689598,
-0.06285493820905685,
-0.06012044474482536,
0.06370645761489868,
-0.014851858839392662,
0.0595196969807148,
0.14023031294345856,
-0.08714787662029266,
-0.10089857876300812,
-0.03969234228134155,
0.07561225444078445,
0.049674149602651596,
0.05041225254535675,
-0.0341767743229866,
-0.021858597174286842,
0.009825832210481167,
0.025151345878839493,
0.045358311384916306,
0.026922816410660744,
0.06371249258518219,
-0.0700102373957634,
0.01982763037085533,
0.045505963265895844,
0.020079847425222397,
0.026140188798308372,
-0.024515794590115547,
0.01192479208111763,
0.03893918916583061,
0.03746921941637993,
-0.0028185504488646984,
0.007840395905077457,
-0.008624982088804245,
0.009053963236510754,
0.0035479555372148752,
0.002802841830998659,
0.10423913598060608,
-0.04961519315838814,
-0.00960452388972044,
8.011423790975909e-34,
0.05963745340704918,
0.008350919000804424,
0.0738014206290245,
0.04782222956418991,
0.024956082925200462,
0.006533435545861721,
-0.05839715152978897,
0.00885479710996151,
-0.03012324497103691,
-0.002823425689712167,
-0.03658679500222206,
-0.008644876070320606,
-0.07014349848031998,
0.013239515945315361,
0.06438188999891281,
0.005138953682035208,
-0.13072621822357178,
-0.028065161779522896,
-0.03365510702133179,
0.017924748361110687,
0.05639578029513359,
-0.05386389046907425,
0.046310871839523315,
0.03262085095047951,
-0.02041378803551197,
0.05257230997085571,
0.066493920981884,
-0.07160907983779907,
-0.043804049491882324,
-0.017129823565483093,
-0.12794621288776398,
-0.10017789900302887,
-0.07497040182352066,
-0.0031650562305003405,
0.06246316060423851,
-0.07992462068796158,
-0.025782324373722076,
-0.059376273304224014,
0.026706097647547722,
0.0028085699304938316,
-0.04346175119280815,
-0.022763174027204514,
-0.02847713604569435,
-0.025259457528591156,
-0.08783924579620361,
0.0008564242743887007,
-0.03233441337943077,
0.03200215846300125,
-0.09265245497226715,
-0.007495115511119366,
0.06575330346822739,
0.0007603750564157963,
0.012933406047523022,
-0.13478796184062958,
0.01413528248667717,
-0.018358437344431877,
0.020019983872771263,
0.009508354589343071,
-0.029597148299217224,
-0.009874009527266026,
0.039872776716947556,
0.05415725335478783,
0.008922243490815163,
0.08461800217628479,
-0.018487390130758286,
0.041853323578834534,
-0.03232044726610184,
0.10821791738271713,
0.09551378339529037,
-0.030757000669836998,
0.00304747698828578,
0.05530661344528198,
-0.026321012526750565,
-0.04702650010585785,
0.018376482650637627,
0.05171358212828636,
0.045566458255052567,
-0.06910249590873718,
0.045558031648397446,
0.025890471413731575,
0.05028358846902847,
-0.009211081080138683,
-0.03560369834303856,
-0.053593121469020844,
-0.025521423667669296,
-0.054369308054447174,
0.04707130044698715,
-0.14326274394989014,
-0.003823868464678526,
-0.05179931968450546,
-0.08957820385694504,
-0.0018453217344358563,
-0.024399440735578537,
-0.008144485764205456,
-0.022465892136096954,
-1.2268297503261315e-33,
-0.03219675272703171,
-0.05172312259674072,
-0.017860060557723045,
0.04662403091788292,
0.03285130113363266,
-0.0446871779859066,
0.018444038927555084,
0.020780116319656372,
0.03684408590197563,
-0.017986848950386047,
0.001136857084929943,
-0.04965689405798912,
-0.03764056786894798,
0.005239671561866999,
0.007342829369008541,
-0.030598843470215797,
-0.04338766261935234,
-0.0061569660902023315,
-0.021205086261034012,
0.04481495916843414,
0.00112685258500278,
0.057277776300907135,
-0.03279392048716545,
-0.018852071836590767,
-0.11684533208608627,
0.07788149267435074,
-0.02316669374704361,
0.02643827721476555,
-0.03276342153549194,
0.047490157186985016,
0.024164531379938126,
-0.02385564334690571,
0.019419953227043152,
0.007056500297039747,
-0.0645783469080925,
0.053248707205057144,
0.08758757263422012,
-0.06830793619155884,
0.00639674486592412,
0.1076233983039856,
0.12266658246517181,
-0.020925424993038177,
-0.06278099119663239,
-0.02759273536503315,
-0.08220379054546356,
0.026870032772421837,
-0.03605850785970688,
-0.037147827446460724,
-0.03105318360030651,
0.05996417626738548,
0.012142766267061234,
0.0008993922383524477,
-0.08867132663726807,
0.029259705916047096,
-0.034025534987449646,
-0.09116114675998688,
-0.05140262842178345,
-0.0043964022770524025,
0.0006322143017314374,
0.001131779165007174,
-0.05273276939988136,
0.04235124960541725,
-0.06266491115093231,
0.017157582566142082,
0.07225733995437622,
0.043926600366830826,
-0.005736523307859898,
-0.028422651812434196,
-0.03535354137420654,
-0.0426374226808548,
-0.013765149749815464,
-0.03271031752228737,
-0.008770938031375408,
0.016874738037586212,
0.05328723415732384,
0.05892956629395485,
-0.003000048454850912,
0.010758312419056892,
0.021212875843048096,
-0.05846059322357178,
0.00873287208378315,
-0.11565280705690384,
-0.005167019087821245,
0.11087054759263992,
0.06592313200235367,
0.053643155843019485,
0.013271874748170376,
0.08985304832458496,
0.0680561289191246,
0.03524424508213997,
0.053491558879613876,
0.1316164880990982,
0.0662127286195755,
0.07354573905467987,
-0.033470332622528076,
-3.992153452259117e-8,
-0.08121752738952637,
0.03998834639787674,
-0.03115960769355297,
0.05006429925560951,
-0.025982415303587914,
-0.009916532784700394,
-0.038399722427129745,
-0.01710977964103222,
0.032220885157585144,
-0.04153463616967201,
0.05598002299666405,
-0.004441448953002691,
-0.04154052585363388,
-0.0383337140083313,
0.013717425055801868,
0.041653625667095184,
-0.06284686177968979,
0.024388790130615234,
-0.004145143553614616,
0.03155207633972168,
0.06049095466732979,
-0.01874213106930256,
-0.057036444544792175,
0.04714824631810188,
-0.03160989657044411,
-0.08367578685283661,
-0.06585937738418579,
0.06049066409468651,
0.02583315037190914,
-0.05358937755227089,
0.019966794177889824,
0.01503591425716877,
0.007722980808466673,
-0.05558718368411064,
0.05747876688838005,
0.015876533463597298,
0.009009303525090218,
-0.030493682250380516,
0.03971898555755615,
0.026342041790485382,
-0.026156553998589516,
0.031122948974370956,
-0.05104468762874603,
0.031621549278497696,
0.017030643299221992,
0.005242275074124336,
-0.07170482724905014,
-0.0983438715338707,
-0.03756126016378403,
-0.019973043352365494,
-0.04019875451922417,
0.016009511426091194,
0.04234742000699043,
0.10846167057752609,
0.025469843298196793,
0.02890068106353283,
0.01158919557929039,
0.01892913319170475,
0.012182153761386871,
0.006513621192425489,
0.07678958028554916,
0.03859400376677513,
-0.04853006452322006,
0.014589548110961914
] | https://github.com/haotianxia/SPORTU | 13 | 0 | 0 | 0 |
SIM: Surface-based fMRI Analysis for Inter-Subject Multimodal Decoding from Movie-Watching Experiments | https://openreview.net/forum?id=OJsMGsO6yn | [
"Simon Dahan",
"Gabriel Bénédict",
"Logan Zane John Williams",
"Yourong Guo",
"Daniel Rueckert",
"Robert Leech",
"Emma Claire Robinson"
] | Poster | Current AI frameworks for brain decoding and encoding, typically train and test models within the same datasets. This limits their utility for cognitive training (neurofeedback) for which it would be useful to pool experiences across individuals to better simulate stimuli not sampled during training. A key obstacle to model generalisation is the degree of variability of inter-subject cortical organisation, which makes it difficult to align or compare cortical signals across participants. In this paper we address this through use of surface vision transformers, which build a generalisable model of cortical functional dynamics, through encoding the topography of cortical networks and their interactions as a moving image across a surface. This is then combined with tri-modal self-supervised contrastive (CLIP) alignment of audio, video, and fMRI modalities to enable the retrieval of visual and auditory stimuli from patterns of cortical activity (and vice-versa). We validate our approach on 7T task-fMRI data from 174 healthy participants engaged in the movie-watching experiment from the Human Connectome Project (HCP). Results show that it is possible to detect which movie clips an individual is watching purely from their brain activity, even for individuals and movies *not seen during training*. Further analysis of attention maps reveals that our model captures individual patterns of brain activity that reflect semantic and visual systems. This opens the door to future personalised simulations of brain function. Code \& pre-trained models will be made available at https://github.com/metrics-lab/sim. | movie-watching experiment, fMRI, cortical analysis, surface-based transformers, multimodal learning, contrastive learning, self-supervised learning, generalization, encoding/decoding | A surface-based deep learning fMRI model that generalises encoding and decoding of audio-visual stimuli from movie-watching experiments to unseen subjects and unseen stimuli | 13,684 | 2501.16471 | [
-0.00926306564360857,
-0.164290651679039,
0.05202658846974373,
-0.028558921068906784,
0.04731552675366402,
0.050463151186704636,
0.019000519067049026,
-0.015210350044071674,
0.02674029767513275,
-0.04102775454521179,
-0.053133610635995865,
-0.08301761001348495,
-0.03993861749768257,
0.09879834949970245,
-0.06050233542919159,
-0.08042638748884201,
0.051995862275362015,
0.05547922104597092,
-0.06708502024412155,
-0.013992229476571083,
0.06414076685905457,
0.020186657086014748,
-0.007387829944491386,
0.0023316333536058664,
-0.017743345350027084,
0.06733996421098709,
0.05431007221341133,
-0.04758666083216667,
-0.01884438470005989,
-0.02421351708471775,
0.09908686578273773,
0.013827014714479446,
0.035755276679992676,
0.030781256034970284,
-0.06331565976142883,
0.0059447381645441055,
-0.0316803865134716,
-0.03204832598567009,
0.04753114655613899,
-0.0630754753947258,
-0.03511857986450195,
0.035500261932611465,
-0.0012281776871532202,
-0.026965977624058723,
0.05331956595182419,
0.05615238845348358,
-0.020078929141163826,
-0.11573508381843567,
-0.02465040609240532,
0.0038927181158214808,
-0.04273039475083351,
0.056339725852012634,
-0.02408922091126442,
0.10078047960996628,
-0.005518920253962278,
0.08842190355062485,
0.05683419480919838,
-0.008754526264965534,
0.0035100565291941166,
0.024357115849852562,
0.023487158119678497,
-0.039577506482601166,
0.06935921311378479,
-0.026910871267318726,
0.06322532147169113,
0.005058032926172018,
0.04237990826368332,
-0.023617148399353027,
-0.04057277366518974,
-0.06799198687076569,
-0.08104126155376434,
0.03775128349661827,
-0.07854309678077698,
-0.05218365043401718,
0.05219937860965729,
0.09235666692256927,
0.01677539199590683,
-0.08230220526456833,
0.009722305461764336,
-0.10412564873695374,
0.05370474234223366,
0.06675689667463303,
0.06871486455202103,
0.010928711853921413,
0.06706544756889343,
0.0022634915076196194,
-0.030661316588521004,
0.04361633583903313,
-0.01990831270813942,
0.04982542619109154,
-0.1027679294347763,
-0.050260305404663086,
-0.08656086772680283,
0.005014084279537201,
0.025438468903303146,
0.012001873925328255,
0.000991534674540162,
-0.027326589450240135,
0.047824833542108536,
0.02717752195894718,
-0.009237143211066723,
0.007267743349075317,
0.01993386261165142,
0.03260914608836174,
-0.005992910359054804,
-0.012305180542171001,
0.11326209455728531,
0.020063014701008797,
-0.0005186572670936584,
-0.028215652331709862,
-0.004077418707311153,
0.0550147108733654,
-0.017798643559217453,
0.009371211752295494,
0.019471347332000732,
-0.009038851596415043,
0.005784326232969761,
0.06147913634777069,
0.07886137068271637,
0.02262437716126442,
0.044720668345689774,
-0.009968833066523075,
-0.019830459728837013,
0.035333141684532166,
0.07377448678016663,
-0.0567987859249115,
-0.07395520806312561,
1.4949252060956514e-33,
-0.03838907182216644,
0.03107408992946148,
0.09253959357738495,
0.07298710197210312,
-0.038982462137937546,
-0.02841990813612938,
-0.00955851748585701,
0.013615507632493973,
0.06294643133878708,
0.061026472598314285,
-0.07782338559627533,
0.11232119053602219,
-0.04558180272579193,
0.05393817648291588,
-0.03436826542019844,
0.005022893194109201,
-0.10906403511762619,
-0.006957561708986759,
-0.07515240460634232,
0.0004993491456843913,
0.08955322951078415,
0.01611296646296978,
-0.0162749532610178,
-0.00839143618941307,
0.005511751398444176,
-0.0329391285777092,
-0.022792400792241096,
-0.07167181372642517,
-0.0004539914079941809,
-0.006352232303470373,
-0.14172481000423431,
0.04483996331691742,
0.010530729778110981,
-0.060376305133104324,
0.02287977933883667,
-0.04343871399760246,
0.027415316551923752,
-0.028358444571495056,
-0.07139763981103897,
0.005984735209494829,
0.02090856060385704,
0.07691758126020432,
-0.026421086862683296,
-0.03265603631734848,
-0.047665029764175415,
-0.07176531106233597,
0.03098447434604168,
0.07597639411687851,
-0.014802731573581696,
-0.037715185433626175,
-0.004865375347435474,
0.04773670434951782,
-0.011033156886696815,
-0.15191175043582916,
0.028024541214108467,
-0.0011265386128798127,
0.007876166142523289,
-0.00591127248480916,
0.10110661387443542,
0.002676035277545452,
0.013248040340840816,
-0.007482344750314951,
-0.01219017244875431,
0.061035219579935074,
0.06221102178096771,
0.08456012606620789,
0.02832747995853424,
0.06289797276258469,
0.0455453023314476,
0.0056924717500805855,
-0.08384552597999573,
0.08254110813140869,
-0.03146553784608841,
-0.03611593320965767,
-0.03554081916809082,
-0.011316976509988308,
0.095613032579422,
-0.051361083984375,
-0.019577814266085625,
-0.022964810952544212,
0.031379472464323044,
-0.01677517220377922,
-0.0005248612724244595,
-0.03909067437052727,
-0.02801329828798771,
0.0038705463521182537,
0.011334870010614395,
-0.09179985523223877,
-0.07899381965398788,
-0.02034681662917137,
-0.0027198318857699633,
0.01008425559848547,
0.08419723063707352,
-0.05147029459476471,
-0.021028930321335793,
-2.070123244627798e-33,
-0.05040137097239494,
-0.017270753160119057,
-0.07783710211515427,
-0.00984436646103859,
0.04913678020238876,
-0.03628641366958618,
0.09038589149713516,
-0.03368770331144333,
-0.04567398503422737,
-0.005131937563419342,
-0.03023969568312168,
-0.07934477925300598,
0.02960321120917797,
0.026379575952887535,
0.016748886555433273,
0.021105393767356873,
-0.021239152178168297,
0.014036450535058975,
0.024763988330960274,
0.021081743761897087,
0.0212427768856287,
0.1317402720451355,
-0.03441409021615982,
-0.0452779158949852,
-0.07682732492685318,
0.06278501451015472,
-0.047988031059503555,
0.00954018160700798,
0.025380276143550873,
0.04547560214996338,
-0.011684821918606758,
0.00222413195297122,
-0.048769064247608185,
-0.03266551345586777,
-0.013134622946381569,
0.10467860102653503,
-0.0044783856719732285,
-0.0724332183599472,
-0.06511158496141434,
0.020846404135227203,
0.005825288128107786,
-0.004055030643939972,
-0.013506850227713585,
0.007376612164080143,
-0.014390912838280201,
-0.007375671062618494,
-0.05368945002555847,
-0.033720482140779495,
-0.04085835814476013,
0.06748874485492706,
-0.008127320557832718,
0.08479613810777664,
-0.13706053793430328,
-0.12594792246818542,
-0.04435722902417183,
-0.05945000797510147,
0.0013072277652099729,
-0.030694246292114258,
0.013968335464596748,
0.03370793163776398,
0.02076728269457817,
-0.03566422313451767,
-0.0355876088142395,
-0.08630245923995972,
-0.02459370531141758,
0.09178078174591064,
-0.004047445487231016,
-0.046453967690467834,
0.03853115439414978,
0.02244798094034195,
-0.0637076124548912,
-0.07491369545459747,
0.06434396654367447,
-0.05214732885360718,
0.07188503444194794,
-0.04665996506810188,
0.04580406844615936,
0.03731544315814972,
-0.016171181574463844,
0.03355566784739494,
0.01792060025036335,
-0.05055822432041168,
0.020378848537802696,
0.04752439633011818,
0.03678952530026436,
0.06481460481882095,
-0.025545185431838036,
0.04415551573038101,
0.013106713071465492,
0.013537829741835594,
-0.018993815407156944,
-0.02627376653254032,
-0.026109736412763596,
-0.004140886012464762,
-0.007580428849905729,
-4.9121890555170467e-8,
-0.10160509496927261,
0.0020405473187565804,
0.1071215271949768,
-0.001803387887775898,
0.006372760981321335,
-0.018667425960302353,
0.017393972724676132,
-0.023093516007065773,
-0.028936974704265594,
-0.04113321751356125,
0.05203424021601677,
-0.10274484008550644,
-0.0003020190342795104,
0.040622372180223465,
0.05617562681436539,
0.08027474582195282,
-0.03135690838098526,
0.059744905680418015,
-0.0032287552021443844,
0.02790737710893154,
0.03560195118188858,
0.019363008439540863,
0.006140822544693947,
-0.02374975010752678,
0.026482651010155678,
-0.005555097479373217,
-0.027750037610530853,
0.029005266726017,
0.016014017164707184,
-0.018991582095623016,
0.05891355499625206,
0.041110433638095856,
0.00037087113014422357,
-0.04953593760728836,
-0.09410789608955383,
-0.015851084142923355,
0.09917739033699036,
0.02171553112566471,
-0.047425419092178345,
0.023158133029937744,
-0.026289450004696846,
0.08063094317913055,
-0.03615216538310051,
0.0659840852022171,
0.10387825220823288,
0.03273734077811241,
0.06203114241361618,
-0.05741685628890991,
0.036893438547849655,
0.02778959646821022,
-0.05142386257648468,
0.014627806842327118,
0.02162148617208004,
0.14563514292240143,
0.10205771028995514,
0.01489490456879139,
0.04380377009510994,
-0.007833362556993961,
0.014580333605408669,
-0.04203721508383751,
-0.02682429738342762,
0.07691299170255661,
-0.08969448506832123,
-0.047714024782180786
] | https://github.com/metrics-lab/sim | 1 | 0 | 0 | 0 |
Why In-Context Learning Models are Good Few-Shot Learners? | https://openreview.net/forum?id=iLUcsecZJp | [
"Shiguang Wu",
"Yaqing Wang",
"Quanming Yao"
] | Poster | We explore in-context learning (ICL) models from a learning-to-learn perspective. Unlike studies that identify specific learning algorithms in ICL models, we compare ICL models with typical meta-learners to understand their superior performance. We theoretically prove the expressiveness of ICL models as learning algorithms and examine their learnability and generalizability.
Our findings show that ICL with transformers
can effectively construct data-dependent learning algorithms instead of directly follow existing ones
(including gradient-based, metric-based, and amortization-based meta-learners).
The construction of such learning algorithm is determined by the pre-training process, as a function fitting the training distribution, which raises generalizability as an important issue.
With above understanding, we propose strategies to transfer techniques for classical deep networks to meta-level to further improve ICL. As examples, we implement meta-level meta-learning for domain adaptability with limited data and meta-level curriculum learning for accelerated convergence during pre-training, demonstrating their empirical effectiveness. | In-Context Learning, Meta-Learning | null | 13,664 | null | [
-0.05352171137928963,
-0.0653427466750145,
0.019477328285574913,
0.06529216468334198,
0.10878583043813705,
-0.009832017123699188,
0.012277227826416492,
-0.009764957241714,
-0.0015161704504862428,
-0.05718464404344559,
-0.07160510122776031,
0.0009166455129161477,
0.011189733631908894,
0.08434613049030304,
0.021973533555865288,
-0.05036517605185509,
0.07021261006593704,
0.08748725056648254,
-0.1252395212650299,
-0.01914016529917717,
0.01928919367492199,
0.001362903625704348,
0.0603567436337471,
-0.0028063564095646143,
0.02065441571176052,
0.0160339567810297,
-0.01593712531030178,
-0.02902870625257492,
-0.009475511498749256,
0.008763029240071774,
0.057794634252786636,
-0.009652247652411461,
-0.0769403874874115,
0.015137595124542713,
-0.044914234429597855,
0.08004558831453323,
-0.009622718207538128,
0.0935089960694313,
-0.004309870768338442,
-0.006794142536818981,
0.018217457458376884,
-0.025628263130784035,
-0.0007700166897848248,
0.0039604404009878635,
0.05925920605659485,
0.009718109853565693,
-0.00845944695174694,
-0.06279664486646652,
-0.017731010913848877,
0.04880808666348457,
-0.12555652856826782,
-0.004853768739849329,
-0.07498607784509659,
0.09929054230451584,
0.02835591323673725,
0.06327272951602936,
0.024798467755317688,
-0.03104338049888611,
-0.04867272451519966,
-0.00105976068880409,
-0.03855392336845398,
-0.11025755107402802,
-0.07558148354291916,
0.012548514641821384,
-0.08247952908277512,
-0.003813415765762329,
-0.002239816589280963,
0.051836442202329636,
0.0030704704113304615,
-0.036889851093292236,
-0.017909061163663864,
0.09252021461725235,
-0.06798645108938217,
0.07236216217279434,
0.07331748306751251,
0.04292640462517738,
0.06805945932865143,
-0.0038490951992571354,
0.050655487924814224,
-0.08137071877717972,
0.0700688287615776,
0.08451443910598755,
0.009884090162813663,
-0.019247768446803093,
0.11418966948986053,
0.01449502445757389,
-0.01738191582262516,
0.002592365723103285,
-0.021463127806782722,
-0.014417738653719425,
-0.004861652851104736,
-0.02596266381442547,
-0.025001585483551025,
-0.007909118197858334,
0.05383140966296196,
-0.002643394749611616,
0.03028912842273712,
-0.10811897367238998,
-0.04930240288376808,
0.13273707032203674,
-0.012257891707122326,
-0.005983950104564428,
-0.038794420659542084,
-0.019689636304974556,
0.014939692802727222,
-0.02506900392472744,
0.004462980665266514,
0.09219802916049957,
-0.021271077916026115,
-0.08590003848075867,
0.08757640421390533,
0.01663157530128956,
0.04250844940543175,
-0.06479214876890182,
0.03645860403776169,
0.015233255922794342,
-0.02187778800725937,
-0.0777667760848999,
0.05406644567847252,
0.09427240490913391,
-0.08201408386230469,
0.028631439432501793,
-0.03207385540008545,
0.028235670179128647,
0.0038874747697263956,
-0.0961821898818016,
-0.0964074581861496,
8.517398803898054e-34,
0.040353141725063324,
0.05621323361992836,
0.09329132735729218,
0.04426780715584755,
0.006681329105049372,
-0.054417684674263,
0.022708704695105553,
0.015159493312239647,
0.008467195555567741,
0.02023635432124138,
0.061853423714637756,
0.06412168592214584,
-0.0958654060959816,
0.047322385013103485,
0.05713414028286934,
0.04111088439822197,
-0.08100529760122299,
-0.006016854662448168,
0.0225240308791399,
-0.008125534281134605,
0.06825482100248337,
-0.09242688864469528,
0.002591861644759774,
-0.026158731430768967,
-0.007187888491898775,
0.046475864946842194,
0.028225934132933617,
0.022919537499547005,
-0.023242373019456863,
0.03450322151184082,
-0.06282702088356018,
0.014367451891303062,
0.02198580652475357,
0.048559967428445816,
0.00652011064812541,
-0.008952075615525246,
-0.00035891178413294256,
-0.010571908205747604,
-0.018455492332577705,
-0.02438029646873474,
0.007479052059352398,
0.03543713316321373,
-0.006603730842471123,
-0.008898915722966194,
-0.13135741651058197,
-0.0704161524772644,
-0.012298251502215862,
-0.009332552552223206,
-0.05164546146988869,
-0.023153482005000114,
0.07506371289491653,
-0.04449383541941643,
-0.08466669917106628,
-0.0551268644630909,
0.01822853460907936,
0.05983947217464447,
0.002766759367659688,
0.06271413713693619,
0.0364513024687767,
0.0608057864010334,
0.051709458231925964,
-0.034959159791469574,
-0.04442315548658371,
0.13295236229896545,
-0.02167162112891674,
0.034394752234220505,
0.00623677670955658,
0.05848783627152443,
0.010674985125660896,
-0.05590709298849106,
-0.0498541034758091,
0.028842387720942497,
-0.06899981945753098,
-0.04049468785524368,
0.06435513496398926,
-0.013620619662106037,
0.030582694336771965,
-0.11933756619691849,
0.004220408387482166,
0.09993744641542435,
-0.0018628776306286454,
0.06718211621046066,
-0.013241451233625412,
0.035348281264305115,
-0.06106678023934364,
-0.02728787437081337,
-0.004465320613235235,
-0.07780293375253677,
0.005840838421136141,
-0.05382872372865677,
-0.08411825448274612,
-0.019073255360126495,
-0.03823632374405861,
-0.008590731769800186,
-0.02268165722489357,
-1.3822019100075174e-33,
0.0016352154780179262,
0.04058639705181122,
-0.04705340415239334,
0.06126847863197327,
-0.007644639816135168,
-0.0032252022065222263,
0.025618933141231537,
0.0330527238547802,
-0.029715796932578087,
-0.03354090824723244,
-0.025369251146912575,
-0.033585142344236374,
0.025051288306713104,
-0.01947863958775997,
-0.004680367652326822,
-0.024590307846665382,
-0.05703498795628548,
-0.06421332061290741,
-0.01764012686908245,
0.05469611659646034,
0.08341372013092041,
0.06584401428699493,
-0.08260795474052429,
0.04320666193962097,
-0.07032773643732071,
0.0017898426158353686,
-0.10694526880979538,
0.07923532277345657,
0.038548216223716736,
0.042834699153900146,
0.0051465826109051704,
-0.020424706861376762,
-0.01127452403306961,
-0.0581706203520298,
-0.06077425926923752,
0.0993775799870491,
0.026294950395822525,
-0.03342369198799133,
-0.08410807698965073,
0.09308267384767532,
0.06932275742292404,
-0.05722331628203392,
-0.06616261601448059,
-0.0050976951606571674,
0.03454391658306122,
-0.02531365491449833,
-0.004258448723703623,
-0.020206840708851814,
0.08403616398572922,
0.015400735661387444,
-0.06055940315127373,
-0.013585901819169521,
-0.0811324268579483,
-0.03467226400971413,
-0.06965862214565277,
-0.046559061855077744,
0.07799530774354935,
-0.01613183692097664,
0.10081110149621964,
0.012699796818196774,
-0.011627848260104656,
-0.07083197683095932,
-0.010723493993282318,
-0.01602969504892826,
-0.040148284286260605,
-0.03288084641098976,
-0.024309325963258743,
0.0031978965271264315,
0.030190790072083473,
0.05445103347301483,
0.00117291952483356,
0.027902644127607346,
-0.016622040420770645,
-0.07865680754184723,
-0.05014071986079216,
-0.011623644270002842,
0.043205611407756805,
-0.0002092067152261734,
-0.022339187562465668,
-0.008877567946910858,
0.03348863869905472,
-0.009344163350760937,
0.008388230577111244,
0.04052333906292915,
0.09033501893281937,
0.08854091912508011,
0.03206894174218178,
0.06088554486632347,
0.0013871756382286549,
-0.016176410019397736,
-0.03869745135307312,
0.027399767190217972,
-0.027377024292945862,
0.004710515029728413,
-0.07775300741195679,
-4.10600335953859e-8,
-0.11275062710046768,
0.0668676421046257,
0.017235934734344482,
0.021109189838171005,
0.01395636796951294,
0.04049401357769966,
-0.014064930379390717,
0.08045230060815811,
0.05350935831665993,
-0.007098423317074776,
0.034119755029678345,
0.023957032710313797,
0.015396270900964737,
-0.033199794590473175,
-0.04126863554120064,
0.10440932214260101,
0.06156764179468155,
0.004395004361867905,
0.027740931138396263,
0.002843499416485429,
0.018763381987810135,
0.053374242037534714,
0.07396616041660309,
-0.07869722694158554,
0.0684254914522171,
-0.14525271952152252,
-0.03171246498823166,
0.06460065394639969,
0.049239158630371094,
0.022627128288149834,
-0.08118034154176712,
0.015582132153213024,
0.04900430887937546,
-0.03349660709500313,
0.03270614147186279,
0.08004088699817657,
0.027839699760079384,
-0.051051199436187744,
-0.03835279121994972,
-0.11196671426296234,
-0.01758686639368534,
0.08646751940250397,
-0.04944944754242897,
0.007378438487648964,
-0.01598759926855564,
0.05509675666689873,
0.03233484923839569,
-0.03788537532091141,
-0.01982327550649643,
0.015500892885029316,
0.006913331802934408,
-0.008996053598821163,
-0.019419126212596893,
0.003651553997769952,
0.07817354053258896,
0.025242069736123085,
-0.01983608491718769,
-0.06371589750051498,
0.01120197493582964,
0.061550743877887726,
-0.035838332027196884,
0.03077612817287445,
-0.056104548275470734,
-0.00068341672886163
] | 0 | 0 | 0 | 0 |
|
Evaluating Large Language Models through Role-Guide and Self-Reflection: A Comparative Study | https://openreview.net/forum?id=E36NHwe7Zc | [
"Lili Zhao",
"Yang Wang",
"Qi Liu",
"Mengyun Wang",
"Wei Chen",
"Zhichao Sheng",
"Shijin Wang"
] | Poster | Large Language Models fine-tuned with Reinforcement Learning from Human Feedback (RLHF-LLMs) can over-rely on aligned preferences without truly gaining self-knowledge, leading to hallucination and biases. If an LLM can better access its knowledge and know what it knows, it can avoid making false or unsupported claims. Therefore, it is crucial to evaluate whether LLMs have the ability to know what they know, as it can help to ensure accuracy and faithfulness in real-world applications. Inspired by research in Educational Psychology, surface learners who don’t really know are easily affected by teacher and peer guidance, we treat LLM as a student, incorporate role guidance in prompts to explore whether LLMs really know. Specifically, we propose a novel strategy called Role-Guided and Self-Reflection (RoSe) to fully assess whether LLM “knows it knows”. We introduce multiple combinations of different roles and strong reminder in prompts combined with self-reflection to explore what local information in prompt LLMs rely on and whether LLMs remain unaffected by external guidance with varying roles. Our findings reveal that LLMs are very sensitive to the strong reminder information. Role guidance can help LLMs reduce their reliance on strong reminder. Meanwhile, LLMs tend to trust the role of authority more when guided by different roles. Following these findings, we propose a double-calibrated strategy with verbalized confidence to extract well-calibrated data from closed-source LLM and fine-tune open-source LLMs. Extensive experiments conducted on fine-tuning open-source LLMs demonstrate the effectiveness of double-calibrated strategy in mitigating the reliance of LLMs on local information. For a thorough comparison, we not only employ public JEC-QA and openBookQA datasets, but also construct EG-QA which contains English Grammar multiple-choice question-answering and 14 key knowledge points for assessing self-knowledge and logical reasoning. | LLMs, Verbalized confidence, Shortcut learning | Evaluating large language models through role-guide and self-reflection strategy to make comparative study | 13,663 | null | [
0.0019853729754686356,
-0.029286781325936317,
0.061978936195373535,
0.0328221395611763,
0.04294809699058533,
0.001242262776941061,
0.04119503125548363,
0.014220443554222584,
0.01159283984452486,
-0.003954101819545031,
-0.10997200012207031,
-0.022426938638091087,
0.00914820097386837,
0.03165561333298683,
0.06692885607481003,
0.0007049139821901917,
0.08215931057929993,
0.031585708260536194,
-0.07331380248069763,
-0.047794632613658905,
0.026333585381507874,
-0.03663814440369606,
0.04973149299621582,
-0.0007864769431762397,
-0.04632418975234032,
-0.025456685572862625,
-0.02527640387415886,
-0.03647078946232796,
-0.0018447303446009755,
-0.05398689955472946,
0.0576360858976841,
0.02287508174777031,
-0.01561553031206131,
-0.009402420371770859,
-0.04647265747189522,
0.06339041888713837,
-0.010448535904288292,
0.051718540489673615,
0.015010942704975605,
-0.07597372680902481,
-0.037553075700998306,
0.018665282055735588,
0.042204760015010834,
0.005651839077472687,
0.09007789194583893,
-0.05950630456209183,
0.032793089747428894,
0.005040805786848068,
-0.08860306441783905,
-0.002287796465680003,
-0.056791581213474274,
-0.10065720230340958,
-0.011004049330949783,
-0.008562883362174034,
0.006692009046673775,
0.09633634984493256,
0.05508654564619064,
0.0676996260881424,
-0.04692591354250908,
-0.037246569991111755,
-0.04245203360915184,
-0.11786560714244843,
-0.08685518056154251,
0.015920130535960197,
-0.09521307796239853,
0.012858766131103039,
-0.0502634234726429,
0.09178207069635391,
-0.020313723012804985,
0.012757125310599804,
0.0039161876775324345,
0.006038335617631674,
0.08642914891242981,
0.012428163550794125,
0.09252485632896423,
-0.011065410450100899,
-0.008128714747726917,
-0.047630585730075836,
0.03756754845380783,
-0.03114580363035202,
-0.04972787946462631,
0.028081966564059258,
-0.007364673539996147,
0.01756417751312256,
0.01683400385081768,
-0.012714104726910591,
0.06123563274741173,
0.037054043263196945,
0.011615450493991375,
0.0005713885766454041,
0.03294968605041504,
-0.04831952229142189,
-0.05217910185456276,
0.03181043639779091,
-0.004339586943387985,
0.04618280753493309,
0.008636634796857834,
-0.08143120259046555,
-0.08187645673751831,
0.06737068295478821,
0.045402441173791885,
0.0720374658703804,
-0.031833380460739136,
-0.008059192448854446,
-0.09005233645439148,
-0.058276932686567307,
-0.04318945109844208,
0.06182601675391197,
0.05161589756608009,
-0.05209709703922272,
0.0017528432654216886,
0.011359740048646927,
0.010753476992249489,
0.09663928300142288,
0.028342584148049355,
-0.02380584180355072,
0.04070854187011719,
-0.0008730294066481292,
0.003260384313762188,
-0.01076452061533928,
0.047936681658029556,
0.0059071979485452175,
0.12605270743370056,
0.02223191224038601,
0.02698751725256443,
-0.03000427782535553,
-0.03145580738782883,
-2.745094132730308e-35,
0.07172699272632599,
0.0923231989145279,
0.01461600512266159,
0.0866442322731018,
0.026391619816422462,
-0.028020035475492477,
0.055125996470451355,
0.009899095632135868,
-0.00626772828400135,
-0.00046723574632778764,
0.02443913370370865,
0.027489349246025085,
0.02569415792822838,
0.04123501479625702,
0.0005945058655925095,
0.12101805210113525,
-0.050578027963638306,
0.047475408762693405,
-0.06773611903190613,
-0.04206792637705803,
0.07448403537273407,
0.053583208471536636,
0.03647644445300102,
-0.14312049746513367,
0.03838244825601578,
-0.06745792180299759,
0.008315199986100197,
0.01355462521314621,
0.017197487875819206,
0.05664394423365593,
0.008505216799676418,
-0.02827298641204834,
-0.005283435340970755,
0.02218022570014,
-0.0327322781085968,
-0.012157137505710125,
-0.08110760897397995,
-0.08374275267124176,
0.018673444166779518,
-0.07383587956428528,
0.04254889115691185,
-0.03944516181945801,
0.05672675371170044,
0.06794823706150055,
-0.08874022215604782,
-0.10160014778375626,
-0.034702666103839874,
0.016472535207867622,
-0.08739344775676727,
0.015116414986550808,
0.005858362652361393,
0.03033611923456192,
-0.04609635844826698,
-0.08252911269664764,
-0.01603093184530735,
0.01693127304315567,
-0.015806443989276886,
0.07550211250782013,
-0.060491759330034256,
-0.046719055622816086,
0.03091091476380825,
0.04805528372526169,
-0.018907658755779266,
0.08035002648830414,
0.027496708557009697,
0.007090030703693628,
-0.03177898749709129,
-0.017636919394135475,
0.08579012751579285,
-0.05996060371398926,
-0.10937704145908356,
0.0036468282341957092,
-0.033428870141506195,
-0.017440292984247208,
-0.01641542837023735,
-0.12078311294317245,
0.010623467154800892,
-0.06059384346008301,
0.11590126901865005,
0.010935109108686447,
0.09495285898447037,
-0.043227195739746094,
0.020901678130030632,
-0.028517166152596474,
-0.04190700501203537,
-0.10751603543758392,
0.05403446406126022,
-0.057309892028570175,
-0.019621698185801506,
0.006181583274155855,
-0.034217823296785355,
-0.03768142685294151,
-0.022467803210020065,
0.07485950738191605,
-0.04417785629630089,
-2.0141878969568964e-33,
0.04047233983874321,
-0.04632838815450668,
0.030345333740115166,
0.02252449467778206,
-0.018017603084445,
-0.053303178399801254,
-0.014943580143153667,
0.0136572802439332,
-0.029758770018815994,
-0.0844176858663559,
-0.062448155134916306,
-0.004835902247577906,
0.04139094427227974,
0.09823399782180786,
-0.03549044206738472,
-0.0503951795399189,
0.03150496259331703,
0.00497922208160162,
0.03049960546195507,
-0.025409111753106117,
-0.041338980197906494,
0.0576518215239048,
-0.04974230006337166,
0.10425722599029541,
-0.04217283055186272,
-0.03027232363820076,
0.010799739509820938,
0.008963070809841156,
0.007504360284656286,
0.04385422170162201,
0.03413617983460426,
-0.04065413028001785,
0.021786712110042572,
-0.07951710373163223,
-0.044955458492040634,
0.09428267925977707,
0.011727222241461277,
-0.01083787064999342,
-0.09615343809127808,
0.07422146201133728,
0.060451894998550415,
-0.006595676299184561,
-0.027015451341867447,
-0.10380789637565613,
-0.05227939039468765,
-0.028664182871580124,
0.0068378872238099575,
0.042376648634672165,
0.040046326816082,
0.03551774099469185,
-0.0013435420114547014,
-0.043931182473897934,
0.05200531706213951,
0.01785573549568653,
0.002475342946127057,
-0.03124392032623291,
0.04747992008924484,
-0.027848532423377037,
0.12166588753461838,
-0.018466560170054436,
-0.044593945145606995,
0.00604853592813015,
0.02101803757250309,
-0.010624034330248833,
0.008971196599304676,
-0.04001394659280777,
-0.05106443911790848,
0.018978215754032135,
0.06279487907886505,
-0.02490321919322014,
0.06976451724767685,
-0.0511133074760437,
-0.020772326737642288,
-0.08067499846220016,
0.048972807824611664,
-0.00009780649997992441,
-0.0496085099875927,
-0.0675797238945961,
-0.1207534521818161,
-0.051435694098472595,
-0.010278809815645218,
-0.01002406608313322,
0.037535496056079865,
0.05615302920341492,
0.05924440175294876,
0.028965022414922714,
0.019279489293694496,
0.13624350726604462,
0.03652438148856163,
-0.012722273357212543,
0.035298146307468414,
0.055039264261722565,
-0.05921463668346405,
0.017224298790097237,
-0.05167726054787636,
-4.7400273928133174e-8,
-0.0034835513215512037,
-0.0032337105367332697,
0.04241955652832985,
0.09946025162935257,
0.04145587608218193,
0.011377065442502499,
-0.025600682944059372,
-0.024177860468626022,
-0.01763276569545269,
0.00019896581943612546,
0.01689734309911728,
0.010048563592135906,
-0.03824161738157272,
-0.07303965836763382,
0.06620359420776367,
0.04968781769275665,
0.01034900825470686,
0.011963206343352795,
-0.03840077295899391,
-0.06466590613126755,
0.0624917708337307,
-0.07074946910142899,
-0.005340693052858114,
-0.0803651362657547,
0.057276323437690735,
-0.015055358409881592,
-0.057339414954185486,
0.0916595607995987,
-0.025554563850164413,
0.04459515959024429,
0.05098869651556015,
-0.06248828396201134,
-0.015169982798397541,
-0.006747218314558268,
0.028371037915349007,
0.09138566255569458,
0.008910184726119041,
-0.06391328573226929,
0.08412197232246399,
0.03621835634112358,
0.025653336197137833,
0.08194773644208908,
-0.0030001522973179817,
0.11031549423933029,
-0.015669237822294235,
-0.06522141396999359,
-0.01988334394991398,
-0.09486681967973709,
-0.026665646582841873,
-0.041981201618909836,
0.0020252359099686146,
-0.09054228663444519,
-0.01824258267879486,
0.03113868460059166,
0.05847176909446716,
0.07314544916152954,
-0.017663728445768356,
0.01055161189287901,
-0.012483674101531506,
0.0355440117418766,
0.09481202065944672,
0.049809519201517105,
-0.07997186481952667,
0.03937570005655289
] | 0 | 0 | 0 | 0 |
|
Local Loss Optimization in the Infinite Width: Stable Parameterization of Predictive Coding Networks and Target Propagation | https://openreview.net/forum?id=g6syfIrVuS | [
"Satoki Ishikawa",
"Rio Yokota",
"Ryo Karakida"
] | Poster | Local learning, which trains a network through layer-wise local targets and losses, has been studied as an alternative to backpropagation (BP) in neural computation. However, its algorithms often become more complex or require additional hyperparameters due to the locality, making it challenging to identify desirable settings where the algorithm progresses in a stable manner.
To provide theoretical and quantitative insights, we introduce maximal update parameterization ($\mu$P) in the infinite-width limit for two representative designs of local targets: predictive coding (PC) and target propagation (TP). We verify that $\mu$P enables hyperparameter transfer across models of different widths.
Furthermore, our analysis reveals unique and intriguing properties of $\mu$P that are not present in conventional BP. By analyzing deep linear networks, we find that PC's gradients interpolate between first-order and Gauss-Newton-like gradients, depending on the parameterization.
We demonstrate that, in specific standard settings, PC in the infinite-width limit behaves more similarly to the first-order gradient.
For TP, even with the standard scaling of the last layer differing from classical $\mu$P, its local loss optimization favors the feature learning regime over the kernel regime. | deep learning, feature learning, local learning, predictive coding, target propagation, infinite width, maximal update parameterization (muP) | We derive the parameterization of major local learning methods that enable feature learning in infinite-width neural networks and demonstrate its benefits. | 13,642 | 2411.02001 | [
-0.007425152231007814,
-0.09238986670970917,
0.0429253876209259,
0.0780433863401413,
0.09010292589664459,
0.024233335629105568,
-0.036919817328453064,
-0.0890142098069191,
-0.021516412496566772,
-0.04952627792954445,
-0.030944718047976494,
0.015208360739052296,
-0.003219759091734886,
-0.01192149706184864,
-0.08704297989606857,
0.028775226324796677,
0.09344536066055298,
0.09761453419923782,
-0.09901859611272812,
0.019030138850212097,
0.028013888746500015,
-0.0005676042637787759,
0.026432160288095474,
0.011601800099015236,
0.047623779624700546,
-0.0062329997308552265,
-0.011166254058480263,
-0.03713137283921242,
0.04120156541466713,
-0.01795809529721737,
0.017957305535674095,
-0.036913007497787476,
-0.08442404121160507,
0.017471618950366974,
-0.02944234386086464,
0.06879810243844986,
-0.046290937811136246,
0.03351772204041481,
0.04507048800587654,
0.011817396618425846,
0.02686101570725441,
0.03990788757801056,
-0.028180396184325218,
0.08105222880840302,
0.06649024784564972,
-0.024872323498129845,
0.07900336384773254,
-0.016371678560972214,
-0.04607329145073891,
-0.0575370118021965,
0.023761825636029243,
-0.01260231714695692,
-0.03990264981985092,
0.047342002391815186,
-0.02170456387102604,
-0.04161613807082176,
0.023398688063025475,
0.0822516530752182,
-0.03671218827366829,
0.047435369342565536,
-0.08187806606292725,
-0.09507366269826889,
-0.004128708969801664,
-0.0553324930369854,
0.015353469178080559,
-0.04653000086545944,
-0.011187423020601273,
0.028677649796009064,
-0.009911076165735722,
0.029555968940258026,
0.004529672674834728,
0.10534537583589554,
-0.06643374264240265,
0.04121842980384827,
0.0867132619023323,
0.06517200171947479,
0.10495827347040176,
0.03869493678212166,
-0.03852642700076103,
-0.03613893315196037,
0.041728269308805466,
-0.050145674496889114,
-0.04556338116526604,
-0.07630904018878937,
0.10272951424121857,
0.003762547392398119,
0.006140482146292925,
0.02241632342338562,
0.07396086305379868,
-0.0030792427714914083,
-0.040267594158649445,
-0.023692093789577484,
-0.03690513223409653,
-0.02550722286105156,
0.05142844468355179,
-0.07832731306552887,
-0.03988500311970711,
-0.09812077134847641,
-0.026486635208129883,
0.06617284566164017,
-0.04835807904601097,
-0.0015115387504920363,
0.05545276030898094,
0.01530370395630598,
0.07279669493436813,
-0.04062806814908981,
0.017240703105926514,
0.05329177528619766,
0.09053654968738556,
-0.06161915138363838,
0.021591734141111374,
0.04960808530449867,
0.019794950261712074,
0.02389329858124256,
0.036268990486860275,
-0.043835703283548355,
0.001921047456562519,
-0.008625384420156479,
-0.016082169488072395,
0.032533954828977585,
-0.11484234780073166,
-0.006960784085094929,
-0.1324547976255417,
-0.013197270222008228,
-0.013160530477762222,
-0.01424056850373745,
-0.07631435245275497,
2.1503299792492874e-33,
0.012777520343661308,
0.024531260132789612,
-0.01964392513036728,
-0.011101800948381424,
0.035560108721256256,
0.043135371059179306,
0.05367066711187363,
-0.018352873623371124,
0.021635381504893303,
-0.013240435160696507,
-0.048162978142499924,
-0.08845203369855881,
-0.05961974710226059,
0.14331784844398499,
0.08594097197055817,
0.000713723711669445,
-0.025252241641283035,
0.05203329026699066,
-0.005184200592339039,
-0.006903383415192366,
0.02337687648832798,
-0.0008609485812485218,
-0.06373085081577301,
-0.047331877052783966,
-0.09048189222812653,
-0.06458287686109543,
-0.001355987973511219,
0.01851728744804859,
-0.05578015372157097,
0.028645001351833344,
-0.07980230450630188,
-0.005867707077413797,
-0.02857705019414425,
0.018196381628513336,
-0.01514546200633049,
-0.012743009254336357,
-0.030316011980175972,
-0.054288577288389206,
0.03524782508611679,
0.021092673763632774,
-0.00021541392197832465,
0.03708898648619652,
0.0480460561811924,
-0.01890466921031475,
-0.06781709939241409,
-0.15662725269794464,
-0.03357004374265671,
0.05274627357721329,
-0.10629095882177353,
-0.05576949566602707,
-0.0009695332264527678,
0.0011722557246685028,
-0.04712904244661331,
-0.03428056091070175,
0.036513812839984894,
-0.05620637908577919,
0.028957216069102287,
-0.016897553578019142,
0.019738120958209038,
0.03298235312104225,
0.06487496942281723,
-0.023165730759501457,
-0.0029061443638056517,
0.055187076330184937,
0.0004679887497331947,
0.03671091049909592,
0.0036348358262330294,
0.009710409678518772,
0.06795630604028702,
0.0033965439070016146,
0.024698613211512566,
0.031211303547024727,
0.051658641546964645,
-0.048714932054281235,
0.034348733723163605,
0.00479670288041234,
-0.03707585111260414,
-0.05818355083465576,
0.04290751367807388,
0.03377201780676842,
-0.061575181782245636,
0.09550058096647263,
-0.04025643691420555,
0.045098163187503815,
-0.0346093624830246,
0.003334748325869441,
-0.007001889869570732,
-0.10998089611530304,
-0.08183930069208145,
-0.005635663401335478,
-0.11141685396432877,
-0.02854933962225914,
0.01466013677418232,
-0.05416138470172882,
-0.030093805864453316,
-2.6197529953326136e-33,
-0.05704456567764282,
0.05335765704512596,
-0.0280719343572855,
0.021522527560591698,
-0.06308968365192413,
-0.013393913395702839,
-0.09010292589664459,
0.03877279907464981,
-0.02366442233324051,
0.002026900416240096,
0.009777747094631195,
-0.07711148262023926,
-0.0591726116836071,
0.013070954941213131,
0.061333734542131424,
0.021493300795555115,
-0.008475723676383495,
0.06778944283723831,
0.016995590180158615,
-0.12150108069181442,
0.037990689277648926,
0.043231308460235596,
-0.04719759523868561,
0.07334998995065689,
0.0035658034030348063,
-0.048761460930109024,
-0.040163811296224594,
0.09456389397382736,
0.02866324782371521,
-0.03422529250383377,
-0.08036118000745773,
0.006572755519300699,
-0.07099234312772751,
0.04424101486802101,
-0.031959131360054016,
0.0920281857252121,
0.01466762088239193,
0.029988493770360947,
0.012956315651535988,
0.07098924368619919,
0.01680399291217327,
-0.00044293043902143836,
0.029312176629900932,
0.05173084884881973,
0.017033684998750687,
-0.01798381842672825,
0.0021530012600123882,
-0.040246013551950455,
0.039586830884218216,
0.04181801900267601,
0.030108856037259102,
0.04485274478793144,
-0.005873562302440405,
0.008982491679489613,
-0.06940451264381409,
-0.003912684042006731,
-0.03749183937907219,
0.035173989832401276,
0.10791805386543274,
-0.0090577881783247,
-0.03487826883792877,
-0.05618546903133392,
0.000935710093472153,
-0.002314672339707613,
-0.09029505401849747,
0.04030309617519379,
0.018089428544044495,
0.045273683965206146,
0.04334833100438118,
-0.010371539741754532,
0.028402667492628098,
-0.001610519946552813,
0.012122332118451595,
0.06013505533337593,
-0.1073589026927948,
-0.03490143269300461,
-0.0030853678472340107,
-0.036751117557287216,
-0.03149246796965599,
-0.010928908362984657,
0.08303631097078323,
-0.01552648562937975,
-0.02475505694746971,
0.028737453743815422,
0.14278489351272583,
0.03525964915752411,
0.01984582096338272,
0.11022017151117325,
0.028566991910338402,
-0.020210538059473038,
0.01545098889619112,
0.02357582002878189,
0.033781152218580246,
-0.08379469066858292,
-0.03255335986614227,
-5.1720068228178206e-8,
-0.11351653933525085,
0.08359257876873016,
-0.05460560321807861,
0.06429851800203323,
-0.0065847355872392654,
0.0055167535319924355,
0.05112862214446068,
0.09426651149988174,
-0.00548881571739912,
0.00768701545894146,
0.07926255464553833,
0.022237692028284073,
0.03301470726728439,
-0.018238401040434837,
0.00391722097992897,
0.05828077718615532,
-0.02094968780875206,
-0.05040181800723076,
0.05739597603678703,
0.01174528431147337,
-0.01740814372897148,
0.0262716393917799,
-0.009590478613972664,
0.017933949828147888,
0.07222671806812286,
-0.10580211132764816,
-0.004207650199532509,
0.11385851353406906,
0.04993293061852455,
0.023102818056941032,
0.05111593380570412,
0.0053842258639633656,
0.020552173256874084,
0.031108656898140907,
-0.01837538368999958,
0.05050168186426163,
0.005650981795042753,
-0.01983746886253357,
-0.020874274894595146,
0.02294190786778927,
-0.0068114870227873325,
0.0002896337246056646,
0.040711816400289536,
-0.011077454313635826,
-0.004003855399787426,
-0.016860734671354294,
0.05088122561573982,
-0.03338431194424629,
-0.04619022086262703,
-0.019811924546957016,
0.11633885651826859,
0.06392426043748856,
-0.05002845823764801,
0.013078744523227215,
0.11850397288799286,
-0.051792651414871216,
-0.056291814893484116,
-0.08435069024562836,
0.01884467899799347,
0.10695018619298935,
-0.01223601121455431,
0.09405945986509323,
-0.08858932554721832,
0.01995515637099743
] | 0 | 0 | 0 | 0 |
|
Towards Faster Decentralized Stochastic Optimization with Communication Compression | https://openreview.net/forum?id=CMMpcs9prj | [
"Rustem Islamov",
"Yuan Gao",
"Sebastian U Stich"
] | Poster | Communication efficiency has garnered significant attention as it is considered the main bottleneck for large-scale decentralized Machine Learning applications in distributed and federated settings. In this regime, clients are restricted to transmitting small amounts of compressed information to their neighbors over a communication graph. Numerous endeavors have been made to address this challenging problem by developing algorithms with compressed communication for decentralized non-convex optimization problems. Despite considerable efforts, current theoretical understandings of the problem are still very limited, and existing algorithms all suffer from various limitations. In particular, these algorithms typically rely on strong, and often infeasible assumptions such as bounded data heterogeneity or require large batch access while failing to achieve linear speedup with the number of clients. In this paper, we introduce MoTEF, a novel approach that integrates communication compression with $\textbf{Mo}$mentum $\textbf{T}$racking and $\textbf{E}$rror $\textbf{F}$eedback. MoTEF is the first algorithm to achieve an asymptotic rate matching that of distributed SGD under arbitrary data heterogeneity, hence resolving a long-standing theoretical obstacle in decentralized optimization with compressed communication. We provide numerical experiments to validate our theoretical findings and confirm the practical superiority of MoTEF. | Optimization, Decentralized Learning, Federated Learning, Communication Compression | null | 13,638 | 2405.20114 | [
-0.08396818488836288,
0.03480612486600876,
-0.015317067503929138,
-0.021092893555760384,
0.029984481632709503,
-0.12733422219753265,
0.027073834091424942,
-0.07438036799430847,
-0.005915646441280842,
0.01021637488156557,
-0.017469823360443115,
0.06802653521299362,
0.050726331770420074,
0.0009920386364683509,
0.012965851463377476,
0.05129222944378853,
0.1244051530957222,
0.06679659336805344,
-0.0531652607023716,
0.010560259222984314,
0.0025588301941752434,
-0.036701545119285583,
-0.004510327707976103,
-0.0048675877042114735,
0.06356826424598694,
-0.05526323989033699,
-0.0006431448273360729,
-0.04218949377536774,
-0.028372956439852715,
0.014574919827282429,
0.08303669840097427,
-0.013201459310948849,
0.0025355673860758543,
-0.026985106989741325,
-0.03664539009332657,
0.08234983682632446,
0.022307123988866806,
0.004704474005848169,
-0.00971596036106348,
0.07961460947990417,
0.041916172951459885,
0.004075538832694292,
-0.03899295628070831,
0.020093148574233055,
0.0006857355474494398,
0.03649149835109711,
0.09313828498125076,
0.0334969200193882,
0.00989604089409113,
-0.05356261134147644,
-0.04199295490980148,
-0.017258472740650177,
-0.08087204396724701,
0.07528220117092133,
-0.018129967153072357,
-0.06435862183570862,
-0.021808404475450516,
0.0010394906857982278,
0.006726128049194813,
0.028732415288686752,
0.007303479593247175,
-0.061162762343883514,
0.0009531556279398501,
0.038953132927417755,
-0.03298046439886093,
-0.011524591594934464,
0.07326476275920868,
0.08039332926273346,
-0.05891670659184456,
0.03735743463039398,
0.034298378974199295,
-0.045580778270959854,
-0.04936761409044266,
0.07450266182422638,
0.020817775279283524,
0.04505343735218048,
0.0893038734793663,
0.018077075481414795,
0.06162578985095024,
0.003312569111585617,
-0.017758913338184357,
0.03277706727385521,
0.0066094365902245045,
-0.03907787799835205,
0.017243163660168648,
-0.1128014400601387,
0.004483248107135296,
-0.009954272769391537,
0.051211558282375336,
-0.07731093466281891,
-0.0382264144718647,
0.06856369972229004,
-0.07561903446912766,
0.013013793155550957,
-0.03178858011960983,
-0.04149029031395912,
0.06849358975887299,
-0.02561921440064907,
-0.04861036315560341,
0.15738476812839508,
-0.018634745851159096,
0.015319550409913063,
-0.03330413997173309,
-0.02586679719388485,
-0.03270935267210007,
-0.033728938549757004,
0.00959211029112339,
0.10402502119541168,
-0.010469933040440083,
-0.007117652799934149,
-0.0109748225659132,
0.04943978413939476,
0.02513755112886429,
0.03875907510519028,
0.0065015037544071674,
-0.03919339179992676,
-0.03227074816823006,
0.04320893809199333,
-0.03870750591158867,
0.0634324699640274,
-0.05625114217400551,
-0.03662468120455742,
-0.015570955350995064,
0.035746823996305466,
-0.0463513508439064,
0.08830608427524567,
-0.052105613052845,
4.643294153902209e-33,
0.04228268936276436,
0.07199519127607346,
0.05766841396689415,
-0.09026401489973068,
0.057422246783971786,
0.03681863471865654,
0.027766035869717598,
0.006718690972775221,
-0.03418607637286186,
0.014628897421061993,
-0.038046013563871384,
-0.06972997635602951,
0.001858468633145094,
0.021657686680555344,
0.005451477598398924,
-0.06119955703616142,
-0.033177800476551056,
-0.006032899953424931,
0.02876453474164009,
-0.10791122168302536,
0.07340160757303238,
-0.04207978397607803,
-0.027797291055321693,
0.012133796699345112,
0.04665166139602661,
-0.0773734301328659,
0.013061788864433765,
-0.09599393606185913,
0.09045698493719101,
0.03644949197769165,
-0.07420814037322998,
-0.022615883499383926,
0.0956328734755516,
0.0173021387308836,
0.03298050910234451,
-0.022038329392671585,
-0.02290222980082035,
-0.08209873735904694,
-0.05815902352333069,
-0.06887853890657425,
-0.018588466569781303,
-0.007680435664951801,
-0.02266368828713894,
-0.04595133662223816,
-0.11900048702955246,
-0.0626441091299057,
0.06451714783906937,
0.04474634677171707,
-0.015211858786642551,
-0.0902908593416214,
0.038511913269758224,
-0.017307722941040993,
-0.07255824655294418,
0.05655970796942711,
-0.008796514943242073,
0.01800062507390976,
0.0348028764128685,
0.02761843428015709,
0.07461801916360855,
0.09830118715763092,
-0.02165592834353447,
-0.052581533789634705,
0.010159718804061413,
0.0012401409912854433,
0.04622313007712364,
0.0013662411365658045,
-0.07344600558280945,
0.02872850000858307,
0.0943506583571434,
-0.06778029352426529,
0.023276614025235176,
0.01327509619295597,
0.07778412848711014,
-0.0308303814381361,
0.07402832806110382,
-0.02805330604314804,
0.08846546709537506,
-0.01994595117866993,
-0.052663370966911316,
0.004483056254684925,
-0.06754019111394882,
0.016596471890807152,
-0.044876083731651306,
-0.04437403753399849,
-0.02236996218562126,
-0.06041767820715904,
-0.027909833937883377,
-0.03376662731170654,
-0.024000810459256172,
0.014970175921916962,
-0.09600730985403061,
-0.01327001117169857,
-0.03220537304878235,
-0.032965660095214844,
0.03234381973743439,
-3.39400557381316e-33,
-0.03440062329173088,
0.02801058627665043,
0.004603814333677292,
0.03875519335269928,
0.07702364027500153,
-0.06947596371173859,
-0.06241718307137489,
-0.06739102303981781,
-0.0746462345123291,
-0.052306726574897766,
-0.11686902493238449,
-0.0711231604218483,
0.04326235130429268,
-0.04328056052327156,
-0.008136690594255924,
-0.02228502929210663,
0.028941022232174873,
0.051861245185136795,
0.0011391745647415519,
0.04399196803569794,
-0.014820090495049953,
0.033925674855709076,
-0.04967048019170761,
0.020518876612186432,
0.12541118264198303,
0.053359996527433395,
-0.08662894368171692,
0.09375203400850296,
0.05814051628112793,
-0.0037775179371237755,
-0.04107111692428589,
-0.021697690710425377,
-0.12666626274585724,
-0.09970749169588089,
-0.0292096845805645,
0.062079865485429764,
-0.005825731437653303,
0.11820670962333679,
-0.04794260486960411,
0.11440110206604004,
0.0755801871418953,
-0.02861190214753151,
-0.10285399854183197,
0.05440283194184303,
0.05817443132400513,
0.009197216480970383,
-0.08653471618890762,
-0.010928291827440262,
0.09121953696012497,
-0.009533902630209923,
0.01766095869243145,
0.022469621151685715,
-0.060886137187480927,
0.038583677262067795,
0.011989949271082878,
-0.00007669991464354098,
-0.03867647796869278,
0.04544427618384361,
0.04979871213436127,
0.015687720850110054,
-0.007555273827165365,
-0.026575863361358643,
-0.046320073306560516,
0.03536128252744675,
0.04257751628756523,
-0.05299338325858116,
0.04174622520804405,
0.02689971961081028,
0.04977736994624138,
0.06988818943500519,
-0.04087337478995323,
-0.0019243748392909765,
-0.02755201794207096,
-0.005641300231218338,
0.05424843728542328,
0.04124872013926506,
0.04701755940914154,
0.03496861830353737,
-0.06592625379562378,
0.0404529869556427,
0.05390181019902229,
0.08566509187221527,
0.026184789836406708,
-0.005669343750923872,
0.030043158680200577,
0.0633973553776741,
0.08353295177221298,
-0.0033074193634092808,
0.035954512655735016,
-0.04234990477561951,
-0.012929179705679417,
0.01592220738530159,
-0.010698765516281128,
0.03031756915152073,
0.006936538498848677,
-4.4995445591666794e-8,
-0.0522986575961113,
0.001436219783499837,
-0.017398720607161522,
-0.05552445724606514,
0.10509297996759415,
-0.007329077459871769,
0.01732608675956726,
0.014042975381016731,
0.01363570336252451,
-0.021231086924672127,
0.050544627010822296,
-0.033826883882284164,
-0.029095882549881935,
-0.053706638514995575,
-0.031688056886196136,
0.04652063176035881,
-0.03782705217599869,
-0.0983399823307991,
0.0007055645692162216,
-0.0012712223688140512,
-0.0011835475452244282,
-0.00786237046122551,
-0.0055009410716593266,
0.09984546899795532,
0.027113817632198334,
-0.09706679731607437,
0.03242260590195656,
0.010784948244690895,
0.04822830483317375,
0.056940220296382904,
-0.15436041355133057,
0.014580287970602512,
0.05695609375834465,
-0.04312121495604515,
0.041600070893764496,
0.07446621358394623,
-0.025383738800883293,
-0.0010705051245167851,
-0.02029426209628582,
-0.02500070072710514,
0.0606658011674881,
0.05584012717008591,
0.011159224435687065,
-0.04157143086194992,
0.04128776490688324,
-0.006591852754354477,
0.021093150600790977,
0.0482633113861084,
0.0025090964045375586,
-0.03711732476949692,
0.020185105502605438,
0.02836344577372074,
0.0004866349045187235,
0.062237728387117386,
0.05984247103333473,
-0.04995393007993698,
-0.009275291115045547,
-0.09452846646308899,
0.08159231394529343,
0.03475628048181534,
0.015365607105195522,
-0.02530834637582302,
-0.07435423135757446,
-0.017066186293959618
] | https://github.com/mlolab/motef | 2 | 0 | 0 | 0 |
Group-robust Sample Reweighting for Subpopulation Shifts via Influence Functions | https://openreview.net/forum?id=aQj9Ifxrl6 | [
"Rui Qiao",
"Zhaoxuan Wu",
"Jingtan Wang",
"Pang Wei Koh",
"Bryan Kian Hsiang Low"
] | Poster | Machine learning models often have uneven performance among subpopulations
(a.k.a., groups) in the data distributions. This poses a significant challenge for the
models to generalize when the proportions of the groups shift during deployment.
To improve robustness to such shifts, existing approaches have developed strategies
that train models or perform hyperparameter tuning using the group-labeled data
to minimize the worst-case loss over groups. However, a non-trivial amount of
high-quality labels is often required to obtain noticeable improvements. Given
the costliness of the labels, we propose to adopt a different paradigm to enhance
group label efficiency: utilizing the group-labeled data as a target set to optimize
the weights of other group-unlabeled data. We introduce Group-robust Sample
Reweighting (GSR), a two-stage approach that first learns the representations from
group-unlabeled data, and then tinkers the model by iteratively retraining its last
layer on the reweighted data using influence functions. Our GSR is theoretically
sound, practically lightweight, and effective in improving the robustness to sub-
population shifts. In particular, GSR outperforms the previous state-of-the-art
approaches that require the same amount or even more group labels. Our code is
available at https://github.com/qiaoruiyt/GSR. | distribution shift, subpopulation shift, spurious correlation, influence function, sample reweighting, data selection | We introduce Group-robust Sample Reweighting (GSR), which uses group-labeled data to guide the iterative retraining of the model its on group-unlabeled data reweighted using influence functions. | 13,637 | 2503.07315 | [
-0.08020087331533432,
-0.0768146961927414,
0.007337305694818497,
0.10831514745950699,
0.07158072292804718,
0.031702302396297455,
-0.004247075412422419,
-0.06118251755833626,
-0.008513242937624454,
-0.07998093217611313,
-0.011895892210304737,
-0.06774743646383286,
0.03955389931797981,
-0.022867223247885704,
0.0004502605297602713,
-0.024490388110280037,
0.15971407294273376,
0.0509808212518692,
-0.09583467245101929,
-0.046222880482673645,
-0.006827079225331545,
-0.025808341801166534,
0.040524084120988846,
0.04332299530506134,
0.05827554687857628,
-0.05213670805096626,
-0.05792785808444023,
0.01950024627149105,
0.04550208896398544,
-0.042985230684280396,
-0.02556423842906952,
0.04754641652107239,
-0.05116478353738785,
-0.046302247792482376,
-0.053828585892915726,
0.033600613474845886,
-0.04066113755106926,
0.04645669832825661,
0.05868355929851532,
0.05934930220246315,
0.01905529387295246,
0.03931186720728874,
0.014454868622124195,
0.012862573377788067,
-0.04350396990776062,
0.010684266686439514,
0.04591771215200424,
-0.0266555342823267,
-0.0612703301012516,
0.1153460219502449,
0.053593091666698456,
-0.04644409567117691,
-0.02068668231368065,
0.0667266994714737,
-0.034298885613679886,
-0.0027078993152827024,
0.0675763189792633,
-0.023821622133255005,
-0.011619770899415016,
0.03985302150249481,
-0.04533415287733078,
-0.06401003897190094,
0.019626310095191002,
-0.02085835672914982,
-0.0908806249499321,
-0.00022305453603621572,
0.04133690893650055,
0.05385465919971466,
-0.03708520904183388,
0.09033872187137604,
0.079742930829525,
0.09910174459218979,
-0.059870198369026184,
0.025243597105145454,
0.009443052113056183,
0.06949155777692795,
-0.0152528565376997,
-0.002775807399302721,
0.07575596868991852,
-0.07643073052167892,
0.05901753902435303,
0.01899556629359722,
-0.028042888268828392,
0.04754899442195892,
0.12729544937610626,
-0.059143662452697754,
0.018152175471186638,
-0.007299026008695364,
-0.017438825219869614,
0.06418842077255249,
-0.08351244777441025,
0.09674883633852005,
0.11052146553993225,
-0.010951600968837738,
-0.10012412816286087,
0.008131206035614014,
-0.00334205967374146,
-0.09089790284633636,
0.022432241588830948,
0.11382688581943512,
-0.0946853831410408,
0.11421991139650345,
0.05545928329229355,
-0.030633334070444107,
-0.03261807933449745,
-0.08224227279424667,
-0.03284376859664917,
0.04969388619065285,
0.058733779937028885,
-0.008540963754057884,
0.0220399908721447,
0.10057961940765381,
-0.017780574038624763,
0.0044584921561181545,
-0.005037711467593908,
-0.04227606579661369,
0.0661209300160408,
-0.0002746790705714375,
-0.0670846551656723,
0.052136775106191635,
-0.03542314097285271,
-0.030960526317358017,
-0.008034651167690754,
0.015560497529804707,
0.04693484306335449,
0.06314972043037415,
-0.08864961564540863,
6.2301884768027615e-34,
0.021498460322618484,
-0.02200084924697876,
0.05226478353142738,
0.00601380318403244,
0.061666395515203476,
-0.11620606482028961,
-0.07892242819070816,
-0.007291340734809637,
-0.014795971103012562,
0.019047515466809273,
-0.015490283258259296,
0.0018511125817894936,
-0.014418847858905792,
0.03662637621164322,
0.03832800313830376,
-0.007304069586098194,
-0.008937768638134003,
0.10580287873744965,
-0.078789621591568,
-0.060059867799282074,
-0.06528373807668686,
-0.029467908665537834,
0.01492067240178585,
-0.0630573257803917,
0.04086032882332802,
0.018720651045441628,
-0.03649644926190376,
0.007110150996595621,
0.01692485809326172,
0.007138163782656193,
-0.03548683971166611,
-0.041946250945329666,
0.05823851376771927,
-0.006790537387132645,
-0.013210204429924488,
-0.04169313237071037,
0.0019392581889405847,
-0.0620524138212204,
0.05619548261165619,
0.08356592804193497,
0.022275695577263832,
0.018708979710936546,
0.06303544342517853,
-0.04209638386964798,
0.014004471711814404,
0.00234807375818491,
-0.010171454399824142,
0.0034298812970519066,
-0.0525871105492115,
-0.03740643709897995,
0.009301498532295227,
-0.03611001744866371,
-0.018053095787763596,
-0.03350275382399559,
-0.07159938663244247,
-0.020801423117518425,
-0.007785600610077381,
0.005024098791182041,
-0.012331174686551094,
-0.004735057707875967,
0.0051748487167060375,
0.05160398408770561,
-0.020141521468758583,
0.08897769451141357,
0.08487749844789505,
0.0188896581530571,
0.014959285035729408,
0.04989183694124222,
0.006721112877130508,
0.01653984561562538,
0.04004643112421036,
0.050271011888980865,
-0.06081199645996094,
-0.020708104595541954,
0.06776861846446991,
0.038423433899879456,
0.04399393871426582,
-0.047635696828365326,
-0.01623786799609661,
0.03221980854868889,
0.011695032939314842,
-0.04022746905684471,
-0.0008936059894040227,
-0.03678956255316734,
-0.01031379122287035,
-0.02437077835202217,
0.025045862421393394,
-0.056802909821271896,
-0.03721307963132858,
0.01993616670370102,
-0.061992526054382324,
0.009659278206527233,
-0.06766515225172043,
-0.02349812164902687,
-0.011914820410311222,
-2.883728251567005e-34,
-0.0629565641283989,
0.05442187562584877,
0.02584957517683506,
0.014063230715692043,
0.0032328220549970865,
0.003115258179605007,
0.01640866883099079,
0.05269767716526985,
0.019063809886574745,
-0.00952848419547081,
-0.09155993163585663,
-0.04316454753279686,
0.04009683430194855,
-0.032775554805994034,
0.004307146649807692,
-0.021313782781362534,
-0.05236202850937843,
0.053062908351421356,
-0.010683711618185043,
-0.0025789865758270025,
0.030025972053408623,
0.05098038539290428,
-0.05541856214404106,
0.13058997690677643,
0.020636772736907005,
0.009236096404492855,
-0.061510100960731506,
0.09016720950603485,
0.044650837779045105,
-0.022125398740172386,
-0.07966696470975876,
-0.03943658992648125,
0.007011739537119865,
-0.11638064682483673,
-0.03434659168124199,
0.07684270292520523,
0.02426346205174923,
0.04431094601750374,
-0.053848251700401306,
0.10163096338510513,
-0.04296967759728432,
0.03382716700434685,
-0.07572776079177856,
0.0007531660376116633,
-0.03360116854310036,
0.0017859083600342274,
0.0038270961958914995,
-0.02591097168624401,
-0.018062451854348183,
-0.021883171051740646,
-0.07513213902711868,
0.030089275911450386,
-0.0967811793088913,
0.10468431562185287,
-0.035175323486328125,
-0.04815706983208656,
-0.05372963473200798,
-0.07410209625959396,
0.021805988624691963,
0.02419966831803322,
-0.10814154148101807,
0.03137814626097679,
-0.0017513192724436522,
-0.03518860042095184,
-0.0311266016215086,
-0.05120016634464264,
0.031645651906728745,
-0.0442374125123024,
0.020643871277570724,
-0.00897679291665554,
0.04292350634932518,
-0.027803651988506317,
0.036105960607528687,
-0.01029693242162466,
-0.08811849355697632,
-0.052858199924230576,
0.001483298372477293,
-0.03914276137948036,
-0.013332243077456951,
-0.02808016911149025,
0.10152172297239304,
-0.03685395419597626,
0.0038243515882641077,
0.06514837592840195,
0.002045562956482172,
0.062005456537008286,
0.0739026591181755,
0.0815432146191597,
0.03115689381957054,
0.014611070044338703,
-0.05960420146584511,
-0.01657327450811863,
0.007935473695397377,
0.05253006890416145,
-0.03310025483369827,
-4.136368758622666e-8,
-0.053329408168792725,
0.10080889612436295,
0.041173648089170456,
0.05852273479104042,
0.06374254822731018,
0.0055479505099356174,
-0.03707609698176384,
0.11754288524389267,
-0.00577834527939558,
0.0008817030466161668,
0.005258896853774786,
-0.009886350482702255,
0.005353420507162809,
-0.019975554198026657,
-0.041227906942367554,
0.09157960116863251,
0.03175879269838333,
0.07201962172985077,
-0.012437562458217144,
0.02663731947541237,
-0.0680788978934288,
0.019986312836408615,
0.09014905244112015,
-0.05348660424351692,
0.050398457795381546,
-0.12236690521240234,
-0.016664644703269005,
0.011371076107025146,
0.12095383554697037,
-0.03847550228238106,
-0.07403333485126495,
0.007928385399281979,
-0.006433373317122459,
0.0740092322230339,
-0.0018717159982770681,
0.06326763331890106,
-0.08070017397403717,
-0.014021743088960648,
0.02502022311091423,
0.020379124209284782,
-0.005388852674514055,
0.08516336977481842,
-0.0474420040845871,
0.037407539784908295,
0.01915295049548149,
-0.003539122873917222,
0.013795717619359493,
-0.06926535815000534,
-0.0010852698469534516,
-0.05945868417620659,
0.02596185728907585,
-0.09713543206453323,
-0.029872672632336617,
0.0518721342086792,
0.00946932565420866,
-0.014597278088331223,
-0.07914737612009048,
0.0011032555485144258,
0.0565422959625721,
0.057123638689517975,
0.01907070353627205,
-0.014447236433625221,
-0.06121839955449104,
-0.0027449384797364473
] | https://github.com/qiaoruiyt/gsr | 6 | 0 | 0 | 0 |
Endless Jailbreaks with Bijection Learning | https://openreview.net/forum?id=xP1radUi32 | [
"Brian R.Y. Huang",
"Maximilian Li",
"Leonard Tang"
] | Poster | Despite extensive safety measures, LLMs are vulnerable to adversarial inputs, or jailbreaks, which can elicit unsafe behaviors. In this work, we introduce bijection learning, a powerful attack algorithm which automatically fuzzes LLMs for safety vulnerabilities using randomly-generated encodings whose complexity can be tightly controlled. We leverage in-context learning to teach models bijective encodings, pass encoded queries to the model to bypass built-in safety mechanisms, and finally decode responses back into English. Our attack is extremely effective on a wide range of frontier language models. By controlling complexity parameters such as number of key-value mappings in the encodings, we find a close relationship between the capability level of the attacked LLM and the average complexity of the most effective bijection attacks. Our work highlights that new vulnerabilities in frontier models can emerge with scale: more capable models are more severely jailbroken by bijection attacks. | jailbreaking, redteaming, AI safety, AI alignment, adversarial robustness, adversarial attacks | We jailbreak frontier language models with a novel state-of-the-art encoding-based jailbreak, and we derive inverse scaling laws regarding the efficacy of our jailbreak. | 13,633 | 2410.01294 | [
0.01657455787062645,
-0.06422857195138931,
-0.026609372347593307,
-0.025852102786302567,
-0.004057788290083408,
0.020845701918005943,
0.0563499853014946,
-0.049597956240177155,
0.013586738146841526,
0.007247263565659523,
-0.026838529855012894,
-0.05680761858820915,
0.07572534680366516,
-0.0015758980298414826,
0.022507842630147934,
0.07110723853111267,
0.08132659643888474,
0.08493123948574066,
-0.06797434389591217,
0.006574107799679041,
0.03777360916137695,
0.046240344643592834,
-0.00425298186019063,
-0.015141069889068604,
-0.0015437895199283957,
-0.028486691415309906,
-0.06320298463106155,
-0.0140719935297966,
0.01885271817445755,
-0.03189074248075485,
0.011115310713648796,
0.007118565961718559,
0.007373744621872902,
0.015756772831082344,
-0.0151071697473526,
0.00003105041469098069,
-0.002676825737580657,
-0.02162904292345047,
0.03001396916806698,
-0.06274926662445068,
-0.042393479496240616,
0.06863295286893845,
0.04012460634112358,
-0.006520012393593788,
0.00479894457384944,
-0.06628314405679703,
-0.013548132963478565,
-0.009686020202934742,
-0.01620536483824253,
-0.05178714916110039,
-0.036266863346099854,
0.002801788505166769,
-0.034843601286411285,
0.028572412207722664,
-0.031192585825920105,
-0.0817827507853508,
0.004721015691757202,
0.0451853983104229,
-0.01412389986217022,
0.07213696092367172,
-0.07093989104032516,
-0.0011667382204905152,
-0.0008623410831205547,
0.02064589411020279,
0.0014192743692547083,
-0.05679117515683174,
-0.001906482269987464,
0.09086273610591888,
-0.0036850145552307367,
0.018652966246008873,
0.006158207543194294,
-0.013681967742741108,
-0.07922621816396713,
-0.01452919002622366,
0.05414067953824997,
0.09135835617780685,
-0.0010805749334394932,
-0.0330854132771492,
0.05683191493153572,
-0.08291932940483093,
0.01796519197523594,
-0.05496424809098244,
0.08064036816358566,
0.045080967247486115,
0.12282468378543854,
0.0408267080783844,
-0.026133179664611816,
0.02109847217798233,
0.08194469660520554,
0.06230740249156952,
-0.03816975653171539,
-0.011672343127429485,
0.12017598003149033,
0.07061344385147095,
0.10004761070013046,
-0.03280370682477951,
0.018399428576231003,
-0.08076917380094528,
-0.06736640632152557,
0.08660238236188889,
-0.03278083726763725,
0.08225757628679276,
0.022241363301873207,
-0.10191154479980469,
0.027177324518561363,
0.003869659034535289,
0.043212976306676865,
-0.017185501754283905,
0.0006179611664265394,
-0.06887542456388474,
0.04333483800292015,
0.07365468144416809,
0.06635567545890808,
-0.008665249682962894,
-0.004614047706127167,
-0.03203862905502319,
-0.013608911074697971,
-0.027269965037703514,
-0.02065211348235607,
0.05584634467959404,
-0.026704683899879456,
0.0880781039595604,
-0.04534530267119408,
-0.015378784388303757,
-0.05680534616112709,
-0.057006657123565674,
-0.037950921803712845,
1.8050591867509572e-33,
0.039771731942892075,
0.0480220802128315,
-0.02607298083603382,
0.06394007802009583,
0.03303270787000656,
-0.03255295380949974,
0.03193867951631546,
0.059478070586919785,
-0.031443074345588684,
0.0021887999027967453,
0.009823331609368324,
0.037717364728450775,
0.025211721658706665,
0.03218180686235428,
0.10174617916345596,
0.056893665343523026,
-0.024719243869185448,
-0.02600695751607418,
-0.010981501080095768,
-0.05295441672205925,
0.017044754698872566,
-0.04980581998825073,
-0.01460152119398117,
-0.07398099452257156,
0.025866476818919182,
0.054485224187374115,
0.002299422863870859,
0.011393614113330841,
-0.017944056540727615,
0.054000724107027054,
-0.12914519011974335,
0.017674554139375687,
-0.06073455139994621,
-0.014720200560986996,
0.05058320239186287,
-0.08946358412504196,
0.011267468333244324,
-0.00105069181881845,
-0.06539981812238693,
-0.042131438851356506,
-0.022154126316308975,
-0.024499436840415,
-0.04220827296376228,
0.03314409404993057,
0.029111772775650024,
-0.018668079748749733,
-0.05886823683977127,
-0.011318930424749851,
-0.06204192712903023,
-0.023721802979707718,
-0.00891821924597025,
-0.0073870448395609856,
-0.04367683082818985,
-0.02872626483440399,
-0.05429106205701828,
0.034358881413936615,
0.022074971348047256,
0.027207110077142715,
0.021641019731760025,
0.1620611548423767,
-0.07310813665390015,
0.042559247463941574,
0.03224914148449898,
0.05371272563934326,
0.03775867447257042,
-0.02381843514740467,
-0.01939273811876774,
-0.06374947726726532,
0.04122983664274216,
-0.020343853160738945,
-0.04306711256504059,
-0.012239455245435238,
-0.0634988397359848,
0.027403464540839195,
-0.06740950793027878,
-0.023100582882761955,
0.06953015178442001,
-0.07091695815324783,
0.011396883986890316,
-0.0071763209998607635,
0.05426614359021187,
-0.04407443106174469,
0.09394045174121857,
-0.003436430124565959,
-0.04347313567996025,
-0.07905244082212448,
0.0018065269105136395,
-0.10895933210849762,
-0.03646009415388107,
0.05524842441082001,
-0.06588460505008698,
-0.07264628261327744,
0.010184337384998798,
-0.018488574773073196,
0.005677308887243271,
-2.739133813879225e-33,
-0.05178258195519447,
-0.07643106579780579,
-0.10384781658649445,
0.046699367463588715,
-0.06691457331180573,
-0.11082347482442856,
0.0005453714984469116,
0.0649341270327568,
-0.014803471975028515,
-0.02854531817138195,
-0.1110682487487793,
0.002038830192759633,
0.12741994857788086,
-0.016836440190672874,
0.07095148414373398,
-0.06519951671361923,
-0.012081173248589039,
0.026470566168427467,
-0.007117654196918011,
0.08705764263868332,
0.02616356685757637,
0.0870814397931099,
-0.06533379852771759,
0.08946152776479721,
-0.01194197777658701,
0.03665844723582268,
-0.08330327272415161,
-0.003087975550442934,
0.05304166302084923,
0.10447404533624649,
-0.02732190303504467,
-0.03228558227419853,
0.019133422523736954,
-0.016767356544733047,
-0.10685213655233383,
0.04624180123209953,
0.11373879015445709,
-0.01160721480846405,
-0.032505299896001816,
0.044066693633794785,
0.07396484911441803,
-0.03549157828092575,
-0.10143016278743744,
-0.006330380216240883,
-0.010295290499925613,
-0.03191646561026573,
-0.006042038090527058,
-0.004152128007262945,
0.05744372308254242,
-0.030897395685315132,
0.058347463607788086,
-0.010042103007435799,
-0.004088548012077808,
0.007799169979989529,
-0.014228980988264084,
-0.12726140022277832,
0.01817874237895012,
-0.04537365585565567,
0.06124311685562134,
0.019710712134838104,
-0.03826361149549484,
-0.04794631525874138,
0.012898849323391914,
-0.046676598489284515,
0.046542439609766006,
-0.07790850847959518,
-0.07840394973754883,
0.008000367321074009,
-0.011599809862673283,
-0.0813882052898407,
0.09999623894691467,
-0.031177116557955742,
-0.041705869138240814,
0.016913732513785362,
-0.03643587976694107,
-0.014248848892748356,
-0.010327187366783619,
-0.0836627408862114,
-0.039343200623989105,
0.01656809262931347,
0.07596058398485184,
-0.006324864458292723,
0.045981328934431076,
0.100377656519413,
0.06840323656797409,
0.003098083892837167,
0.0921679362654686,
-0.016265954822301865,
0.012500612996518612,
-0.006796734873205423,
-0.022674893960356712,
0.03992113843560219,
-0.07564583420753479,
0.046803273260593414,
-0.02535158395767212,
-3.6844955531023516e-8,
-0.03041171282529831,
0.029032113030552864,
-0.03752864524722099,
0.06430632621049881,
0.04995325952768326,
0.055033326148986816,
-0.12285706400871277,
0.043737802654504776,
0.014023736119270325,
0.014361551031470299,
0.06698956340551376,
0.05089130997657776,
-0.02851833961904049,
-0.012976680882275105,
-0.03737263008952141,
0.1090148314833641,
-0.023343859240412712,
0.017559727653861046,
-0.02803395316004753,
0.007906015031039715,
-0.006718193180859089,
0.06134607642889023,
0.02522099018096924,
-0.053239669650793076,
0.01597769744694233,
-0.10838005691766739,
-0.04653672128915787,
0.02032523788511753,
0.10384204238653183,
0.11788629740476608,
-0.024807462468743324,
-0.05078864470124245,
-0.011361079290509224,
-0.016595739871263504,
-0.05512161925435066,
0.0953078344464302,
-0.02956770919263363,
-0.043958812952041626,
0.004698540084064007,
-0.046214692294597626,
0.04015412926673889,
-0.006085602100938559,
-0.050596803426742554,
0.0034571411088109016,
-0.06724418699741364,
-0.011637424118816853,
0.02393556945025921,
-0.062159955501556396,
0.06944815814495087,
0.013896491378545761,
0.04443775489926338,
-0.04396878927946091,
0.031294334679841995,
0.04740333929657936,
0.03321559727191925,
0.04044762998819351,
-0.03203839436173439,
-0.050931982696056366,
0.060588911175727844,
0.0882091373205185,
0.031622264534235,
0.024400796741247177,
0.04946010559797287,
-0.017344681546092033
] | 0 | 0 | 0 | 0 |
|
GotenNet: Rethinking Efficient 3D Equivariant Graph Neural Networks | https://openreview.net/forum?id=5wxCQDtbMo | [
"Sarp Aykent",
"Tian Xia"
] | Poster | Understanding complex three-dimensional (3D) structures of graphs is essential for accurately modeling various properties, yet many existing approaches struggle with fully capturing the intricate spatial relationships and symmetries inherent in such systems, especially in large-scale, dynamic molecular datasets. These methods often must balance trade-offs between expressiveness and computational efficiency, limiting their scalability. To address this gap, we propose a novel Geometric Tensor Network (GotenNet) that effectively models the geometric intricacies of 3D graphs while ensuring strict equivariance under the Euclidean group E(3). Our approach directly tackles the expressiveness-efficiency trade-off by leveraging effective geometric tensor representations without relying on irreducible representations or Clebsch-Gordan transforms, thereby reducing computational overhead. We introduce a unified structural embedding, incorporating geometry-aware tensor attention and hierarchical tensor refinement that iteratively updates edge representations through inner product operations on high-degree steerable features, allowing for flexible and efficient representations for various tasks. We evaluated models on QM9, rMD17, MD22, and Molecule3D datasets, where the proposed model consistently outperforms state-of-the-art methods in both scalar and high-degree property predictions, demonstrating exceptional robustness across diverse datasets, and establishes GotenNet as a versatile and scalable framework for 3D equivariant Graph Neural Networks. | graph neural networks, computational physics, 3D graphs | GotenNet: An efficient framework that uses high-degree steerable features to model complex 3D molecular structures while maintaining E(3) equivariance. | 13,629 | null | [
-0.05830257013440132,
-0.1340695470571518,
-0.0024940732400864363,
-0.02251659519970417,
0.03829960897564888,
-0.02174423448741436,
-0.07482638210058212,
-0.05131387338042259,
-0.013241445645689964,
-0.0562216155230999,
-0.036435291171073914,
-0.0504954494535923,
-0.020478803664445877,
0.15150436758995056,
-0.009611842222511768,
0.04429951310157776,
0.012263234704732895,
0.11347398161888123,
-0.08459441363811493,
0.036483656615018845,
-0.008604759350419044,
0.03369443491101265,
0.03546866029500961,
0.06647753715515137,
0.08664066344499588,
0.023668197914958,
0.0835380032658577,
-0.03789573907852173,
-0.008477067574858665,
0.027579370886087418,
0.07900968194007874,
-0.024993756785988808,
-0.03823093697428703,
-0.030983077362179756,
-0.05201594531536102,
0.007055426947772503,
-0.05462813004851341,
0.007878071628510952,
-0.04044213518500328,
-0.005383580457419157,
0.010745744220912457,
0.050862882286310196,
0.016210423782467842,
0.09888797253370285,
0.07106483727693558,
0.061375848948955536,
-0.009262403473258018,
-0.00617356738075614,
0.03303622454404831,
0.010671514086425304,
-0.025689393281936646,
-0.03932105749845505,
-0.07509442418813705,
-0.009030720219016075,
0.02284977026283741,
-0.00188559724483639,
-0.019247692078351974,
-0.03397650271654129,
-0.0171920545399189,
-0.022564109414815903,
0.08620362728834152,
-0.04532397910952568,
0.028796985745429993,
-0.020871400833129883,
-0.016867581754922867,
0.04162323474884033,
0.05352676287293434,
0.004666768945753574,
-0.018208811059594154,
0.055072929710149765,
-0.007882331497967243,
-0.0025491989217698574,
-0.12174102663993835,
-0.005653164815157652,
0.03748561069369316,
0.0794893428683281,
0.08563673496246338,
0.018195539712905884,
-0.002889821073040366,
-0.0055880616419017315,
0.00076152483234182,
0.05391406640410423,
0.0792485922574997,
0.02642812207341194,
0.011419102549552917,
0.019560661166906357,
-0.047275226563215256,
0.07194891571998596,
-0.01771211065351963,
-0.008257986046373844,
0.011260478757321835,
0.02330821566283703,
0.021902009844779968,
-0.08264362812042236,
-0.009517576545476913,
0.03845387324690819,
0.06685390323400497,
-0.03193534538149834,
-0.001904781674966216,
0.08284126222133636,
0.0072946310974657536,
0.0065220114775002,
0.007935216650366783,
-0.027229785919189453,
0.014889581128954887,
0.010006600990891457,
0.04508798569440842,
0.0679938942193985,
-0.045858319848775864,
-0.06425031274557114,
-0.022892622277140617,
0.05955527722835541,
-0.017170941457152367,
-0.016218606382608414,
-0.018292339518666267,
0.006770689971745014,
-0.008792352862656116,
-0.010338194668293,
0.11873408406972885,
0.10712055116891861,
-0.050589557737112045,
0.06738755851984024,
-0.01511253509670496,
0.04590592160820961,
0.011444107629358768,
0.039937909692525864,
-0.13398809731006622,
3.330144271810847e-33,
0.027291787788271904,
0.035437703132629395,
0.0829077884554863,
0.034511201083660126,
0.02051682583987713,
0.025448065251111984,
-0.039649598300457,
0.0025115502066910267,
-0.019540423527359962,
0.040433287620544434,
-0.13099448382854462,
0.011951998807489872,
0.04180897772312164,
0.07025111466646194,
-0.02954982966184616,
-0.033475544303655624,
-0.01079075038433075,
-0.05375109985470772,
-0.02391059324145317,
-0.0228935107588768,
-0.01923619955778122,
-0.016545787453651428,
0.015979144722223282,
0.05659104138612747,
-0.05359797552227974,
0.008320964872837067,
-0.038312479853630066,
0.05380388721823692,
-0.016363294795155525,
-0.020026342943310738,
-0.049247175455093384,
-0.004466138314455748,
-0.08702005445957184,
-0.009092044085264206,
-0.04336544871330261,
-0.06670975685119629,
0.0015820653643459082,
-0.09571957588195801,
0.054570019245147705,
-0.0936625674366951,
0.0028353515081107616,
0.03224185109138489,
-0.018866991624236107,
-0.04997606575489044,
-0.09193450212478638,
0.04036998003721237,
0.017908599227666855,
0.02157462202012539,
-0.042882826179265976,
-0.023860404267907143,
0.027288924902677536,
0.04412074759602547,
0.018002014607191086,
0.005278969183564186,
-0.017051734030246735,
-0.02777925878763199,
0.03842327743768692,
0.0567847341299057,
0.06068800762295723,
-0.007545454427599907,
-0.019654717296361923,
0.06269742548465729,
0.014096816070377827,
0.0004730331711471081,
-0.030807968229055405,
0.05281209573149681,
-0.054194994270801544,
0.09435450285673141,
0.07199130207300186,
0.013109703548252583,
-0.017570974305272102,
0.09837940335273743,
-0.02636708691716194,
-0.053499314934015274,
-0.0017655311385169625,
0.0037617934867739677,
-0.0283071119338274,
-0.13880622386932373,
0.02364196442067623,
0.08181940019130707,
-0.07953081279993057,
-0.0028394998516887426,
-0.0024539262522011995,
-0.027324628084897995,
-0.036469560116529465,
-0.11931236833333969,
0.031228696927428246,
-0.07566095143556595,
0.02864452637732029,
-0.00654413178563118,
-0.03436368331313133,
-0.10358577966690063,
0.017001353204250336,
-0.006967151537537575,
0.016021866351366043,
-2.9291100476904228e-33,
0.010390548035502434,
0.023340903222560883,
0.0033053881488740444,
0.010915549471974373,
0.029183687642216682,
-0.04713283106684685,
-0.05921095237135887,
0.005903902929276228,
-0.00427011726424098,
-0.0008273163693957031,
0.045472923666238785,
-0.03660915791988373,
0.024104012176394463,
-0.07030702382326126,
0.0897812619805336,
0.025993136689066887,
-0.06838060915470123,
-0.043674908578395844,
0.011695143766701221,
-0.026711707934737206,
0.032907355576753616,
-0.010709324851632118,
-0.07393602281808853,
0.04080755263566971,
-0.036136966198682785,
0.06380485743284225,
-0.010926387272775173,
-0.05736921355128288,
-0.023998575285077095,
0.030341364443302155,
-0.04665130376815796,
-0.07109973579645157,
-0.05373966321349144,
0.01832501031458378,
-0.026509752497076988,
0.044936180114746094,
0.06014031916856766,
0.0007717469125054777,
-0.010489072650671005,
-0.025493411347270012,
-0.030303074046969414,
-0.03972882404923439,
-0.06098124757409096,
0.11214525997638702,
0.01591266505420208,
0.09144680947065353,
-0.013939104042947292,
0.020964153110980988,
-0.025927690789103508,
0.012173270806670189,
-0.05006423220038414,
0.046563178300857544,
0.03437730669975281,
-0.006269638426601887,
0.013672934845089912,
-0.08668632805347443,
0.02196926437318325,
-0.021125605329871178,
0.09376703947782516,
-0.0045253983698785305,
-0.05730618163943291,
-0.11970355361700058,
-0.07770749181509018,
0.025540834292769432,
-0.07327128946781158,
-0.05322795361280441,
-0.021189486607909203,
-0.05811544507741928,
0.07217719405889511,
0.02706069126725197,
-0.02017389051616192,
-0.07322211563587189,
-0.020930001512169838,
-0.0005031440523453057,
0.032605722546577454,
-0.050221554934978485,
0.0457337312400341,
-0.009207826107740402,
-0.010212631896138191,
-0.03610634058713913,
0.08717192709445953,
-0.0666307583451271,
0.029130572453141212,
0.08593346923589706,
0.10070767253637314,
0.022478925064206123,
0.023623589426279068,
0.039379019290208817,
0.03211057558655739,
0.033624619245529175,
0.007880566641688347,
0.06014125794172287,
-0.06462642550468445,
0.14103169739246368,
0.10560375452041626,
-3.838539441858302e-8,
-0.06320592761039734,
0.015938829630613327,
0.06134762614965439,
-0.0749700665473938,
0.007524098735302687,
0.012599729001522064,
0.10260077565908432,
0.1555846929550171,
-0.06746314465999603,
0.046068184077739716,
0.06584953516721725,
-0.09075332432985306,
-0.027231240645051003,
-0.027196625247597694,
0.04386518523097038,
0.07825417816638947,
0.02294156700372696,
-0.003985732328146696,
0.031136106699705124,
0.05855869874358177,
-0.0455256924033165,
0.006501722149550915,
0.015831246972084045,
0.014197021722793579,
0.04118925333023071,
-0.09044154733419418,
-0.04199419170618057,
0.09091746807098389,
0.03466345742344856,
-0.012529687955975533,
-0.03281496837735176,
-0.0026007662527263165,
0.039964403957128525,
0.04925760254263878,
-0.09174548089504242,
0.007968191988766193,
0.018556341528892517,
-0.0533931590616703,
-0.07755109667778015,
-0.0007612102781422436,
-0.05234223231673241,
0.09241262823343277,
-0.02733135037124157,
-0.005656884517520666,
0.021889125928282738,
0.05043067783117294,
0.05401803180575371,
0.06481976807117462,
0.021724434569478035,
-0.04464617371559143,
-0.01836726814508438,
-0.04074384644627571,
-0.058950379490852356,
0.04399468004703522,
-0.011642918922007084,
0.025704937055706978,
-0.0886809304356575,
-0.02117869257926941,
0.05855952575802803,
-0.0059386733919382095,
-0.04225850850343704,
-0.02587912417948246,
-0.05002521723508835,
-0.11635152995586395
] | 0 | 0 | 0 | 0 |
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.