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
|
|---|---|---|---|---|---|---|---|---|---|
Toward Generalizing Visual Brain Decoding to Unseen Subjects | https://openreview.net/forum?id=At9JmGF3xy | [
"Xiangtao Kong",
"Kexin Huang",
"Ping Li",
"Lei Zhang"
] | Poster | Visual brain decoding aims to decode visual information from human brain activities. Despite the great progress, one critical limitation of current brain decoding research lies in the lack of generalization capability to unseen subjects. Prior work typically focuses on decoding brain activity of individuals based on the observation that different subjects exhibit different brain activities, while it remains unclear whether brain decoding can be generalized to unseen subjects. This study aims to answer this question. We first consolidate an image-fMRI dataset consisting of stimulus-image and fMRI-response pairs, involving 177 subjects in the movie-viewing task of the Human Connectome Project (HCP). This dataset allows us to investigate the brain decoding performance with the increase of participants. We then present a learning paradigm that applies uniform processing across all subjects, instead of employing different network heads or tokenizers for individuals as in previous methods, so that we can accommodate a large number of subjects to explore the generalization capability across different subjects. A series of experiments are conducted and we have the following findings. First, the network exhibits clear generalization capabilities with the increase of training subjects. Second, the generalization capability is common to popular network architectures (MLP, CNN and Transformer). Third, the generalization performance is affected by the similarity between subjects. Our findings reveal the inherent similarities in brain activities across individuals. With the emergence of larger and more comprehensive datasets, it is possible to train a brain decoding foundation model in the future. Codes and models can be found at https://github.com/Xiangtaokong/TGBD}{https://github.com/Xiangtaokong/TGBD. | Visual brain decoding, fMRI - image retrieval, Generalizing to unseen subjects | The study explores the generalization capability of visual brain decoding across unseen subjects, showing that the generalization can be enhanced by increasing the training subjects and affected by the subject similarity. | 1,426 | 2410.14445 | [
0.03591996058821678,
-0.10806650668382645,
0.06600933521986008,
0.024180883541703224,
0.03994531184434891,
0.022673197090625763,
0.0424572229385376,
-0.048521753400564194,
0.03381386026740074,
-0.07228768616914749,
-0.056361302733421326,
-0.027071965858340263,
-0.048563167452812195,
0.035819198936223984,
-0.017175698652863503,
-0.028319688513875008,
0.060982197523117065,
0.048640962690114975,
-0.07416143268346786,
-0.011281011626124382,
0.013806339353322983,
-0.053045421838760376,
0.04856158047914505,
-0.05689592659473419,
0.05454152449965477,
0.019716812297701836,
0.01272941567003727,
-0.03626420348882675,
-0.0029343024361878633,
0.00038481943192891777,
0.08528589457273483,
-0.004516247194260359,
0.009899089112877846,
0.04512966051697731,
-0.03829285874962807,
0.06522832810878754,
-0.03453295677900314,
0.03908982127904892,
0.03172655403614044,
-0.051497314125299454,
0.0018384975846856833,
0.038751885294914246,
0.056849606335163116,
-0.01177408266812563,
0.03949689492583275,
0.06607508659362793,
0.016087835654616356,
-0.1071958988904953,
-0.03950494900345802,
-0.09281416237354279,
-0.039242908358573914,
0.05321548506617546,
-0.07965313643217087,
0.0912674069404602,
0.018793633207678795,
0.0075652035884559155,
-0.037801727652549744,
0.020589347928762436,
-0.08388090878725052,
0.06315889954566956,
-0.088353231549263,
-0.04743602126836777,
0.00384799693711102,
-0.021649926900863647,
0.07825575023889542,
0.05261887237429619,
0.025998545810580254,
-0.011375612579286098,
-0.04107656702399254,
0.015931591391563416,
-0.008998285047709942,
0.043692804872989655,
-0.07804835587739944,
-0.023626647889614105,
0.023577623069286346,
0.10195022821426392,
-0.00007541146624134853,
-0.031049828976392746,
0.06382566690444946,
-0.10246749222278595,
0.04760095477104187,
0.05681714415550232,
0.00861816294491291,
0.009768168441951275,
0.11892355233430862,
0.008242860436439514,
-0.042003732174634933,
0.049946196377277374,
-0.06947636604309082,
0.07037346810102463,
-0.13365818560123444,
-0.04089466482400894,
0.0038714283145964146,
-0.0278641227632761,
0.02950260229408741,
0.004744144156575203,
0.0010719929123297334,
-0.05166392773389816,
0.019315339624881744,
0.08111152797937393,
-0.03077777661383152,
0.027498260140419006,
0.046325311064720154,
-0.03282395377755165,
-0.015816127881407738,
0.06734420359134674,
0.11268626153469086,
0.0571051687002182,
0.07045502960681915,
-0.022514302283525467,
0.013062158599495888,
0.023650066927075386,
-0.050500161945819855,
0.01784541644155979,
0.016472604125738144,
-0.03465358167886734,
0.04632621631026268,
0.052459292113780975,
0.0928235575556755,
0.06250664591789246,
0.01726566255092621,
-0.0212432611733675,
0.005197994410991669,
-0.023487020283937454,
0.004183499608188868,
-0.038664620369672775,
-0.12882544100284576,
3.1996728477352194e-33,
-0.040565814822912216,
-0.023552222177386284,
0.07360675185918808,
0.02841138280928135,
-0.029434125870466232,
0.02762185037136078,
-0.048683710396289825,
-0.03900826722383499,
0.02602253295481205,
0.07301948219537735,
-0.06358665972948074,
-0.016538109630346298,
-0.04921473562717438,
0.16456574201583862,
0.012285024859011173,
-0.01386539451777935,
-0.05161648243665695,
0.06653659790754318,
-0.020098790526390076,
-0.06307149678468704,
0.03725995495915413,
-0.007469106465578079,
-0.011290132999420166,
-0.05289120972156525,
0.016240809112787247,
-0.08233386278152466,
0.009782892651855946,
-0.10689903795719147,
-0.01790875755250454,
-0.02236570231616497,
-0.12850145995616913,
0.04145341366529465,
0.05394989624619484,
0.0032261074520647526,
-0.012312992475926876,
-0.0583261214196682,
0.05217773839831352,
-0.015884803608059883,
-0.014749646186828613,
0.028681235387921333,
0.021882211789488792,
0.04162244498729706,
-0.0357014536857605,
-0.011711216531693935,
-0.05552596598863602,
0.0057544950395822525,
0.00001347487068414921,
0.04460473731160164,
-0.07526250183582306,
-0.013918522745370865,
-0.01055880356580019,
-0.014242880046367645,
-0.046899013221263885,
-0.11123932898044586,
0.019741103053092957,
-0.010306321084499359,
0.013935948722064495,
0.019604366272687912,
0.10158240795135498,
-0.025485675781965256,
0.047926392406225204,
0.006164295598864555,
0.019252562895417213,
0.052716754376888275,
0.04102245718240738,
0.10432867705821991,
-0.01185949519276619,
-0.026346391066908836,
-0.0212120171636343,
-0.02971964329481125,
-0.07237696647644043,
0.10254096984863281,
-0.02423861436545849,
-0.11065346002578735,
0.02747936360538006,
0.005644918419420719,
0.06751375645399094,
-0.022050322964787483,
-0.024456188082695007,
0.03525589779019356,
0.0032990993931889534,
-0.05959074944257736,
-0.049171626567840576,
-0.015649842098355293,
-0.05777372419834137,
0.03230612352490425,
0.03562154248356819,
-0.07552582025527954,
-0.05549030378460884,
-0.03573326766490936,
0.013835210353136063,
0.0006078008445911109,
0.004738387651741505,
-0.031051496043801308,
-0.02620026096701622,
-3.1181045600476916e-33,
-0.07096484303474426,
0.03872489184141159,
-0.037808168679475784,
-0.039581988006830215,
0.04719507694244385,
-0.0715772733092308,
0.048123471438884735,
-0.055760253220796585,
-0.01954086869955063,
0.01325500663369894,
0.015248029492795467,
-0.10953538864850998,
0.015820126980543137,
-0.028495948761701584,
0.011159445159137249,
-0.03719426691532135,
0.024502858519554138,
0.026206111535429955,
0.027107151225209236,
0.044119153171777725,
-0.03141191974282265,
0.15382413566112518,
-0.07374761253595352,
-0.028137778863310814,
0.030230367556214333,
0.05650973320007324,
-0.038541585206985474,
0.09736214578151703,
0.005815199110656977,
-0.012031053192913532,
-0.05081915110349655,
0.035438377410173416,
-0.07411214709281921,
-0.027865078300237656,
-0.025050831958651543,
0.1497502326965332,
0.006748087704181671,
-0.015016240067780018,
-0.09577874094247818,
0.0740274116396904,
0.028493324294686317,
-0.034092262387275696,
-0.06139783933758736,
0.04688538238406181,
-0.001361010130494833,
-0.061361268162727356,
-0.05990941449999809,
-0.035397496074438095,
-0.028892068192362785,
0.06925678253173828,
0.013699537143111229,
0.08637972176074982,
-0.11244221031665802,
-0.06643584370613098,
-0.05726541578769684,
0.0032633806113153696,
0.055404867976903915,
-0.0004675035015679896,
0.06146235764026642,
-0.0023383686784654856,
0.01802009344100952,
-0.10102028399705887,
-0.02064460888504982,
0.011506248265504837,
-0.02284904010593891,
0.022758781909942627,
-0.04000715911388397,
0.008293353952467442,
-0.012019853107631207,
0.029459847137331963,
-0.017862964421510696,
-0.08694804459810257,
0.04420287162065506,
0.014904841780662537,
0.08675988018512726,
0.01949824020266533,
-0.006431129295378923,
0.03179997205734253,
-0.006804371252655983,
0.07115580886602402,
-0.010032843798398972,
-0.034857917577028275,
-0.03626818582415581,
0.06796894967556,
0.08530399948358536,
0.04135418310761452,
-0.046941231936216354,
0.044381458312273026,
0.012867651879787445,
-0.020497357472777367,
-0.03170832619071007,
-0.007450376637279987,
-0.04412677139043808,
0.00728716840967536,
-0.08243235945701599,
-4.569723088820865e-8,
-0.10686889290809631,
-0.04757588729262352,
-0.01698557659983635,
0.013338570483028889,
0.05917797610163689,
-0.028374718502163887,
-0.0655079036951065,
0.03788771852850914,
-0.043917059898376465,
-0.06121569871902466,
0.07819583266973495,
-0.033994823694229126,
-0.05789368599653244,
-0.0036860285326838493,
0.041025932878255844,
0.03123214840888977,
-0.046464260667562485,
0.02447134256362915,
-0.00874366145581007,
0.010049835778772831,
0.0379968099296093,
-0.00811759103089571,
-0.008839803747832775,
-0.03623175621032715,
0.013178559951484203,
-0.043061576783657074,
-0.031705040484666824,
0.051776278764009476,
-0.05925135686993599,
0.007453737314790487,
0.05237396061420441,
0.08396070450544357,
0.05521003529429436,
-0.04678146541118622,
0.006992572918534279,
0.02933281660079956,
0.033900704234838486,
0.044131603091955185,
-0.07894136756658554,
0.06078732758760452,
-0.00864422507584095,
-0.029199611395597458,
-0.0010502496734261513,
0.058418042957782745,
0.06125595420598984,
-0.006050897762179375,
0.07591061294078827,
-0.03535548970103264,
0.024874182417988777,
-0.030782774090766907,
0.05396363511681557,
0.0035386118106544018,
0.04045349732041359,
0.09488985687494278,
0.11074981838464737,
-0.031643789261579514,
0.04391985759139061,
0.016318000853061676,
0.006977465935051441,
0.04512336850166321,
0.018987935036420822,
0.08978479355573654,
-0.08555644750595093,
-0.024350494146347046
] |
THE ROBUSTNESS OF DIFFERENTIABLE CAUSAL DISCOVERY IN MISSPECIFIED SCENARIOS | https://openreview.net/forum?id=iaP7yHRq1l | [
"Huiyang Yi",
"Yanyan He",
"Duxin Chen",
"Mingyu Kang",
"He Wang",
"Wenwu Yu"
] | Poster | Causal discovery aims to learn causal relationships between variables from targeted data, making it a fundamental task in machine learning. However, causal discovery algorithms often rely on unverifiable causal assumptions, which are usually difficult to satisfy in real-world data, thereby limiting the broad application of causal discovery in practical scenarios. Inspired by these considerations, this work extensively benchmarks the empirical performance of various mainstream causal discovery algorithms, which assume i.i.d. data, under eight model assumption violations. Our experimental results show that differentiable causal discovery methods exhibit robustness under the metrics of Structural Hamming Distance and Structural Intervention Distance of the inferred graphs in commonly used challenging scenarios, except for scale variation. We also provide the theoretical explanations for the performance of differentiable causal discovery methods. Finally, our work aims to comprehensively benchmark the performance of recent differentiable causal discovery methods under model assumption violations, and provide the standard for reasonable evaluation of causal discovery, as well as to further promote its application in real-world scenarios. | Differentiable causal discovery, model assumption violations, benchmark | Causal discovery algorithms are typically based on untestable causal assumptions. We study the performance of several mainstream algorithms under model assumption violations and find that differentiable causal discovery exhibits robust performance. | 1,418 | null | [
-0.07977819442749023,
-0.023890145123004913,
0.010785642080008984,
0.08814401179552078,
0.09317117184400558,
-0.004778966307640076,
-0.0874820202589035,
-0.055163707584142685,
-0.03332555666565895,
0.04491181671619415,
-0.0067639658227562904,
-0.02811950258910656,
0.030990757048130035,
0.05466423183679581,
-0.06002333015203476,
-0.022293979302048683,
0.08958220481872559,
-0.0029668163042515516,
-0.04586271196603775,
0.021421322599053383,
-0.07932765036821365,
0.011457690969109535,
-0.01872909441590309,
0.051100946962833405,
0.023727240040898323,
-0.050651632249355316,
0.026305435225367546,
0.01785936765372753,
-0.022244030609726906,
0.011635617353022099,
-0.039369117468595505,
-0.014495342038571835,
-0.09827711433172226,
-0.09801516681909561,
0.006349194794893265,
0.07736564427614212,
0.04989057406783104,
0.02872767485678196,
0.01874700002372265,
-0.02439098432660103,
0.05318965017795563,
0.005528499372303486,
0.012317988090217113,
0.022572999820113182,
0.04604963958263397,
-0.011905389837920666,
-0.010441298596560955,
0.028439179062843323,
-0.03296688571572304,
-0.046090081334114075,
-0.1021772250533104,
-0.04876621812582016,
0.00820496678352356,
-0.008657719939947128,
0.03359581157565117,
-0.047522127628326416,
0.01121735479682684,
-0.006500221323221922,
0.028720494359731674,
0.03161321207880974,
-0.014400147832930088,
-0.10687509924173355,
-0.07340957224369049,
-0.04939989000558853,
0.015199805609881878,
0.02987811714410782,
-0.0009130710968747735,
0.05344361811876297,
-0.048595476895570755,
0.07614578306674957,
-0.012751269154250622,
0.1158553808927536,
-0.10860416293144226,
0.09057489782571793,
0.11254338175058365,
0.0538083016872406,
0.030917100608348846,
0.025718141347169876,
0.00906514935195446,
-0.1145569384098053,
0.009842303581535816,
0.020758481696248055,
-0.025187043473124504,
0.015275668352842331,
0.09143009781837463,
-0.0007203050190582871,
0.03435888886451721,
0.05871850252151489,
0.009845100343227386,
0.09771633893251419,
-0.010972980409860611,
-0.03491983562707901,
0.014372593723237514,
-0.022137043997645378,
0.02438800223171711,
-0.01926564797759056,
0.01613525114953518,
0.0007895971066318452,
0.057997927069664,
0.0721687525510788,
0.021383438259363174,
0.0970749482512474,
0.04304340109229088,
-0.0456269308924675,
0.020507263019680977,
0.018659181892871857,
-0.015286946669220924,
-0.013985683210194111,
0.026243316009640694,
-0.1218755841255188,
-0.027221042662858963,
0.030476916581392288,
0.07131218910217285,
-0.015644311904907227,
-0.006102690007537603,
-0.08986686170101166,
0.04100385308265686,
-0.06055188924074173,
-0.02047586441040039,
-0.009509374387562275,
0.03602586314082146,
0.03396681696176529,
0.021745389327406883,
0.02867879718542099,
-0.054838236421346664,
-0.010309981182217598,
-0.08488591015338898,
1.7891475850156313e-33,
0.11878905445337296,
-0.0035798782482743263,
0.08411820977926254,
0.028475550934672356,
-0.0038965982384979725,
-0.013078389689326286,
-0.10427702218294144,
-0.000825568858999759,
0.011495819315314293,
-0.004335883539170027,
0.005479138344526291,
0.011366114020347595,
-0.01924181543290615,
0.00519477017223835,
0.0781983956694603,
0.056917134672403336,
-0.02304772660136223,
0.09039494395256042,
-0.01370804663747549,
-0.022602729499340057,
0.04980643093585968,
-0.11717664450407028,
-0.05432761833071709,
-0.007216117810457945,
0.025723159313201904,
0.04619261249899864,
0.020417267456650734,
-0.005976911634206772,
-0.023394977673888206,
0.009162080474197865,
-0.041558507829904556,
-0.012565389275550842,
0.07147035747766495,
0.11216531693935394,
-0.0024977384600788355,
-0.010904324240982533,
-0.11077592521905899,
-0.05811792239546776,
-0.02374412678182125,
0.0226678978651762,
-0.010095118544995785,
-0.01693035289645195,
0.015764491632580757,
-0.011326774954795837,
-0.07160868495702744,
-0.023625627160072327,
-0.02618594281375408,
-0.06420092284679413,
-0.060026269406080246,
-0.07309284061193466,
0.05134217068552971,
0.05495832860469818,
0.016801167279481888,
-0.09903188049793243,
-0.003563696052879095,
0.05621012672781944,
-0.02791089005768299,
0.018781624734401703,
0.0019559767097234726,
0.03688701242208481,
0.010465381667017937,
0.012301815673708916,
-0.002248039934784174,
0.06383872777223587,
-0.04867168515920639,
0.0951780304312706,
-0.06731399148702621,
-0.022191859781742096,
-0.03326087445020676,
0.06335553526878357,
-0.012696490623056889,
0.03824273869395256,
-0.0518716424703598,
-0.07432296872138977,
0.04462168365716934,
-0.0208242479711771,
-0.02284209243953228,
-0.07225373387336731,
-0.012826076708734035,
-0.017106976360082626,
-0.0804067999124527,
-0.04448665305972099,
0.008666189387440681,
0.009554886259138584,
-0.08892671018838882,
-0.025871146470308304,
0.0048391083255410194,
-0.024838348850607872,
-0.06568634510040283,
-0.05359100550413132,
-0.06852775812149048,
0.02312205359339714,
-0.03769581392407417,
0.03358234837651253,
0.07620669156312943,
-3.572329903542824e-33,
-0.06539749354124069,
0.02164078690111637,
0.09860961884260178,
-0.018110530450940132,
-0.005149429198354483,
-0.02035183273255825,
0.008289203979074955,
-0.004402988124638796,
0.04662267491221428,
-0.01509881392121315,
0.022768212482333183,
-0.057338129729032516,
0.009256940335035324,
-0.0328993983566761,
0.008375684730708599,
0.0810447409749031,
0.016083260998129845,
0.024585716426372528,
-0.03215384855866432,
0.0025686270091682673,
0.04729807749390602,
0.038088053464889526,
-0.05132812634110451,
0.023840416222810745,
0.03985949233174324,
-0.03450235351920128,
-0.03133609890937805,
0.0015458136331290007,
-0.023807281628251076,
-0.009701850824058056,
-0.09888579696416855,
0.016248174011707306,
-0.026455318555235863,
-0.04900001734495163,
-0.04681592807173729,
0.13825079798698425,
0.06511489301919937,
-0.03161650896072388,
-0.05835598334670067,
0.014886545017361641,
0.013273540884256363,
0.06423153728246689,
-0.08346792310476303,
-0.008528496138751507,
0.051503945142030716,
0.04917266219854355,
-0.05060996487736702,
0.06309520453214645,
0.03654024004936218,
0.010639481246471405,
-0.0314192958176136,
0.06319042295217514,
0.06617667526006699,
0.07010748237371445,
-0.0833083838224411,
-0.06760053336620331,
0.02718508429825306,
-0.022997193038463593,
0.05762922391295433,
0.03712791949510574,
-0.13277854025363922,
-0.04208803176879883,
-0.030841661617159843,
0.05733117833733559,
-0.04250180348753929,
-0.09627102315425873,
-0.004366613924503326,
-0.05895376577973366,
0.020459257066249847,
0.01298869401216507,
0.0002347122790524736,
0.0004849485121667385,
-0.044963955879211426,
-0.07010926306247711,
-0.01667492464184761,
-0.016590530052781105,
0.010511472821235657,
-0.023338191211223602,
-0.09291944652795792,
0.02704545296728611,
0.12076067924499512,
0.0273246131837368,
0.07812128216028214,
0.009097428061068058,
0.04109190031886101,
0.05838580057024956,
0.09766804426908493,
0.03002394735813141,
0.009146503172814846,
0.0599970668554306,
-0.09566020220518112,
-0.034909382462501526,
-0.09403321892023087,
0.05656762048602104,
-0.029594661667943,
-3.707204854208612e-8,
-0.07413583993911743,
0.02988128922879696,
0.011630733497440815,
-0.005159769207239151,
0.07581693679094315,
-0.0011031916365027428,
-0.10186362266540527,
0.14313161373138428,
-0.008632550947368145,
0.061150625348091125,
0.004705050494521856,
-0.011641350574791431,
0.040117297321558,
0.007050167769193649,
-0.04512258619070053,
-0.0077711488120257854,
-0.008293801918625832,
-0.03392479196190834,
-0.04020330682396889,
0.028707481920719147,
0.020604753866791725,
0.048132460564374924,
-0.028842246159911156,
0.023644931614398956,
0.10512705892324448,
-0.15200866758823395,
-0.018314303830266,
0.018161462619900703,
0.011053862050175667,
0.007552214898169041,
-0.04544433578848839,
0.06270673871040344,
0.06388449668884277,
-0.015563682653009892,
0.006590263918042183,
0.11824259907007217,
-0.005514959804713726,
0.007909520529210567,
-0.047231875360012054,
-0.0237690769135952,
0.007911492139101028,
0.06603199243545532,
-0.04671132192015648,
0.0390097014605999,
0.027484485879540443,
-0.02098868228495121,
-0.018424607813358307,
0.021757597103714943,
-0.027796868234872818,
-0.017585381865501404,
0.023825429379940033,
-0.01732887327671051,
0.04550975188612938,
-0.024996256455779076,
0.13452641665935516,
0.0907842218875885,
0.0047400896437466145,
-0.03876202553510666,
0.034048743546009064,
-0.001407817006111145,
0.09060713648796082,
-0.0006603652145713568,
-0.014570514671504498,
0.01632268913090229
] |
A Sanity Check for AI-generated Image Detection | https://openreview.net/forum?id=ODRHZrkOQM | [
"Shilin Yan",
"Ouxiang Li",
"Jiayin Cai",
"Yanbin Hao",
"Xiaolong Jiang",
"Yao Hu",
"Weidi Xie"
] | Poster | With the rapid development of generative models, discerning AI-generated content has evoked increasing attention from both industry and academia. In this paper, we conduct a sanity check on whether the task of AI-generated image detection has been solved. To start with, we present Chameleon dataset, consisting of AI-generated images that are genuinely challenging for human perception. To quantify the generalization of existing methods, we evaluate 9 off-the-shelf AI-generated image detectors on Chameleon dataset. Upon analysis, almost all models misclassify AI-generated images as real ones. Later, we propose AIDE AI-generated Image DEtector with Hybrid Features, which leverages multiple experts to simultaneously extract visual artifacts and noise patterns. Specifically, to capture the high-level semantics, we utilize CLIP to compute the visual embedding. This effectively enables the model to discern AI-generated images based on semantics and contextual information. Secondly, we select the highest and lowest frequency patches in the image, and compute the low-level patchwise features, aiming to detect AI-generated images by low-level artifacts, for example, noise patterns, anti-aliasing effects. While evaluating on existing benchmarks, for example, AIGCDetectBenchmark and GenImage, AIDE achieves +3.5% and +4.6% improvements to state-of-the-art methods, and on our proposed challenging Chameleon benchmarks, it also achieves promising results, despite the problem of detecting AI-generated images remains far from being solved. | AI-generated image detection, Chameleon dataset, mixture-of-experts | null | 1,417 | 2406.19435 | [
-0.05356830358505249,
-0.06193038821220398,
0.03255091980099678,
0.045301951467990875,
0.13624966144561768,
-0.04804716259241104,
0.0190153568983078,
-0.06628607958555222,
-0.008037306368350983,
-0.12273182719945908,
-0.04720679298043251,
-0.12894047796726227,
0.06565715372562408,
-0.02212878316640854,
-0.04647093266248703,
0.01319261733442545,
0.12111493945121765,
0.04226350411772728,
-0.08003614097833633,
-0.07187659293413162,
0.02036202885210514,
-0.031929340213537216,
0.0054498896934092045,
-0.08715716004371643,
0.06814699620008469,
0.02732582576572895,
0.020794304087758064,
-0.019472448155283928,
0.025044413283467293,
-0.01904197782278061,
0.06988079845905304,
-0.04152628034353256,
0.0493103452026844,
0.09145945310592651,
-0.0031100185588002205,
0.054816097021102905,
0.0020724518690258265,
0.0530482679605484,
0.080559641122818,
-0.004016845487058163,
0.007764836307615042,
0.0228409506380558,
0.02481989376246929,
0.0022410962264984846,
0.0491829439997673,
0.007641293108463287,
-0.061105627566576004,
-0.07855107635259628,
0.003743196837604046,
0.009129278361797333,
-0.04127989336848259,
-0.009583167731761932,
0.026110071688890457,
0.05163944885134697,
-0.034117601811885834,
0.0025467516388744116,
0.04440871253609657,
0.02922360599040985,
0.022415710613131523,
0.04746696725487709,
-0.017207656055688858,
-0.055842723697423935,
0.034984324127435684,
-0.024676335975527763,
-0.021641243249177933,
-0.019318165257573128,
0.006368889939039946,
-0.017333857715129852,
0.031630706042051315,
-0.04814644157886505,
-0.03488573804497719,
0.1120428517460823,
-0.055371981114149094,
0.02320135571062565,
-0.035483088344335556,
0.0812152624130249,
-0.004268039483577013,
-0.02880999445915222,
0.08396420627832413,
-0.13974995911121368,
-0.008241619914770126,
-0.04968627914786339,
0.04386264458298683,
-0.009410575963556767,
0.11807747930288315,
0.03703995421528816,
0.02539062313735485,
0.04120885953307152,
-0.004999758675694466,
0.10968883335590363,
-0.07797442376613617,
-0.010881262831389904,
0.020033184438943863,
0.003995712846517563,
0.09450636804103851,
0.07526895403862,
0.011601113714277744,
-0.11482493579387665,
-0.0026048440486192703,
0.08317060768604279,
-0.049356501549482346,
0.00432547926902771,
0.048928700387477875,
-0.0037558877374976873,
0.06833305209875107,
-0.0027293963357806206,
0.07556858658790588,
-0.02415197715163231,
0.026412775740027428,
0.005145756062120199,
0.018392100930213928,
0.07749897241592407,
0.01809505745768547,
-0.06644929945468903,
0.07946085929870605,
0.035691048949956894,
-0.022837813943624496,
0.04001273214817047,
0.010504860430955887,
-0.007067447528243065,
0.04570557549595833,
0.021893708035349846,
-0.07857444882392883,
0.014969752170145512,
0.08722961694002151,
-0.04234178364276886,
-0.07093945145606995,
-1.5807431584090162e-34,
0.036474112421274185,
-0.018313268199563026,
0.06231234595179558,
0.017726797610521317,
0.0756092518568039,
-0.02463936060667038,
-0.12523312866687775,
0.0663226917386055,
-0.024249950423836708,
0.026537073776125908,
-0.0344255156815052,
0.004650571383535862,
-0.04267162084579468,
0.11298564076423645,
0.004112116061151028,
0.03240646421909332,
-0.014652695506811142,
-0.030860906466841698,
-0.05374755710363388,
0.05230405181646347,
-0.0025943161454051733,
-0.05114833265542984,
0.033998485654592514,
0.0018607833189889789,
-0.02816055715084076,
-0.01719629019498825,
0.06369362026453018,
-0.036479633301496506,
0.028488323092460632,
0.019099337980151176,
-0.11755984276533127,
-0.02281726524233818,
0.044632766395807266,
0.0463278628885746,
-0.04783708602190018,
-0.013693244196474552,
-0.007137654349207878,
0.03449434041976929,
-0.03729920834302902,
0.05881965905427933,
0.002604125766083598,
0.04221450537443161,
-0.006210966035723686,
0.005067487712949514,
-0.07365906983613968,
-0.006542769260704517,
0.0908677726984024,
0.008809353224933147,
-0.028537802398204803,
0.05832025036215782,
0.015550817362964153,
-0.0007009123219177127,
0.05428098887205124,
-0.08210891485214233,
-0.054542720317840576,
-0.05904611200094223,
0.047153159976005554,
-0.0415923073887825,
0.06863715499639511,
-0.056395672261714935,
-0.0010383134940639138,
0.08173022419214249,
0.0020611335057765245,
0.06515423953533173,
-0.00861906073987484,
0.025000104680657387,
0.007699043024331331,
0.07903710007667542,
0.00024062713782768697,
0.03205570951104164,
-0.009518462233245373,
0.022370675578713417,
-0.04760132357478142,
-0.038707129657268524,
0.00919394288212061,
0.01795591600239277,
0.009069351479411125,
-0.034684427082538605,
0.04702160879969597,
0.02475825697183609,
-0.04012655094265938,
-0.05819233879446983,
0.019741719588637352,
-0.11594381183385849,
-0.0601041205227375,
0.011571680195629597,
-0.007459296379238367,
-0.048264600336551666,
0.010587278753519058,
-0.026514297351241112,
-0.06537589430809021,
-0.006019215099513531,
0.010022556409239769,
-0.01511306781321764,
-0.025419149547815323,
9.319321050029019e-34,
0.0302893053740263,
0.03946790099143982,
-0.04335794225335121,
0.04729560762643814,
-0.06877567619085312,
-0.012284276075661182,
0.03546745330095291,
0.05881443992257118,
-0.026238461956381798,
-0.055251941084861755,
-0.010050071403384209,
-0.0095589067786932,
-0.028245048597455025,
-0.07108207792043686,
-0.04317952319979668,
-0.040465857833623886,
-0.019409840926527977,
-0.009246302768588066,
-0.06299831718206406,
0.06004580110311508,
0.008087629452347755,
0.07253072410821915,
-0.0066121360287070274,
0.0012906015617772937,
-0.056723061949014664,
0.08335580676794052,
-0.04853574186563492,
0.01704944111406803,
-0.0004584738635458052,
-0.03649722412228584,
0.012116161175072193,
0.02948453649878502,
0.020912352949380875,
-0.00813451036810875,
-0.0496656708419323,
0.006583013106137514,
0.05582273378968239,
-0.039441339671611786,
0.013040675781667233,
0.085704006254673,
0.0800400823354721,
-0.035768479108810425,
-0.10591353476047516,
0.10804138332605362,
-0.04672180488705635,
-0.09977538138628006,
-0.016933199018239975,
0.06333739310503006,
0.008185244165360928,
0.04143522307276726,
-0.07236696034669876,
0.01957375556230545,
-0.08877798169851303,
0.002637571655213833,
-0.00917759072035551,
-0.036806415766477585,
-0.07603022456169128,
0.04323570802807808,
0.09386629611253738,
0.07066676765680313,
-0.042334575206041336,
-0.09232271462678909,
-0.06431858241558075,
-0.12907001376152039,
-0.010520752519369125,
0.006521930452436209,
-0.035825829952955246,
0.05235178768634796,
0.014814797788858414,
-0.00014266130165196955,
0.05906364321708679,
-0.021355783566832542,
-0.019523167982697487,
0.00039184975321404636,
-0.027679067105054855,
0.043703459203243256,
0.06341829895973206,
-0.020807486027479172,
0.04272722825407982,
-0.023111335933208466,
0.01356411725282669,
-0.0936751440167427,
0.04925479367375374,
0.058889709413051605,
0.032191842794418335,
0.06691667437553406,
-0.09723873436450958,
-0.02385593391954899,
0.026187969371676445,
0.024224162101745605,
-0.09190479665994644,
0.036914926022291183,
-0.0641663670539856,
0.11397840827703476,
-0.015473290346562862,
-4.2890768270353874e-8,
-0.04567229375243187,
-0.01314614899456501,
0.005367760546505451,
-0.024663610383868217,
0.07005752623081207,
-0.038576871156692505,
-0.0485045500099659,
0.06406544893980026,
-0.017747510224580765,
-0.09683579206466675,
0.051854684948921204,
-0.02492937259376049,
-0.026997597888112068,
-0.03203604742884636,
0.05099437013268471,
0.01239637192338705,
0.025127295404672623,
0.11229301989078522,
-0.005609639920294285,
-0.00949475634843111,
0.018486464396119118,
-0.005141557659953833,
0.04445727914571762,
-0.06951874494552612,
0.004123134538531303,
-0.05927443504333496,
-0.050253402441740036,
0.04443548247218132,
0.007134031970053911,
0.008147663436830044,
0.0204754825681448,
0.039033371955156326,
0.01283001434057951,
-0.06339818984270096,
0.05690881982445717,
0.1334546059370041,
0.0014220754383131862,
-0.04420437663793564,
0.005791237112134695,
-0.04911408945918083,
-0.0012007998302578926,
0.03820975124835968,
0.012906677089631557,
-0.03916466236114502,
-0.03358832746744156,
-0.005400151014328003,
0.1036432534456253,
-0.08297191560268402,
0.03776591643691063,
0.045532893389463425,
-0.017803531140089035,
-0.07105836272239685,
0.040206555277109146,
0.0722540095448494,
0.013702831231057644,
-0.06101519986987114,
-0.019833220168948174,
-0.021699456498026848,
0.04046895354986191,
0.112914077937603,
0.03321303799748421,
0.009165020659565926,
0.036055948585271835,
-0.05179556459188461
] |
Neural Phylogeny: Fine-Tuning Relationship Detection among Neural Networks | https://openreview.net/forum?id=jv2zHOalpL | [
"Runpeng Yu",
"Xinchao Wang"
] | Poster | Given a collection of neural networks, can we determine which are parent models and which are child models fine-tuned from the parents?
In this work, we strive to answer this question
via introducing a new task termed as neural phylogeny detection, aimed at identifying the existence and direction of the fine-tuning relationship. Specifically, neural phylogeny detection attempts to identify all parent-child model pairs and determine, within each pair, which model is the parent and which is the child.
We present two approaches for neural phylogeny detection: a learning-free method and a learning-based method. First, we propose a metric that leverages the distance from network parameters to a fake initialization to infer fine-tuning directions. By integrating this metric with traditional clustering algorithms, we propose a series of efficient, learning-free neural phylogeny detection methods. Second, we introduce a transformer-based neural phylogeny detector, which significantly enhances detection accuracy through a learning-based manner. Extensive experiments, ranging from shallow fully-connected networks to open-sourced Stable Diffusion and LLaMA models, progressively validate the effectiveness of both methods. The results demonstrate the reliability of both the learning-free and the learning-based approaches across various learning tasks and network architectures, as well as their ability to detect cross-generational phylogeny between ancestor models and their fine-tuned descendants. | Neural Phylogeny, Finetuning | null | 1,414 | null | [
-0.08933323621749878,
-0.11158910393714905,
0.02948710136115551,
0.02468198537826538,
0.05755796283483505,
0.05314118042588234,
-0.12455093115568161,
-0.08092666417360306,
0.013716979883611202,
-0.06953567266464233,
0.050455491989851,
-0.07839248329401016,
-0.012138658203184605,
0.05766015872359276,
-0.1289972960948944,
-0.005772116594016552,
0.022068258374929428,
0.055265381932258606,
-0.08247315883636475,
-0.05048375204205513,
0.024832172319293022,
0.05906735733151436,
0.05612916126847267,
-0.014530343003571033,
-0.008959565311670303,
0.03142734244465828,
0.05663316324353218,
-0.0050425054505467415,
-0.01644100435078144,
0.01558472029864788,
0.05648664012551308,
0.024610817432403564,
-0.014433090575039387,
0.026625322178006172,
-0.03470645844936371,
0.016167139634490013,
-0.012509351596236229,
0.02023024670779705,
0.02674207091331482,
0.02833373285830021,
0.02962658181786537,
0.04723607748746872,
0.013633319176733494,
-0.006414818577468395,
0.05207603797316551,
0.04054183512926102,
-0.04044092074036598,
0.010338089428842068,
-0.05158708244562149,
-0.02919871173799038,
-0.06512261927127838,
-0.011037909425795078,
-0.03095288760960102,
0.10552505403757095,
0.046696729958057404,
0.07798312604427338,
0.0267579834908247,
-0.011686683632433414,
-0.04334094375371933,
-0.028287382796406746,
-0.012681759893894196,
-0.032751601189374924,
-0.07220856845378876,
-0.04752451926469803,
-0.0398174449801445,
0.09732892364263535,
0.0019873064011335373,
0.06588010489940643,
0.046861130744218826,
0.0009087990620173514,
-0.006863896269351244,
0.09972267597913742,
-0.02679799310863018,
0.08537346869707108,
-0.00813739001750946,
0.07350452989339828,
0.12005221843719482,
0.010844227857887745,
0.027763796970248222,
-0.10663453489542007,
-0.0009549699607305229,
-0.0029119309037923813,
0.0363808237016201,
-0.05327611416578293,
0.12285223603248596,
0.026035374030470848,
-0.007194578181952238,
-0.020818380638957024,
0.03930804878473282,
-0.017387865111231804,
-0.059815824031829834,
-0.005239379592239857,
-0.04346834123134613,
-0.07470287382602692,
0.015468035824596882,
0.019740061834454536,
0.02334105409681797,
-0.04533083364367485,
0.01208405289798975,
0.07550972700119019,
-0.04724521189928055,
-0.05148221552371979,
0.05807632580399513,
-0.03793540596961975,
0.016619814559817314,
0.044773269444704056,
0.050438035279512405,
0.02727983146905899,
0.07741732150316238,
-0.09129532426595688,
0.014738043770194054,
0.026999173685908318,
0.031359102576971054,
0.02197488769888878,
0.01654946058988571,
-0.04030036926269531,
0.040194764733314514,
-0.034875690937042236,
0.03795932978391647,
0.03882327303290367,
0.02649909444153309,
-0.02742115780711174,
-0.03359001502394676,
0.009351125918328762,
0.011046970263123512,
-0.042765192687511444,
-0.18652579188346863,
4.860852811642475e-34,
0.04338221997022629,
0.014651894569396973,
0.07166128605604172,
0.007401331793516874,
0.008212056010961533,
-0.023215895518660545,
-0.07457653433084488,
0.05839505046606064,
-0.06726522743701935,
-0.006429044995456934,
-0.10196526348590851,
0.04160045087337494,
-0.06470122933387756,
0.02987305074930191,
0.042571745812892914,
-0.000027921470973524265,
-0.026949627324938774,
0.06247427314519882,
0.0201811995357275,
-0.08239558339118958,
-0.03214766085147858,
-0.04814935848116875,
0.03893959894776344,
-0.0756492093205452,
-0.02844429761171341,
0.0003573301946744323,
0.0903501957654953,
-0.07538014650344849,
-0.050532419234514236,
-0.0036468757316470146,
-0.08306100964546204,
0.0009636134491302073,
-0.01940370537340641,
0.020584914833307266,
-0.03514841943979263,
-0.0677325651049614,
-0.010765322484076023,
-0.04075298085808754,
-0.03386548161506653,
-0.058304380625486374,
0.017352966591715813,
-0.003661064663901925,
0.03379078209400177,
-0.018748197704553604,
0.005961709190160036,
-0.011155406013131142,
-0.04287818819284439,
-0.049531951546669006,
0.01687369868159294,
-0.010052410885691643,
0.009946892969310284,
-0.026124605908989906,
0.023015128448605537,
0.028952861204743385,
-0.12809640169143677,
0.07191255688667297,
0.0040626004338264465,
0.051149241626262665,
0.013355103321373463,
0.04362916201353073,
0.04540758579969406,
0.014431193470954895,
-0.004920760169625282,
0.019046030938625336,
0.0735723227262497,
-0.000733073684386909,
-0.0411011204123497,
0.08725113421678543,
0.055302947759628296,
0.026438811793923378,
-0.001194508746266365,
-0.02404927648603916,
-0.05503646656870842,
-0.030217185616493225,
0.043029751628637314,
-0.00585342850536108,
0.010925417765974998,
-0.012561177834868431,
-0.034546710550785065,
-0.007820469327270985,
0.0009642710210755467,
0.025296298786997795,
-0.00470099737867713,
-0.01048120204359293,
-0.05697105452418327,
0.010928786359727383,
0.022584684193134308,
-0.000806623138487339,
0.01984839327633381,
0.04830325022339821,
-0.08225790411233902,
-0.03902347385883331,
0.0015543532790616155,
0.0005877458606846631,
0.054666854441165924,
-4.573707551562918e-34,
0.0024206945672631264,
0.00558259105309844,
-0.036409128457307816,
0.04631182551383972,
-0.02089725062251091,
-0.09237732738256454,
-0.03804069757461548,
0.09110161662101746,
-0.005286415107548237,
-0.01060789916664362,
0.03450799360871315,
-0.027110934257507324,
0.03715239465236664,
-0.06224030628800392,
0.08724287897348404,
-0.052585918456315994,
-0.0483088418841362,
0.043332118541002274,
0.10635395348072052,
0.03769988566637039,
0.019255898892879486,
0.026091426610946655,
-0.13300898671150208,
0.009989375248551369,
-0.03779450058937073,
0.04652862250804901,
-0.013058343902230263,
0.10194225609302521,
-0.04635905474424362,
0.09366679936647415,
-0.07063692063093185,
0.08036825805902481,
-0.04896621033549309,
0.014985540881752968,
-0.0172614473849535,
0.07238901406526566,
-0.043516844511032104,
-0.045738354325294495,
0.0506419837474823,
0.05949831008911133,
0.015440852381289005,
0.03929824382066727,
-0.11742634326219559,
0.047968726605176926,
0.007704559247940779,
-0.07241767644882202,
-0.02078446000814438,
0.06403247267007828,
-0.03790423646569252,
0.044478889554739,
-0.0003195695753674954,
0.015293732285499573,
0.0010171589674428105,
-0.024657757952809334,
-0.020840970799326897,
0.055141400545835495,
0.02499185875058174,
0.013708948157727718,
0.04001596197485924,
0.014389709569513798,
-0.05425781384110451,
-0.10251425951719284,
-0.04396364092826843,
-0.050625529140233994,
-0.05528838187456131,
-0.02221612259745598,
0.024475080892443657,
0.03443356230854988,
0.1376732885837555,
0.0064517781138420105,
0.029722828418016434,
0.1036490723490715,
0.024798152968287468,
0.01403516810387373,
-0.036420758813619614,
-0.04420249164104462,
-0.030519234016537666,
-0.05003830045461655,
0.05978338047862053,
-0.026279069483280182,
0.01986047253012657,
-0.06545903533697128,
0.08094609528779984,
-0.0211203470826149,
0.04236825555562973,
-0.020825104787945747,
0.034533776342868805,
0.06242211535573006,
0.10045257955789566,
0.006510103587061167,
0.02838314138352871,
-0.0649552047252655,
0.011082286946475506,
-0.048861313611269,
-0.1313226819038391,
-4.2776850506243136e-8,
-0.029318364337086678,
-0.027340276166796684,
0.03938441723585129,
0.039400894194841385,
0.03913114592432976,
0.009998112916946411,
0.012204937636852264,
0.11376695334911346,
-0.027730854228138924,
-0.04220075160264969,
0.0018590176478028297,
-0.014681046828627586,
-0.03398500755429268,
-0.01769467629492283,
-0.009730125777423382,
0.05823661759495735,
-0.021977191790938377,
0.0343940407037735,
-0.06120184063911438,
0.06353398412466049,
0.01519649289548397,
0.10180146992206573,
0.03555479645729065,
0.10704275965690613,
0.05615243688225746,
-0.09329178929328918,
-0.06329277157783508,
0.08936366438865662,
0.003585711121559143,
-0.02608313038945198,
0.0028102463111281395,
0.030579566955566406,
-0.00457588117569685,
-0.01755508966743946,
0.054149698466062546,
0.13700154423713684,
-0.05600009486079216,
-0.008309970609843731,
-0.03308066353201866,
-0.02085520327091217,
0.0028441580943763256,
0.06457537412643433,
-0.05653293430805206,
0.026617344468832016,
0.023428620770573616,
-0.02782575413584709,
0.08145356178283691,
-0.02672574855387211,
0.021432531997561455,
-0.0430871956050396,
0.02525852993130684,
-0.021480295807123184,
-0.05227883532643318,
-0.02518363855779171,
-0.09311982989311218,
-0.040560342371463776,
-0.0596206896007061,
-0.07306073606014252,
0.02818145416676998,
0.06607536971569061,
0.027410726994276047,
0.0590929239988327,
0.06341229379177094,
-0.06467744708061218
] |
Breaking the log(1/Δ2) Barrier: Better Batched Best Arm Identification with Adaptive Grids | https://openreview.net/forum?id=buxFBI6GG4 | [
"Tianyuan Jin",
"Qin Zhang",
"Dongruo Zhou"
] | Poster | We investigate the problem of batched best arm identification in multi-armed bandits, where we want to find the best arm from a set of $n$ arms while minimizing both the number of samples and batches. We introduce an algorithm that achieves near-optimal sample complexity and features an instance-sensitive batch complexity, which breaks the $\log(1/\Delta_2)$ barrier. The main contribution of our algorithm is a novel sample allocation scheme that effectively balances exploration and exploitation for batch sizes. Experimental results indicate that our approach is more batch-efficient across various setups. We also extend this framework to the problem of batched best arm identification in linear bandits and achieve similar improvements. | Bandits | null | 1,407 | 2501.17370 | [
-0.075112484395504,
0.017199955880641937,
-0.03528151661157608,
0.016509078443050385,
0.05896652117371559,
-0.04984654486179352,
0.03674499690532684,
-0.012425427325069904,
0.02744233049452305,
-0.0007522473460994661,
-0.07331979274749756,
0.017983337864279747,
0.05189194902777672,
-0.04275405779480934,
0.0229969285428524,
-0.016013212502002716,
0.11561751365661621,
0.027187669649720192,
0.001943061826750636,
0.023565810173749924,
-0.06022653728723526,
0.0019548083655536175,
0.10648217797279358,
-0.027940742671489716,
-0.011036477982997894,
-0.07266711443662643,
-0.02939874678850174,
-0.02658269740641117,
0.06906507164239883,
-0.06152849271893501,
0.04918156564235687,
0.036641206592321396,
-0.007143862545490265,
-0.0490688718855381,
0.03980507701635361,
-0.01952408440411091,
-0.09503132104873657,
-0.008359953761100769,
0.05841095373034477,
0.037549879401922226,
0.08387203514575958,
0.08407764136791229,
-0.018165135756134987,
-0.020823797211050987,
0.006591445300728083,
0.034136734902858734,
-0.02464691363275051,
0.02397121861577034,
0.06890740245580673,
0.03772784769535065,
-0.08305612206459045,
-0.0246546920388937,
-0.0017846905393525958,
-0.026566386222839355,
-0.08669482916593552,
-0.03799401968717575,
0.05306555703282356,
0.002664390252903104,
0.00908925011754036,
0.13089342415332794,
0.04189614579081535,
-0.13471634685993195,
0.021049952134490013,
-0.01530252955853939,
0.03335440158843994,
-0.03800179064273834,
0.0498453713953495,
0.05691343918442726,
0.04362143203616142,
0.01719985343515873,
0.0535150021314621,
0.03875796124339104,
-0.09612984210252762,
0.025652648881077766,
0.04379595071077347,
-0.010953954420983791,
0.024747008457779884,
0.028760967776179314,
0.027687136083841324,
-0.04718063771724701,
-0.09540776908397675,
-0.007384548895061016,
0.015015707351267338,
-0.052270155400037766,
0.044650010764598846,
-0.047106992453336716,
-0.01725713163614273,
0.003319086506962776,
0.1589180827140808,
-0.03440563753247261,
-0.020155994221568108,
0.01732015609741211,
-0.01958734542131424,
0.005141881760209799,
0.038808953016996384,
0.04004478454589844,
0.009111297316849232,
0.06450502574443817,
-0.04428355023264885,
0.06455070525407791,
0.014881523326039314,
0.0431685708463192,
0.049876049160957336,
-0.010281341150403023,
-0.0016835795249789953,
-0.019737718626856804,
0.02505042776465416,
-0.026274623349308968,
0.03906571492552757,
-0.0938863754272461,
0.10119705647230148,
0.06345439702272415,
0.03318699449300766,
0.043605249375104904,
0.014295750297605991,
-0.08347798138856888,
0.036370355635881424,
-0.03771015256643295,
-0.0483444519340992,
0.06534771621227264,
0.021283341571688652,
-0.021325169131159782,
0.02163308672606945,
-0.038359202444553375,
0.033779874444007874,
0.10234443098306656,
-0.025236809626221657,
-2.625893208441233e-34,
0.0288867074996233,
0.013156344182789326,
-0.01626047119498253,
-0.05448567494750023,
0.02327575720846653,
-0.006911481264978647,
0.04692874103784561,
-0.06836555898189545,
-0.05689578503370285,
0.0660821720957756,
-0.024330750107765198,
0.03800193965435028,
-0.0484127514064312,
0.0585373193025589,
0.08877404034137726,
-0.0519242025911808,
-0.007007284089922905,
0.08927340060472488,
0.013855447061359882,
-0.05915197730064392,
-0.006110412534326315,
-0.03648071736097336,
0.0060149915516376495,
-0.10144442319869995,
-0.03185839205980301,
-0.0166622344404459,
-0.00026214183890260756,
-0.03416958078742027,
0.015111460350453854,
0.01676422543823719,
-0.05423993989825249,
-0.01880151592195034,
-0.030610820278525352,
0.051407817751169205,
0.062416061758995056,
-0.0010919909691438079,
-0.03705146163702011,
0.009239294566214085,
-0.03125450015068054,
0.06871043890714645,
0.0015362186823040247,
0.03999553620815277,
0.068193219602108,
-0.05494237318634987,
-0.10812422633171082,
-0.10669620335102081,
-0.08159346878528595,
-0.03847932443022728,
-0.04085572808980942,
-0.011637469753623009,
-0.04351016879081726,
-0.029113829135894775,
0.01864025369286537,
-0.01395266130566597,
-0.02072380855679512,
0.005216039717197418,
-0.06064699962735176,
0.05382569879293442,
0.03453434258699417,
0.06535708159208298,
0.02049889974296093,
-0.06251432001590729,
-0.018515247851610184,
0.08006016165018082,
0.017037585377693176,
-0.003808553796261549,
0.006202468182891607,
0.018933530896902084,
0.030478697270154953,
0.013982200063765049,
0.013511530123651028,
-0.07164768874645233,
0.03240414336323738,
-0.007629724219441414,
0.08852231502532959,
-0.06493841111660004,
0.0967402309179306,
0.06829250603914261,
-0.013170265592634678,
-0.021218102425336838,
0.038578253239393234,
0.016602929681539536,
-0.017923206090927124,
-0.021216468885540962,
-0.053532518446445465,
-0.01643572747707367,
-0.007589865010231733,
-0.12806938588619232,
-0.08044901490211487,
-0.028576724231243134,
-0.09033504873514175,
0.010966474190354347,
-0.06158642843365669,
0.025774076581001282,
0.011060749180614948,
-1.7020048827113135e-33,
0.03989063948392868,
-0.003940450958907604,
0.09466585516929626,
-0.03879862278699875,
0.0340205579996109,
-0.03804483637213707,
0.03843751922249794,
-0.06846477836370468,
-0.04792061448097229,
-0.044067978858947754,
-0.08462768793106079,
0.008923192508518696,
0.02615279331803322,
-0.018146025016903877,
0.027380958199501038,
0.0011584853054955602,
0.05048885941505432,
0.10193867981433868,
0.01354883797466755,
-0.03728527948260307,
0.027132166549563408,
0.05727279558777809,
0.02076604589819908,
0.06327231973409653,
0.004878299310803413,
-0.028235122561454773,
-0.03803648054599762,
0.0194555576890707,
0.0033610695973038673,
-0.06962351500988007,
-0.10567906498908997,
-0.028078025206923485,
-0.0494539700448513,
0.0250327717512846,
-0.02919442392885685,
0.024703973904252052,
0.06349217146635056,
0.07361017167568207,
-0.02994079329073429,
0.1472061574459076,
0.08902289718389511,
0.01320855226367712,
-0.013459330424666405,
0.011992027051746845,
-0.0075217378325760365,
-0.02153857611119747,
-0.0927208736538887,
-0.00043217558413743973,
-0.02062322571873665,
0.011669941246509552,
-0.04904472455382347,
0.06265683472156525,
-0.060383494943380356,
0.04992837831377983,
-0.015893390402197838,
-0.1596333086490631,
0.04090623930096626,
0.02405858039855957,
0.049670834094285965,
0.013440805487334728,
-0.04734314605593681,
0.08017630875110626,
-0.09210027754306793,
0.038191549479961395,
0.03142824396491051,
-0.028756126761436462,
-0.059686433523893356,
-0.007707428652793169,
-0.048724886029958725,
0.028860827907919884,
0.039467282593250275,
0.013161388225853443,
0.011816645972430706,
0.005981566850095987,
-0.05449017137289047,
0.056833721697330475,
-0.005069296807050705,
-0.10862802714109421,
-0.01585250534117222,
-0.04305364191532135,
-0.04331754148006439,
0.0542253702878952,
0.03324176371097565,
0.015614226460456848,
0.048776451498270035,
0.0744907557964325,
0.0668209120631218,
0.005437924060970545,
-0.04223731905221939,
-0.07402374595403671,
0.001424306188710034,
-0.07978405058383942,
0.059778161346912384,
0.04609440267086029,
-0.042735595256090164,
-4.052483859595668e-8,
-0.0319642536342144,
0.07865259796380997,
0.03829607367515564,
-0.002280336571857333,
0.06269999593496323,
0.08270620554685593,
-0.11692690849304199,
0.0425335168838501,
-0.05292969197034836,
-0.030108826234936714,
0.11493657529354095,
0.012934613972902298,
0.05016208440065384,
0.009458687156438828,
-0.047771599143743515,
0.015336507931351662,
-0.0932934507727623,
-0.06177655607461929,
-0.020542984828352928,
0.01096118614077568,
-0.02742970734834671,
0.022648461163043976,
0.07437757402658463,
-0.03585611283779144,
0.011707881465554237,
-0.07066434621810913,
-0.038034774363040924,
-0.010899988003075123,
0.06102825701236725,
-0.0030265673995018005,
-0.07798878848552704,
0.01351839117705822,
0.04250434413552284,
0.03745841607451439,
0.0815729945898056,
0.03605528175830841,
-0.031682126224040985,
-0.018292857334017754,
-0.08044019341468811,
-0.0495154932141304,
-0.053938914090394974,
-0.05502503365278244,
-0.012204481288790703,
-0.06540947407484055,
0.0282608512789011,
0.02647794596850872,
-0.05429917573928833,
-0.057306673377752304,
0.027813810855150223,
-0.06904049217700958,
0.06644636392593384,
-0.0651903823018074,
0.03452032804489136,
0.07082699239253998,
0.06217305362224579,
-0.005984711926430464,
-0.046865593641996384,
-0.06843408197164536,
0.06314905732870102,
0.04444989562034607,
0.04355171322822571,
-0.011522434651851654,
-0.06259693205356598,
-0.015472621656954288
] |
Visual Agents as Fast and Slow Thinkers | https://openreview.net/forum?id=ncCuiD3KJQ | [
"Guangyan Sun",
"Mingyu Jin",
"Zhenting Wang",
"Cheng-Long Wang",
"Siqi Ma",
"Qifan Wang",
"Tong Geng",
"Ying Nian Wu",
"Yongfeng Zhang",
"Dongfang Liu"
] | Poster | Achieving human-level intelligence requires refining cognitive distinctions between \textit{System 1} and \textit{System 2} thinking. While contemporary AI, driven by large language models, demonstrates human-like traits, it falls short of genuine cognition. Transitioning from structured benchmarks to real-world scenarios presents challenges for visual agents, often leading to inaccurate and overly confident responses. To address the challenge, we introduce \textbf{\textsc{FaST}}, which incorporates the \textbf{Fa}st and \textbf{S}low \textbf{T}hinking mechanism into visual agents. \textsc{FaST} employs a switch adapter to dynamically select between \textit{System 1/2} modes, tailoring the problem-solving approach to different task complexity. It tackles uncertain and unseen objects by adjusting model confidence and integrating new contextual data. With this novel design, we advocate a \textit{flexible system}, \textit{hierarchical reasoning} capabilities, and a \textit{transparent decision-making} pipeline, all of which contribute to its ability to emulate human-like cognitive processes in visual intelligence. Empirical results demonstrate that \textsc{FaST} outperforms various well-known baselines, achieving 80.8\% accuracy over $VQA^{v2}$ for visual question answering and 48.7\% $GIoU$ score over ReasonSeg for reasoning segmentation, demonstrate \textsc{FaST}'s superior performance. Extensive testing validates the efficacy and robustness of \textsc{FaST}'s core components, showcasing its potential to advance the development of cognitive visual agents in AI systems. | Multimodal Large Language Model, System2 Thinking, Language Agent | null | 1,406 | 2408.08862 | [
0.043871957808732986,
-0.08242830634117126,
-0.0043465858325362206,
0.05591738224029541,
0.063397616147995,
-0.08226212859153748,
0.017192082479596138,
0.057922232896089554,
0.014680822379887104,
0.0030283683445304632,
-0.028137341141700745,
-0.1296866536140442,
0.06872408092021942,
0.009906234219670296,
0.017177492380142212,
0.031308017671108246,
0.1230447068810463,
0.011564478278160095,
-0.04054172337055206,
-0.03519696742296219,
0.0428694486618042,
-0.04757772013545036,
0.008091195486485958,
-0.05389384925365448,
-0.005699647590517998,
0.03576986864209175,
-0.027269184589385986,
-0.05675061419606209,
-0.026044683530926704,
-0.02449849434196949,
0.0317155122756958,
-0.02709531970322132,
0.0199606791138649,
-0.003560013836249709,
-0.06049853935837746,
0.0663876160979271,
0.010597961023449898,
0.006391911767423153,
0.006261812523007393,
-0.04763718694448471,
-0.014894159510731697,
-0.03837329521775246,
0.027602721005678177,
0.01665617898106575,
0.030903469771146774,
0.015231717377901077,
-0.03188023343682289,
-0.05352840572595596,
-0.013523607514798641,
-0.025038503110408783,
-0.16649581491947174,
-0.03137809410691261,
0.00968034379184246,
-0.04671260342001915,
-0.012343467213213444,
0.11115367710590363,
0.006252642255276442,
0.0426115058362484,
-0.021832002326846123,
-0.02721104957163334,
-0.025740087032318115,
-0.07478716969490051,
0.05980726331472397,
-0.017377417534589767,
0.02099248580634594,
0.05079396441578865,
-0.04192398861050606,
-0.03935853764414787,
0.03901282325387001,
-0.026319609954953194,
0.0038869609124958515,
0.10868526995182037,
-0.0073788841255009174,
0.006789408624172211,
-0.010851909406483173,
-0.03719804063439369,
-0.01721450500190258,
-0.035148922353982925,
0.10482428222894669,
-0.06438548117876053,
-0.011254933662712574,
0.03597566857933998,
-0.03927003964781761,
0.060612209141254425,
0.09224504977464676,
0.003986465744674206,
0.005791581701487303,
0.06202429160475731,
-0.007836352102458477,
-0.04124896600842476,
-0.00026493999757803977,
-0.1250331848859787,
0.11279638856649399,
0.004524705465883017,
0.0754459798336029,
0.005206278525292873,
-0.003821812802925706,
-0.08167356997728348,
-0.05729079246520996,
0.04782642424106598,
-0.029917549341917038,
0.020919837057590485,
-0.0006819387199357152,
0.0033768494613468647,
-0.06332394480705261,
-0.05248284712433815,
0.12884698808193207,
0.04959917813539505,
-0.015423999167978764,
-0.10467799752950668,
-0.005481502041220665,
-0.013735484331846237,
0.03257665038108826,
0.007922942750155926,
-0.017419494688510895,
-0.053588103502988815,
0.000963208731263876,
-0.047146447002887726,
0.04221159219741821,
0.03043876215815544,
0.05989841744303703,
0.04275478050112724,
0.048107318580150604,
-0.01347946934401989,
0.09619338810443878,
0.026332782581448555,
0.018839417025446892,
-1.54233730136813e-33,
0.013842637650668621,
0.04901763051748276,
0.09753148257732391,
0.09528673440217972,
0.05484782159328461,
-0.005478288512676954,
-0.02838084287941456,
0.022812534123659134,
-0.05069206282496452,
0.05273432657122612,
-0.0036987115163356066,
-0.0056788865476846695,
-0.10508319735527039,
0.13954798877239227,
0.04887476563453674,
-0.0777847170829773,
-0.018930774182081223,
0.0360916405916214,
-0.031525157392024994,
-0.09490689635276794,
0.006899733562022448,
0.017755692824721336,
0.017012521624565125,
-0.025447070598602295,
0.05966527760028839,
0.0008155372925102711,
0.0532880499958992,
-0.07783667743206024,
0.01907443255186081,
0.0091611472889781,
-0.1329345554113388,
0.05672622099518776,
0.0021861311979591846,
0.040883228182792664,
-0.02200588956475258,
-0.03220737352967262,
0.00045218487503007054,
-0.09232397377490997,
0.046206336468458176,
0.036300111562013626,
-0.040535591542720795,
0.017870711162686348,
0.02903948724269867,
0.012919087894260883,
-0.033620405942201614,
-0.03364367038011551,
0.0047566890716552734,
0.09502840787172318,
-0.08829369395971298,
-0.0214651208370924,
0.04312043637037277,
0.06119824945926666,
0.036617711186409,
-0.0319831557571888,
0.044175852090120316,
-0.05989309400320053,
0.009816749952733517,
0.05752112716436386,
0.01943802461028099,
0.0247654989361763,
-0.011899856850504875,
0.002775557804852724,
-0.0338280089199543,
0.15632715821266174,
0.025233786553144455,
0.10026086866855621,
-0.006903211586177349,
0.053769294172525406,
0.06747569143772125,
-0.0008054051431827247,
-0.047709498554468155,
0.06336212903261185,
-0.051207803189754486,
-0.07468320429325104,
0.013742792420089245,
0.026213228702545166,
0.07352244108915329,
-0.10740948468446732,
-0.001148452516645193,
-0.025641238316893578,
-0.017812339588999748,
-0.061145566403865814,
-0.061968233436346054,
-0.05606702342629433,
0.012293132953345776,
-0.02380986139178276,
0.022980939596891403,
-0.0635157972574234,
-0.029563255608081818,
-0.03518225997686386,
-0.07715969532728195,
-0.05468462035059929,
-0.005151811987161636,
-0.024494897574186325,
0.01112722884863615,
6.721033132555397e-34,
0.014440659433603287,
-0.08915766328573227,
-0.019909970462322235,
0.021925458684563637,
0.010107968002557755,
0.0021155928261578083,
0.000340277241775766,
-0.03494748845696449,
0.0044605010189116,
-0.0700598731637001,
-0.028385188430547714,
0.0036250310949981213,
-0.00639036251232028,
-0.022235721349716187,
0.06334082782268524,
-0.025213472545146942,
0.01941385492682457,
0.01672053337097168,
0.014689535833895206,
-0.003061998402699828,
0.030556192621588707,
0.04082762822508812,
-0.11768590658903122,
-0.0007226515444926918,
-0.05024748668074608,
0.052865952253341675,
-0.057069920003414154,
0.026210561394691467,
-0.0020240405574440956,
0.01693897321820259,
-0.02614382840692997,
-0.03504114970564842,
0.0012505907798185945,
-0.001958176726475358,
-0.00836834404617548,
0.08120498061180115,
0.014680352061986923,
-0.036326318979263306,
-0.08821946382522583,
0.06083035096526146,
0.007972762919962406,
-0.13233572244644165,
-0.09540876746177673,
-0.044288042932748795,
-0.04214600473642349,
0.0354546420276165,
-0.055507466197013855,
-0.04370740428566933,
-0.005369602236896753,
0.020304478704929352,
0.01784728467464447,
0.0035894194152206182,
-0.07477567344903946,
-0.005234908778220415,
-0.04567193239927292,
-0.05019567161798477,
-0.04261382296681404,
0.002418530872091651,
0.023252910003066063,
0.03826479986310005,
-0.007023760583251715,
0.002463824115693569,
0.014525136910378933,
-0.032101355493068695,
0.04868733882904053,
-0.038065869361162186,
-0.05059518665075302,
0.025368062779307365,
0.045826978981494904,
-0.03273149952292442,
0.03552957624197006,
-0.09355620294809341,
0.008410340175032616,
0.021417366340756416,
0.07971495389938354,
0.008277549408376217,
0.00980240199714899,
0.021428175270557404,
0.036505360156297684,
-0.001735784811899066,
0.02173372358083725,
-0.008922480046749115,
0.05400274321436882,
-0.0029480953235179186,
0.016969094052910805,
0.09375708550214767,
-0.019795028492808342,
0.04966338351368904,
0.05055825784802437,
0.030444324016571045,
-0.05473189428448677,
0.007662491872906685,
-0.011654206551611423,
0.0924450010061264,
-0.11258050799369812,
-4.1778889681154396e-8,
-0.026685727760195732,
0.046959612518548965,
0.06950041651725769,
-0.005918971262872219,
0.0243977140635252,
0.012009285390377045,
-0.0577893890440464,
-0.008158614858984947,
-0.05524255335330963,
-0.013523844070732594,
0.03509955480694771,
-0.04089910537004471,
0.06537722051143646,
-0.07986664026975632,
0.09063112735748291,
0.04950331896543503,
-0.026300879195332527,
-0.0279254000633955,
0.02455439604818821,
-0.0236776415258646,
0.08933602273464203,
0.016395246610045433,
-0.047479428350925446,
0.034897368401288986,
-0.05084772780537605,
-0.07208284735679626,
-0.10189088433980942,
0.12265851348638535,
0.013166810385882854,
0.08578747510910034,
0.05941743776202202,
0.058508869260549545,
0.04043116047978401,
0.0579594224691391,
0.044494085013866425,
0.05305715277791023,
0.034895721822977066,
0.004403003957122564,
-0.029482200741767883,
-0.03701677918434143,
0.041168875992298126,
0.07160845398902893,
-0.042707424610853195,
0.017325421795248985,
0.03476420417428017,
-0.011552433483302593,
-0.06295664608478546,
-0.09786436706781387,
-0.0002546787727624178,
-0.026451660320162773,
-0.06598741561174393,
0.054736364632844925,
-0.043497588485479355,
0.15869943797588348,
0.04714582860469818,
0.016700023785233498,
0.0749589130282402,
-0.07526534050703049,
-0.02555866539478302,
0.06470084935426712,
0.05806303769350052,
0.005626981146633625,
-0.014952737838029861,
-0.006962890736758709
] |
Fourier Sliced-Wasserstein Embedding for Multisets and Measures | https://openreview.net/forum?id=BcYt84rcKq | [
"Tal Amir",
"Nadav Dym"
] | Poster | We present the _Fourier Sliced Wasserstein (FSW) embedding_—a novel method to embed multisets and measures over $\mathbb{R}^d$ into Euclidean space.
Our proposed embedding approximately preserves the sliced Wasserstein distance on distributions, thereby yielding geometrically meaningful representations that better capture the structure of the input. Moreover, it is injective on measures and _bi-Lipschitz_ on multisets—a significant advantage over prevalent methods based on sum- or max-pooling, which are provably not bi-Lipschitz, and, in many cases, not even injective.
The required output dimension for these guarantees is near-optimal: roughly $2 N d$, where $N$ is the maximal input multiset size.
Furthermore, we prove that it is _impossible_ to embed distributions over $\mathbb{R}^d$ into Euclidean space in a bi-Lipschitz manner. Thus, the metric properties of our embedding are, in a sense, the best possible.
Through numerical experiments, we demonstrate that our method yields superior multiset representations that improve performance in practical learning tasks. Specifically, we show that (a) a simple combination of the FSW embedding with an MLP achieves state-of-the-art performance in learning the (non-sliced) Wasserstein distance; and (b) replacing max-pooling with the FSW embedding makes PointNet significantly more robust to parameter reduction, with only minor performance degradation even after a 40-fold reduction. | Sliced Wasserstein distance, Euclidean embedding, multiset embedding, bi-Lipschitz, permutation invariant | A bi-Lipschitz Euclidean embedding for multisets that extends injectively to measures, based on the sliced Wasserstein distance | 1,401 | 2405.16519 | [
-0.009329038672149181,
-0.1214379072189331,
-0.020636852830648422,
-0.0747055783867836,
0.05315804108977318,
0.07584148645401001,
-0.04022951051592827,
0.018535083159804344,
0.04021847993135452,
-0.044195063412189484,
0.0096533028408885,
-0.033327873796224594,
0.001986495219171047,
0.10146789252758026,
-0.06695606559515,
-0.04408036544919014,
0.11917894333600998,
0.08079684525728226,
-0.14850910007953644,
0.006341604515910149,
-0.011009647510945797,
0.017411736771464348,
0.0260347668081522,
-0.028551839292049408,
0.04427570477128029,
-0.03763221949338913,
0.017338071018457413,
-0.08353609591722488,
-0.05051169544458389,
-0.0042651682160794735,
0.09279214590787888,
-0.001783930230885744,
0.027474289759993553,
0.0016412172699347138,
-0.04412327706813812,
0.05929429829120636,
0.0012853865046054125,
0.07271029055118561,
0.002438334748148918,
0.007050745189189911,
-0.05178302153944969,
0.09126494079828262,
0.060733914375305176,
0.03526489436626434,
-0.048952627927064896,
0.003456259146332741,
-0.024478567764163017,
-0.05999327078461647,
0.06293567270040512,
0.020636912435293198,
0.002110621426254511,
0.0772954449057579,
-0.0687611848115921,
0.09018932282924652,
0.03683719411492348,
-0.06437589228153229,
-0.014501221477985382,
-0.010064292699098587,
-0.0373595729470253,
0.009530723094940186,
-0.010600939393043518,
-0.10221987962722778,
0.0523659773170948,
-0.01542519312351942,
0.004977057687938213,
0.008119025267660618,
0.02071959525346756,
0.04398709908127785,
-0.035588912665843964,
0.03841223195195198,
-0.06293269246816635,
0.09530957788228989,
-0.1549176573753357,
-0.0074465591460466385,
0.060404468327760696,
0.002445919206365943,
-0.01061202585697174,
0.007260382175445557,
0.038498405367136,
-0.05009082704782486,
0.03530633822083473,
0.015428682789206505,
0.009350081905722618,
0.03201382979750633,
0.08291149884462357,
0.0765705406665802,
-0.04822621867060661,
0.001019096584059298,
0.04878422990441322,
-0.013907164335250854,
-0.04074961319565773,
0.02380681224167347,
0.013167035765945911,
0.043734703212976456,
-0.1067257672548294,
-0.04136216640472412,
-0.012500442564487457,
0.051058411598205566,
0.04343736544251442,
0.05660698935389519,
-0.0125833535566926,
-0.011841051280498505,
-0.0015625000232830644,
-0.0070454273372888565,
0.02358059398829937,
-0.12878243625164032,
0.044323522597551346,
0.008444342762231827,
-0.0167575404047966,
-0.09993353486061096,
0.09498583525419235,
-0.060956694185733795,
-0.006939567159861326,
0.0733867883682251,
0.029450884088873863,
-0.025275005027651787,
0.06024891138076782,
0.021584592759609222,
0.04526283219456673,
0.04824848845601082,
-0.06963735818862915,
0.007467363961040974,
0.011008085682988167,
0.05812501534819603,
0.06119006872177124,
0.07927685976028442,
0.047134868800640106,
3.0675120178722696e-33,
0.001638201531022787,
0.04504988715052605,
0.022324731573462486,
0.014621395617723465,
0.03477499634027481,
0.034187208861112595,
0.009540809318423271,
0.026298467069864273,
0.005849274341017008,
-0.0021932777017354965,
0.0007252543582580984,
0.04365330561995506,
0.006951279006898403,
0.10259833931922913,
0.05636567994952202,
-0.02111881598830223,
-0.0326649434864521,
-0.00025554868625476956,
0.008570417761802673,
-0.026294972747564316,
0.03997763991355896,
-0.029021456837654114,
-0.01031668670475483,
0.00291266106069088,
-0.0514666773378849,
-0.05034635588526726,
0.0017389978747814894,
-0.01877659745514393,
-0.0546068511903286,
0.020019719377160072,
-0.17130154371261597,
0.012702765874564648,
0.0033657681196928024,
-0.000009957423571904656,
0.03512213006615639,
-0.056187164038419724,
-0.018150757998228073,
0.015612226910889149,
-0.056885480880737305,
-0.11030761152505875,
0.01098426803946495,
0.021088475361466408,
-0.03922209143638611,
-0.07194329798221588,
-0.06845017522573471,
-0.026930058375000954,
0.026822436600923538,
0.029260069131851196,
0.026557082310318947,
-0.12052232772111893,
0.06826911121606827,
-0.00865467730909586,
0.0994333028793335,
-0.06889500468969345,
0.02261357381939888,
-0.030476678162813187,
-0.0034914331045001745,
0.010595387779176235,
0.09103090316057205,
0.06098315119743347,
-0.11056454479694366,
0.05459767207503319,
0.023421166464686394,
0.06368797272443771,
0.03908824548125267,
-0.010349060408771038,
0.000140422533149831,
-0.00029718130826950073,
0.10600768774747849,
0.02903647907078266,
0.0030491806101053953,
-0.025915058329701424,
-0.05559348687529564,
-0.06393200159072876,
-0.059800393879413605,
0.055356137454509735,
-0.013808837160468102,
-0.007628760766237974,
-0.002805096795782447,
0.03991539776325226,
0.0159506406635046,
0.007358679547905922,
0.016723277047276497,
-0.08627341687679291,
-0.03887210786342621,
-0.007958724163472652,
0.038657914847135544,
-0.09017467498779297,
-0.04973137006163597,
-0.06694211810827255,
-0.059145599603652954,
0.012568002566695213,
-0.02653702162206173,
-0.0360557995736599,
-0.002911999821662903,
-3.021769308248951e-33,
-0.03807312622666359,
0.05178418010473251,
-0.036508671939373016,
0.04062465578317642,
-0.002510568592697382,
0.015186424367129803,
-0.012119447812438011,
0.06470460444688797,
-0.003164947498589754,
-0.03899826109409332,
-0.028851617127656937,
-0.03863230720162392,
0.03337807580828667,
-0.01639835350215435,
-0.026770588010549545,
0.032926276326179504,
-0.015177788212895393,
-0.04186202213168144,
-0.025726504623889923,
-0.060125682502985,
0.007139527704566717,
-0.03596918284893036,
0.019299408420920372,
0.10412383079528809,
-0.050091471523046494,
0.0937880426645279,
-0.004848981276154518,
-0.01625261828303337,
-0.06077020615339279,
0.021087929606437683,
-0.03484705090522766,
-0.07774759829044342,
-0.03313945606350899,
0.00384720740839839,
-0.0500783734023571,
0.04458647966384888,
0.06520026177167892,
0.09004376828670502,
-0.0704590380191803,
0.018593527376651764,
-0.05848217383027077,
0.012541307136416435,
-0.038831885904073715,
0.032789502292871475,
-0.022562356665730476,
-0.008813970722258091,
-0.021557150408625603,
-0.022424349561333656,
-0.016669925302267075,
0.0305202417075634,
-0.10257487744092941,
0.06733821332454681,
-0.0013049066765233874,
0.11306332796812057,
-0.025314858183264732,
-0.05746212229132652,
-0.011994471773505211,
-0.040224768221378326,
0.08985798805952072,
-0.05513822287321091,
-0.04800672084093094,
-0.10351504385471344,
-0.049515996128320694,
0.04359269142150879,
0.041689950972795486,
-0.058972667902708054,
-0.0014653829857707024,
-0.08059332519769669,
-0.05662284046411514,
0.07225771248340607,
0.008758020587265491,
-0.067176952958107,
0.07777416706085205,
0.008451119996607304,
-0.06661073863506317,
0.012752469629049301,
0.08183979988098145,
-0.022480804473161697,
-0.057052552700042725,
0.10699224472045898,
-0.02336285263299942,
-0.007085595279932022,
0.029482195153832436,
0.04522588476538658,
0.07583120465278625,
-0.008962002582848072,
0.041426047682762146,
-0.05643882602453232,
0.044640764594078064,
-0.028301529586315155,
-0.01232330221682787,
0.04129790887236595,
-0.025416195392608643,
0.06359302997589111,
0.05921154469251633,
-4.1751583523819136e-8,
0.025816068053245544,
-0.0428069531917572,
-0.0052345916628837585,
-0.05986769497394562,
0.006037022918462753,
-0.052684634923934937,
0.014219899661839008,
0.10746975243091583,
-0.0874948725104332,
0.01901310309767723,
0.0535009540617466,
0.007349693216383457,
-0.06261258572340012,
0.05470787361264229,
-0.009596293792128563,
-0.04979877918958664,
0.018639782443642616,
0.06255220621824265,
-0.002531510777771473,
0.032836973667144775,
0.022971434518694878,
0.007508793380111456,
0.06366044282913208,
-0.06321060657501221,
-0.02346895821392536,
-0.006605151109397411,
-0.003856874303892255,
-0.04567212611436844,
0.05834641680121422,
0.024248341098427773,
-0.020600782707333565,
-0.011183901689946651,
0.10156279802322388,
-0.041218433529138565,
-0.056920960545539856,
0.024090953171253204,
-0.1109107956290245,
0.05240010470151901,
-0.11272302269935608,
-0.02645774744451046,
-0.002681708661839366,
0.06698554009199142,
-0.04859747365117073,
0.013795457780361176,
0.0921032726764679,
-0.01861993782222271,
0.02934802696108818,
0.07426460087299347,
0.03946271166205406,
0.11365542560815811,
-0.028528114780783653,
0.01447160542011261,
0.04767134040594101,
0.026985475793480873,
0.01583905890583992,
-0.02052823081612587,
-0.03858838230371475,
-0.04963361471891403,
0.023946547880768776,
-0.02239879034459591,
-0.002779322909191251,
-0.013860481791198254,
-0.0756092444062233,
-0.023605575785040855
] |
AutoEval: Autonomous Evaluation of LLMs for Truth Maintenance and Reasoning Tasks | https://openreview.net/forum?id=iv1TpRCJeK | [
"Rushang Karia",
"Daniel Richard Bramblett",
"Daksh Dobhal",
"Siddharth Srivastava"
] | Poster | This paper presents AutoEval, a novel benchmark for scaling Large Language Model (LLM) assessment in formal tasks with clear notions of correctness, such as truth maintenance in translation and logical reasoning. AutoEval is the first benchmarking paradigm that offers several key advantages necessary for scaling objective evaluation of LLMs without human labeling: (a) ability to evaluate LLMs of increasing sophistication by auto-generating tasks at different levels of difficulty; (b) auto-generation of ground truth that eliminates dependence on expensive and time-consuming human annotation; (c) the use of automatically generated, randomized datasets that mitigate the ability of successive LLMs to overfit to static datasets used in many contemporary benchmarks. Empirical analysis shows that an LLM's performance on AutoEval is highly indicative of its performance on a diverse array of other benchmarks focusing on translation and reasoning tasks, making it a valuable autonomous evaluation paradigm in settings where hand-curated datasets can be hard to obtain and/or update. | Large Language Models, Logical Reasoning, Autoformalization, Informalization, Formal Translation, Truth Maintenance | An automatic and scalable benchmark for evaluating LLMs for truth maintenance w.r.t. formal syntax. | 1,397 | 2410.08437 | [
-0.028162194415926933,
-0.027560735121369362,
-0.0353187657892704,
0.06707478314638138,
0.059888795018196106,
-0.053030095994472504,
0.009060267359018326,
0.012804427184164524,
-0.011274466291069984,
-0.0029717013239860535,
-0.06868983060121536,
-0.11575672775506973,
0.012540165334939957,
0.06577173620462418,
-0.0029615256935358047,
0.005154991056770086,
0.07220838218927383,
0.0012583204079419374,
-0.09943459182977676,
-0.08810129016637802,
-0.016406355425715446,
0.055198438465595245,
0.05235402286052704,
0.03700627014040947,
0.06440552324056625,
-0.03602847829461098,
-0.02503986842930317,
-0.046789318323135376,
0.0400809720158577,
-0.02274613454937935,
-0.030752647668123245,
0.016656555235385895,
0.015218842774629593,
0.049029525369405746,
0.004961366765201092,
0.05940481275320053,
-0.052586864680051804,
0.05650614574551582,
0.0023781200870871544,
-0.07791903614997864,
-0.06405530124902725,
-0.09815729409456253,
0.019297422841191292,
-0.0032134249340742826,
0.060528647154569626,
0.03981183469295502,
-0.04817478731274605,
0.022481994703412056,
-0.008227035403251648,
0.07447773218154907,
-0.10927960276603699,
0.007623468060046434,
0.002443182049319148,
-0.010141289792954922,
-0.05225467309355736,
0.060329511761665344,
0.05351889133453369,
0.02351478300988674,
-0.026507969945669174,
-0.07492443919181824,
-0.06764249503612518,
-0.10034981369972229,
-0.0649108812212944,
0.04558049887418747,
-0.013989709317684174,
-0.0298764668405056,
0.03087298944592476,
0.019650213420391083,
-0.015153137966990471,
0.03963376581668854,
-0.012381904758512974,
0.057602494955062866,
0.023229099810123444,
0.10930849611759186,
-0.021640973165631294,
0.05103045701980591,
-0.050472646951675415,
-0.05767595022916794,
0.05365375801920891,
-0.019741706550121307,
0.0010661501437425613,
0.008954767137765884,
0.04581163451075554,
0.09735899418592453,
0.055760014802217484,
0.007572574075311422,
0.0406467504799366,
0.023777760565280914,
0.01669931970536709,
0.10713593661785126,
0.009769492782652378,
-0.08437544852495193,
0.07487356662750244,
0.05226914584636688,
0.03946845605969429,
0.03626865893602371,
-0.006838517729192972,
-0.09405507892370224,
-0.03940611332654953,
0.05401036515831947,
0.032503530383110046,
0.14248688519001007,
-0.02450370416045189,
-0.09441359341144562,
0.009938038885593414,
-0.014403247274458408,
0.03578120842576027,
0.0007482623332180083,
0.02509365789592266,
-0.1051342785358429,
-0.001949345925822854,
0.04402909055352211,
0.054758112877607346,
-0.07746218144893646,
0.02726621739566326,
-0.01752912625670433,
0.02979784831404686,
-0.02098127268254757,
-0.02001543901860714,
-0.05658221244812012,
-0.053658634424209595,
0.023483047261834145,
0.07378428429365158,
0.017028601840138435,
0.05311218649148941,
-0.02178354188799858,
-0.0017271792748942971,
2.7216273969176194e-33,
0.02694724313914776,
0.04102291539311409,
0.02245989255607128,
0.053298454731702805,
-0.013767554424703121,
-0.040393222123384476,
0.008012142963707447,
0.08739045262336731,
-0.07802169024944305,
-0.02026478201150894,
-0.013274856843054295,
0.04105997458100319,
-0.06527075916528702,
0.030009692534804344,
0.007944437675178051,
-0.016732215881347656,
0.00924797635525465,
0.07914479821920395,
-0.07486316561698914,
-0.043636444956064224,
0.0671452060341835,
-0.015613947063684464,
0.03679896518588066,
-0.033497583121061325,
0.045138631016016006,
0.059202611446380615,
0.09663419425487518,
0.013549357652664185,
-0.05142098292708397,
0.005273872055113316,
-0.0721006914973259,
-0.08778572827577591,
0.07186947762966156,
0.06312258541584015,
0.01599293015897274,
-0.026657216250896454,
-0.019445069134235382,
-0.05769958347082138,
0.04858945310115814,
-0.008773557841777802,
0.012027938850224018,
0.05302537605166435,
0.009638193994760513,
0.01628969982266426,
-0.03881632536649704,
-0.030775373801589012,
-0.035305608063936234,
0.040543101727962494,
-0.032209064811468124,
-0.06999315321445465,
0.07794134318828583,
0.02464229241013527,
-0.03355645760893822,
0.008780626580119133,
-0.004487563855946064,
0.05037892982363701,
0.01436274778097868,
0.039987318217754364,
0.018553873524069786,
0.016735538840293884,
-0.007060617208480835,
0.006874169688671827,
-0.03073194809257984,
0.08801443874835968,
0.02863556705415249,
0.06293878704309464,
-0.02884969487786293,
0.043313298374414444,
0.06622850894927979,
-0.025573058053851128,
0.014820115640759468,
-0.020562615245580673,
0.015259199775755405,
0.00597586203366518,
0.05775269493460655,
0.01223201397806406,
0.0011705169454216957,
-0.0364423543214798,
0.020415164530277252,
0.011753748171031475,
0.026689371094107628,
0.02485615387558937,
-0.015985460951924324,
-0.04215072840452194,
0.024325009435415268,
-0.07039396464824677,
0.030654408037662506,
-0.07233667373657227,
-0.025138961151242256,
0.015590299852192402,
-0.03224599361419678,
-0.048070333898067474,
-0.06219751015305519,
-0.012778328731656075,
-0.03347807377576828,
-2.4211471636184417e-33,
-0.010722474195063114,
-0.0757722482085228,
-0.0710669457912445,
0.1551559418439865,
-0.0037091358099132776,
-0.07408144325017929,
-0.04731960594654083,
0.05664944648742676,
-0.002790851518511772,
-0.048793841153383255,
-0.07047149538993835,
-0.061178795993328094,
0.03126123175024986,
-0.03326457366347313,
-0.017344703897833824,
-0.047016292810440063,
-0.0338820144534111,
-0.0412374883890152,
-0.058470822870731354,
0.13540378212928772,
-0.057792969048023224,
0.10731759667396545,
-0.11229366064071655,
0.04529566690325737,
-0.003165221307426691,
0.026282649487257004,
-0.0416916199028492,
0.03214774280786514,
-0.0050005014054477215,
0.03506585210561752,
-0.00041666958713904023,
-0.007538233418017626,
-0.045063652098178864,
-0.03438112884759903,
-0.0710010677576065,
0.009970484301447868,
0.08772238343954086,
-0.017918309196829796,
-0.05118211731314659,
0.10367627441883087,
0.07301288098096848,
-0.005797088611871004,
-0.08565628528594971,
-0.0548793226480484,
-0.00858847051858902,
0.02536330558359623,
-0.15362460911273956,
-0.01764724589884281,
-0.016092222183942795,
-0.02624964714050293,
-0.024377910420298576,
-0.034737274050712585,
-0.027110211551189423,
0.0432913601398468,
0.068996362388134,
-0.09475845098495483,
-0.07249543815851212,
-0.06227445602416992,
-0.028215765953063965,
0.007663359399884939,
-0.08007583022117615,
0.0008332179859280586,
0.07914601266384125,
-0.03190038725733757,
-0.004146130755543709,
-0.004768765065819025,
-0.060717713087797165,
0.02486935444176197,
-0.06358535587787628,
0.04315074160695076,
0.07106482982635498,
-0.020729124546051025,
-0.057966120541095734,
-0.006589734926819801,
-0.0022520788479596376,
0.011855588294565678,
0.03326964005827904,
-0.03356548026204109,
0.05391745641827583,
-0.10394956916570663,
0.013431644067168236,
-0.025421401485800743,
0.038939595222473145,
0.08097686618566513,
-0.05916375294327736,
0.009784695692360401,
0.06536247581243515,
0.13915641605854034,
0.009310840629041195,
0.14446619153022766,
-0.016001850366592407,
0.01155961211770773,
-0.04617151990532875,
0.06641709804534912,
-0.06076241284608841,
-4.1430698871636196e-8,
-0.054777875542640686,
0.06151922792196274,
-0.004079369828104973,
0.05093475431203842,
0.014701198786497116,
-0.02398083545267582,
-0.05806804075837135,
0.11691179126501083,
0.0051101637072861195,
-0.05549808219075203,
0.014300406910479069,
-0.03290831670165062,
-0.050264280289411545,
-0.042291928082704544,
-0.039582137018442154,
0.049691639840602875,
0.03263010457158089,
0.06297937035560608,
-0.025163471698760986,
-0.02805064618587494,
0.04186994954943657,
0.06207820400595665,
-0.03786731883883476,
-0.013720321469008923,
0.08766819536685944,
-0.04577382281422615,
-0.07683881372213364,
0.04541216790676117,
0.03783319890499115,
-0.015939464792609215,
-0.00974240805953741,
-0.0030822071712464094,
0.023494170978665352,
-0.00682209013029933,
0.09019839763641357,
0.06792246550321579,
-0.030540671199560165,
-0.006042033899575472,
-0.002086860593408346,
-0.027568893507122993,
0.04907280579209328,
0.0928676426410675,
-0.07303853332996368,
0.03291604295372963,
-0.019993534311652184,
-0.029888655990362167,
-0.08084553480148315,
-0.048309262841939926,
-0.006523819640278816,
-0.07969539612531662,
0.007360470946878195,
-0.08759656548500061,
-0.02746771089732647,
0.021744148805737495,
0.05738379433751106,
0.06521337479352951,
-0.014504911378026009,
-0.016998786479234695,
0.05174282193183899,
0.04641086235642433,
0.029472466558218002,
0.001254882663488388,
-0.04025447741150856,
0.009703326970338821
] |
Efficient Action-Constrained Reinforcement Learning via Acceptance-Rejection Method and Augmented MDPs | https://openreview.net/forum?id=AgMpK7z4bz | [
"Wei Hung",
"Shao-Hua Sun",
"Ping-Chun Hsieh"
] | Poster | Action-constrained reinforcement learning (ACRL) is a generic framework for learning control policies with zero action constraint violation, which is required by various safety-critical and resource-constrained applications. The existing ACRL methods can typically achieve favorable constraint satisfaction but at the cost of either high computational burden incurred by the quadratic programs (QP) or increased architectural complexity due to the use of sophisticated generative models. In this paper, we propose a generic and computationally efficient framework that can adapt a standard unconstrained RL method to ACRL through two modifications: (i) To enforce the action constraints, we leverage the classic acceptance-rejection method, where we treat the unconstrained policy as the proposal distribution and derive a modified policy with feasible actions. (ii) To improve the acceptance rate of the proposal distribution, we construct an augmented two-objective Markov decision process (MDP), which include additional self-loop state transitions and a penalty signal for the rejected actions. This augmented MDP incentives the learned policy to stay close to the feasible action sets. Through extensive experiments in both robot control and resource allocation domains, we demonstrate that the proposed framework enjoys faster training progress, better constraint satisfaction, and a lower action inference time simultaneously than the state-of-the-art ACRL methods. We have made the source code publicly available to encourage further research in this direction. | Reinforcement learning, action constraints | null | 1,396 | 2503.12932 | [
-0.03136306256055832,
-0.026255609467625618,
-0.02238314412534237,
-0.01101450715214014,
0.018383771181106567,
0.061647407710552216,
0.001960953464731574,
-0.04283391684293747,
-0.016962118446826935,
0.09444516152143478,
0.012243976816534996,
-0.04133183881640434,
0.06270778179168701,
0.040272679179906845,
0.010722599923610687,
0.06040188670158386,
0.10801081359386444,
0.015004503540694714,
-0.012576811946928501,
0.026714880019426346,
0.04606267437338829,
0.03650864213705063,
0.08950495719909668,
-0.035799361765384674,
-0.029639437794685364,
0.0018022522563114762,
0.03859146684408188,
-0.04768060892820358,
-0.0053673554211854935,
-0.03091152384877205,
0.0641280934214592,
-0.07911446690559387,
0.00007257983816089109,
-0.0920572429895401,
0.009379150345921516,
0.07973071187734604,
-0.08072713762521744,
-0.00969403050839901,
-0.08536428213119507,
-0.06730452924966812,
-0.029654255136847496,
0.02950248308479786,
0.014399152249097824,
-0.020997993648052216,
0.040231723338365555,
0.029625121504068375,
-0.024460740387439728,
-0.029636211693286896,
0.04527609422802925,
-0.038872990757226944,
0.05210159346461296,
0.0006855207611806691,
0.01698596589267254,
0.018544843420386314,
0.0025996314361691475,
0.021226538345217705,
0.01370027381926775,
-0.004596843849867582,
0.006868818774819374,
-0.028673557564616203,
-0.0401049442589283,
-0.10432589799165726,
-0.07163722068071365,
-0.0757393091917038,
-0.11354584991931915,
-0.017094673588871956,
-0.062288522720336914,
-0.018239179626107216,
0.010990307666361332,
0.0684809610247612,
-0.04704153165221214,
0.014033704996109009,
-0.004375454504042864,
-0.008589808829128742,
0.009178114123642445,
0.004503356758505106,
-0.0023379570338875055,
-0.04548422247171402,
0.04055032506585121,
-0.0055910092778503895,
-0.010017553344368935,
-0.02558295615017414,
-0.005682650953531265,
0.03631015121936798,
0.1050540879368782,
-0.04453670233488083,
-0.001209499780088663,
0.0769663080573082,
0.14105463027954102,
0.02875109761953354,
0.04249456897377968,
0.03360657021403313,
0.010879924520850182,
-0.012347585521638393,
0.014088648371398449,
-0.04035944491624832,
0.03665981441736221,
-0.07966326177120209,
-0.1147308349609375,
0.0676443874835968,
0.0388079509139061,
0.018668191507458687,
-0.043641578406095505,
-0.040892161428928375,
0.05582331493496895,
-0.08732923865318298,
-0.04617782682180405,
-0.007594552356749773,
0.05788934603333473,
-0.05895693227648735,
0.047836191952228546,
-0.05318967252969742,
0.041169844567775726,
-0.013169441372156143,
-0.052858125418424606,
0.08308563381433487,
0.011006577871739864,
-0.05421418696641922,
-0.02720584347844124,
0.02753317728638649,
-0.08761181682348251,
-0.06461399793624878,
0.0590670146048069,
0.0341377779841423,
0.044280003756284714,
-0.10276079922914505,
-0.09938699752092361,
-2.5831830446542957e-34,
-0.0020397005137056112,
0.01597987487912178,
0.02110634744167328,
-0.05035479739308357,
0.019038556143641472,
0.037011709064245224,
0.05058736726641655,
0.017595572397112846,
0.016540078446269035,
-0.07190515846014023,
0.025439754128456116,
-0.052223436534404755,
-0.06135530024766922,
0.05222618207335472,
0.09745167195796967,
-0.01747158356010914,
-0.031026821583509445,
0.051564764231443405,
0.0138781713321805,
-0.03204076737165451,
0.028936415910720825,
-0.009780749678611755,
0.0011747279204428196,
0.013122128322720528,
0.03529602661728859,
-0.026147380471229553,
0.03295154124498367,
0.01921861805021763,
0.016071846708655357,
0.009877147153019905,
-0.061875782907009125,
0.042274340987205505,
-0.007431108970195055,
0.06696105003356934,
0.027745679020881653,
0.0009731908212415874,
-0.0833306685090065,
-0.026325616985559464,
-0.018603306263685226,
-0.03304491564631462,
-0.04111265018582344,
-0.018119409680366516,
0.07971666008234024,
0.003790424671024084,
-0.061080481857061386,
-0.08492473512887955,
0.09521424770355225,
-0.00028475906583480537,
-0.035303425043821335,
-0.046458899974823,
0.029280561953783035,
0.08203016221523285,
0.013540777377784252,
-0.0015796718653291464,
-0.05407347157597542,
0.012094882316887379,
-0.013814738020300865,
0.11522240191698074,
-0.04179944097995758,
0.04187610372900963,
0.041179776191711426,
0.0028348101768642664,
0.04798949509859085,
0.09796827286481857,
-0.01810835674405098,
0.05861380696296692,
-0.028802253305912018,
0.057751353830099106,
0.10202055424451828,
-0.02522227168083191,
-0.005425911862403154,
-0.009182886220514774,
0.0746527761220932,
-0.07181674242019653,
0.05611753091216087,
0.00832289643585682,
0.06438665091991425,
-0.07255388051271439,
0.017431998625397682,
0.003924965858459473,
-0.036479879170656204,
0.05138251930475235,
-0.046388179063797,
0.04387733340263367,
0.007577815093100071,
-0.10062408447265625,
0.02712947502732277,
0.014116959646344185,
-0.03782448172569275,
-0.025518115609884262,
-0.10017736256122589,
-0.025576632469892502,
-0.06718044728040695,
0.10025738179683685,
0.08337784558534622,
-7.6028573112961025e-34,
0.11215332895517349,
-0.03993471339344978,
-0.014831402339041233,
-0.04148712754249573,
0.02281465008854866,
-0.029030196368694305,
-0.0391988679766655,
-0.05852636694908142,
0.07030456513166428,
-0.07905346900224686,
-0.048175372183322906,
0.0021209069527685642,
0.01954096183180809,
0.02523750811815262,
-0.0040351650677621365,
-0.040717363357543945,
-0.06596659123897552,
0.07273689657449722,
-0.009150460362434387,
0.05116524174809456,
-0.01092618703842163,
0.07525289803743362,
-0.07549187541007996,
0.05124320089817047,
0.005835200659930706,
0.024081815034151077,
-0.026138843968510628,
0.030290158465504646,
0.030673138797283173,
0.1000373587012291,
-0.07632292062044144,
0.022979306057095528,
-0.07165592163801193,
0.027533259242773056,
-0.004107206128537655,
-0.0020388804841786623,
0.05126264691352844,
0.04141043499112129,
-0.024999195709824562,
0.07693354785442352,
0.08216980844736099,
-0.0791090875864029,
-0.013628208078444004,
0.06836210936307907,
0.08750537037849426,
-0.05219438299536705,
0.003656038548797369,
-0.015416973270475864,
-0.032601743936538696,
-0.005109368357807398,
-0.07358658313751221,
-0.0016263978322967887,
-0.0702248215675354,
0.016133036464452744,
-0.019559863954782486,
-0.08377891033887863,
-0.01568656414747238,
0.0337001197040081,
0.12160434573888779,
-0.0001667104079388082,
-0.07993034273386002,
0.03802606090903282,
-0.031097587198019028,
0.052511751651763916,
0.015324133448302746,
0.04869589954614639,
0.02092168666422367,
-0.0019609504379332066,
0.03915474936366081,
0.025093162432312965,
0.025105968117713928,
0.06566435843706131,
0.05515686795115471,
0.013860983774065971,
0.08486362546682358,
-0.010738866403698921,
0.010036985389888287,
-0.04708923399448395,
0.01860472746193409,
-0.09357108920812607,
0.02377956733107567,
0.06484426558017731,
0.037920549511909485,
0.1029745563864708,
-0.026181044057011604,
0.033072732388973236,
-0.037546124309301376,
0.014173577539622784,
0.018855730071663857,
0.014007173478603363,
-0.005779270548373461,
-0.012695994228124619,
0.03211703896522522,
0.06551021337509155,
-0.1133042648434639,
-4.452361324069898e-8,
-0.01706625148653984,
0.03267938271164894,
0.058002304285764694,
-0.04066016152501106,
0.1067010760307312,
0.057012367993593216,
-0.03112044930458069,
-0.019837046042084694,
0.03138522803783417,
-0.026578929275274277,
0.051465921103954315,
0.05680552124977112,
0.03871715068817139,
-0.08896726369857788,
-0.14337097108364105,
0.03992934897542,
-0.0009094350971281528,
-0.04763916879892349,
0.003220951184630394,
-0.019534539431333542,
-0.026603788137435913,
-0.03937521576881409,
-0.0001899896888062358,
-0.03415024280548096,
-0.009514130651950836,
-0.06255196779966354,
-0.03548778221011162,
0.054903943091630936,
0.08913543820381165,
0.007017588242888451,
0.012949378229677677,
-0.0785011276602745,
0.059070881456136703,
0.046615876257419586,
0.017277171835303307,
0.04181695356965065,
0.006084207445383072,
-0.07330875843763351,
-0.03105175867676735,
-0.07295890152454376,
0.02670738287270069,
0.10925810784101486,
0.0174446739256382,
-0.03107510507106781,
-0.03450491651892662,
0.07038262486457825,
-0.09418872743844986,
-0.1472776234149933,
0.03890696540474892,
-0.031899694353342056,
0.016435997560620308,
0.0005874848575331271,
-0.017353437840938568,
0.05277194082736969,
-0.005196764599531889,
0.010621367953717709,
0.03514716401696205,
-0.11947526782751083,
0.018236622214317322,
0.08857424557209015,
0.012162094004452229,
-0.00845405925065279,
-0.03995214030146599,
0.018854282796382904
] |
TeaserGen: Generating Teasers for Long Documentaries | https://openreview.net/forum?id=G1n50BMqzm | [
"Weihan Xu",
"Paul Pu Liang",
"Haven Kim",
"Julian McAuley",
"Taylor Berg-Kirkpatrick",
"Hao-Wen Dong"
] | Poster | Teasers are an effective tool for promoting content in entertainment, commercial and educational fields. However, creating an effective teaser for long videos is challenging for it requires long-range multimodal modeling capability for the input videos, while necessitating maintaining audiovisual alignments, managing scene transitions and preserving factual accuracy for the output teasers. Due to the lack of a publicly-available dataset, progress along this research direction has been hindered. In this work, we present DocumentaryNet, a collection of 1,269 documentaries paired with their teasers, featuring multimodal data streams of video, speech, music, sound effects and narrations. With DocumentaryNet, we propose a new two-stage system for generating teasers from long documentaries. The proposed TeaserGen system first generates the teaser narration from the transcribed narration from the documentary using a pretrained large language model, and then selects the most relevant visual content to accompany the generated narration through language-vision models. For narration-video matching, we explore two approaches: a pretraining-based model using pretrained contrastive language-vision models and a deep sequential model that learns the mapping between the narrations and visuals. Our experimental results show that the pretraining-based approach is more effective at identifying relevant visual content than directly trained deep autoregressive models. | Teaser Generation, Multimodal Learning, Vision-Language Model | null | 1,391 | 2410.05586 | [
-0.03499023616313934,
-0.09890300780534744,
0.0029060770757496357,
-0.03621501475572586,
0.1292971670627594,
0.09836485236883163,
-0.07745946198701859,
-0.020784590393304825,
0.05382243171334267,
-0.07669180631637573,
-0.052440568804740906,
-0.06811469048261642,
-0.016624918207526207,
0.07751478999853134,
-0.016544204205274582,
-0.03323835879564285,
0.03856861591339111,
0.04922815039753914,
-0.018994539976119995,
-0.015538470819592476,
0.0656484067440033,
0.03445621579885483,
0.0410027951002121,
0.000777245091740042,
0.07814744114875793,
0.016325484961271286,
-0.00958387553691864,
-0.0019022320630028844,
0.06581465154886246,
-0.03361351415514946,
0.02661101520061493,
0.04019060730934143,
0.05797571688890457,
0.05623374506831169,
-0.026769645512104034,
0.045136451721191406,
-0.024050157517194748,
0.025726720690727234,
-0.07277199625968933,
-0.026512062177062035,
0.008762873709201813,
0.015387275256216526,
-0.021476689726114273,
-0.017951825633645058,
0.030542364344000816,
-0.031846094876527786,
-0.02348383329808712,
-0.0074294814839959145,
0.07159954309463501,
0.08233693987131119,
-0.07879779487848282,
-0.005933159030973911,
0.0825626477599144,
0.04329783096909523,
-0.05413945019245148,
-0.011441639624536037,
0.05180807411670685,
-0.03619636595249176,
0.03394380211830139,
-0.051914896816015244,
-0.008012295700609684,
-0.1081286296248436,
0.013098424300551414,
-0.05602984130382538,
0.03482169657945633,
-0.09887760132551193,
0.053864963352680206,
0.035616252571344376,
0.022972628474235535,
0.011670608073472977,
-0.08074742555618286,
0.03014509566128254,
0.046984218060970306,
0.08052391558885574,
-0.012708883732557297,
0.061060771346092224,
0.06051937863230705,
-0.09451588988304138,
-0.04440269619226456,
-0.09466897696256638,
0.04165195673704147,
-0.04340417683124542,
0.09201321005821228,
-0.05771595984697342,
0.023429205641150475,
0.004908471368253231,
-0.006924177519977093,
0.007304580416530371,
-0.05658567324280739,
0.051998354494571686,
-0.16240109503269196,
-0.0062272027134895325,
0.08063444495201111,
-0.007111141923815012,
0.04476076737046242,
0.07959915697574615,
0.0321778766810894,
-0.11893698573112488,
0.05719174072146416,
0.06692750751972198,
0.04592595994472504,
0.01020679622888565,
0.025371836498379707,
-0.07764061540365219,
0.0035030192229896784,
-0.0754244402050972,
0.039648525416851044,
-0.0005481127300299704,
0.015347138047218323,
-0.04637257382273674,
0.050874583423137665,
0.03735873103141785,
0.05034065619111061,
-0.036184120923280716,
0.05374395102262497,
0.020321102812886238,
0.00005265452273306437,
0.015301468782126904,
-0.013533567078411579,
0.054024383425712585,
0.04858040064573288,
0.029727941378951073,
-0.03674156963825226,
0.020507538691163063,
0.0009278883226215839,
-0.035684410482645035,
0.05320599675178528,
-9.48224548540795e-34,
0.0821872428059578,
0.030379097908735275,
0.044185392558574677,
0.07261329144239426,
0.07244755327701569,
-0.041481874883174896,
-0.044675279408693314,
0.0720909833908081,
-0.04071204736828804,
-0.08495061099529266,
-0.006242582108825445,
-0.03457808867096901,
-0.047363776713609695,
0.06049136072397232,
-0.07634817063808441,
-0.016920220106840134,
-0.074006088078022,
0.026552220806479454,
0.017998378723859787,
0.08322762697935104,
-0.04642752930521965,
-0.032709237188100815,
0.029485268518328667,
0.016357293352484703,
0.07131236046552658,
-0.0027604722417891026,
0.0706201046705246,
-0.09828000515699387,
-0.02241666242480278,
-0.018264926970005035,
-0.12235845625400543,
-0.022710638120770454,
0.0709473192691803,
-0.03614110127091408,
0.08352657407522202,
-0.050793930888175964,
-0.02297978289425373,
-0.026051009073853493,
0.016328442841768265,
-0.05769556760787964,
-0.055982764810323715,
-0.031701359897851944,
-0.045745283365249634,
-0.02184215560555458,
-0.06918547302484512,
-0.029849248006939888,
0.026410669088363647,
0.06881891936063766,
-0.07274572551250458,
0.05004461482167244,
0.0773286372423172,
0.042721476405858994,
-0.04102439805865288,
-0.11113143712282181,
-0.02035890705883503,
-0.036564357578754425,
0.06120605021715164,
-0.05632542446255684,
0.033380407840013504,
-0.06722643226385117,
-0.005869770422577858,
0.09045985341072083,
0.031229887157678604,
0.04616380110383034,
0.02527407743036747,
0.04518699645996094,
0.03467639908194542,
0.05526791885495186,
0.09518244117498398,
0.045589372515678406,
-0.0965687707066536,
0.0021212627179920673,
0.03480585291981697,
-0.026713797822594643,
-0.013290323317050934,
0.10208016633987427,
-0.01153334230184555,
-0.05700189620256424,
0.007001488469541073,
0.06485460698604584,
-0.023304330185055733,
-0.001186109147965908,
0.04674598574638367,
-0.05159175023436546,
-0.04770970344543457,
-0.00708096893504262,
0.01815664954483509,
-0.09045606851577759,
0.03384542092680931,
-0.06414517760276794,
-0.025948675349354744,
0.01511333603411913,
-0.04699258878827095,
0.011062917299568653,
0.055313579738140106,
1.2680655411400004e-33,
0.06342670321464539,
0.1010504886507988,
-0.0540371909737587,
-0.02970639429986477,
-0.009271752089262009,
-0.020051121711730957,
0.004619320388883352,
0.05384649336338043,
-0.022494113072752953,
-0.05562068894505501,
-0.010107580572366714,
-0.07244511693716049,
-0.012361369095742702,
-0.06630223989486694,
-0.01407658588141203,
-0.02012602984905243,
0.06789632141590118,
-0.019162127748131752,
-0.0350327268242836,
0.03204214200377464,
0.04132070764899254,
0.04132549464702606,
-0.09704659134149551,
-0.0299505814909935,
-0.07823199778795242,
-0.00779740372672677,
-0.011875372380018234,
0.016870228573679924,
-0.03168383240699768,
-0.02033645287156105,
0.03282593563199043,
-0.06321115046739578,
-0.007607067469507456,
-0.026314735412597656,
-0.07397247105836868,
0.02711779996752739,
0.07734814286231995,
-0.04871257022023201,
-0.01649574004113674,
0.0697898119688034,
0.08005046844482422,
-0.023875564336776733,
-0.025114960968494415,
-0.0411488376557827,
-0.07855811715126038,
-0.07148398458957672,
-0.03975566849112511,
0.01946825161576271,
-0.000757694011554122,
-0.0020337922032922506,
-0.046798426657915115,
0.020394502207636833,
-0.055835846811532974,
-0.0418512187898159,
0.009569588117301464,
-0.13307926058769226,
-0.09363503009080887,
0.028473297134041786,
0.07002388685941696,
-0.020071115344762802,
-0.04209998995065689,
-0.04728187993168831,
-0.03802208974957466,
-0.06244294345378876,
0.018398266285657883,
-0.026495633646845818,
-0.014922826550900936,
-0.01790894940495491,
0.01696016453206539,
-0.01568656414747238,
0.06757119297981262,
-0.014118689112365246,
0.050906792283058167,
0.05843333527445793,
0.0016350713558495045,
0.023851383477449417,
0.033163927495479584,
-0.01214563474059105,
-0.008755117654800415,
-0.07986203581094742,
-0.007915935479104519,
-0.05825423076748848,
0.023301269859075546,
0.04013465717434883,
0.07910680770874023,
0.14473189413547516,
0.0388590469956398,
0.1054479330778122,
0.03511570394039154,
0.11921804398298264,
-0.033950075507164,
0.05439922586083412,
0.034073587507009506,
0.03730180487036705,
0.03660621866583824,
-3.2412518891078435e-8,
-0.0733630433678627,
0.038487933576107025,
-0.023837365210056305,
-0.006688861176371574,
-0.028070712462067604,
-0.02646135911345482,
0.048338882625103,
0.03539450839161873,
0.06824375689029694,
-0.07944319397211075,
0.010778208263218403,
0.027660837396979332,
0.030558766797184944,
-0.02939138561487198,
-0.0768570601940155,
0.034905221313238144,
-0.03822958469390869,
0.01870073564350605,
0.011775209568440914,
-0.020909402519464493,
0.0457463376224041,
0.0422801673412323,
-0.02084757387638092,
-0.020858943462371826,
-0.008867481723427773,
0.00845708791166544,
0.03948824107646942,
0.0345410481095314,
0.03834143280982971,
-0.01821374148130417,
-0.0459507517516613,
0.029921548441052437,
-0.06351587176322937,
-0.060585737228393555,
0.007878278382122517,
0.01350394356995821,
0.0009107504156418145,
0.012143325991928577,
0.02770255319774151,
-0.008682372979819775,
0.03180177882313728,
0.01965060643851757,
-0.03886030986905098,
0.018753819167613983,
-0.03274320065975189,
0.029354549944400787,
-0.0025189295411109924,
-0.10909836739301682,
0.07377929240465164,
-0.0282520093023777,
-0.047369446605443954,
-0.09629456698894501,
0.05261784791946411,
0.049157559871673584,
0.08995779603719711,
0.026476161554455757,
-0.011613107286393642,
0.07349478453397751,
0.02448822557926178,
0.06877235323190689,
-0.023506999015808105,
0.014561088755726814,
-0.007059508468955755,
0.01936214603483677
] |
CameraCtrl: Enabling Camera Control for Video Diffusion Models | https://openreview.net/forum?id=Z4evOUYrk7 | [
"Hao He",
"Yinghao Xu",
"Yuwei Guo",
"Gordon Wetzstein",
"Bo Dai",
"Hongsheng Li",
"Ceyuan Yang"
] | Poster | Controllability plays a crucial role in video generation, as it allows users to create and edit content more precisely. Existing models, however, lack control of camera pose that serves as a cinematic language to express deeper narrative nuances. To alleviate this issue, we introduce \method, enabling accurate camera pose control for video diffusion models. Our approach explores effective camera trajectory parameterization along with a plug-and-play camera pose control module that is trained on top of a video diffusion model, leaving other modules of the base model untouched. Moreover, a comprehensive study on the effect of various training datasets is conducted, suggesting that videos with diverse camera distributions and similar appearance to the base model indeed enhance controllability and generalization. Experimental results demonstrate the effectiveness of \method in achieving precise camera control with different video generation models, marking a step forward in the pursuit of dynamic and customized video storytelling from textual and camera pose inputs. | camera viewpoints control in Video Generation | null | 1,390 | null | [
0.02170647494494915,
-0.13294929265975952,
0.0011296861339360476,
-0.04569419473409653,
0.0673593208193779,
0.028620434924960136,
-0.022806527093052864,
-0.008099943399429321,
0.07591591775417328,
-0.028618695214390755,
0.02441857010126114,
-0.015370186418294907,
-0.009459196589887142,
0.07587578147649765,
-0.01777159608900547,
-0.08468925952911377,
0.0787554681301117,
0.06130919232964516,
-0.11371269077062607,
0.018480004742741585,
-0.029312055557966232,
-0.07149006426334381,
0.07537465542554855,
-0.027097627520561218,
0.01127238292247057,
0.005948350764811039,
0.021214988082647324,
-0.020576825365424156,
0.046814173460006714,
-0.05088721588253975,
0.05867414176464081,
-0.04095873609185219,
-0.025749316439032555,
0.036358222365379333,
-0.07624851167201996,
0.025208650156855583,
-0.019637353718280792,
-0.021742355078458786,
-0.051653891801834106,
-0.018990321084856987,
0.024156395345926285,
0.08838950842618942,
0.07108015567064285,
0.015494510531425476,
0.07504287362098694,
0.05442172661423683,
-0.009403811767697334,
0.009510408155620098,
-0.03527950122952461,
0.026257796213030815,
-0.07932977378368378,
-0.035264693200588226,
0.04951540380716324,
0.03118583746254444,
-0.034807417541742325,
-0.002447386272251606,
0.03235103189945221,
-0.009035690687596798,
0.07576632499694824,
-0.050127264112234116,
-0.04972575977444649,
-0.02730689011514187,
-0.022969696670770645,
0.04048258066177368,
-0.025394277647137642,
-0.03364141657948494,
0.11481905728578568,
0.03024423122406006,
-0.009819725528359413,
0.05052582547068596,
-0.08355657011270523,
0.04641617462038994,
0.015523376874625683,
0.01585610955953598,
-0.025967927649617195,
0.015036549419164658,
0.015216308645904064,
-0.08522740751504898,
-0.03305138647556305,
-0.09148412942886353,
0.12182427197694778,
-0.028878966346383095,
0.035268764942884445,
0.051311343908309937,
-0.019474707543849945,
-0.03455827385187149,
-0.006643633358180523,
0.023785710334777832,
0.009533625096082687,
0.01848638989031315,
-0.12319070100784302,
-0.01728087104856968,
0.021395130082964897,
-0.09489170461893082,
-0.030035117641091347,
0.03622298315167427,
-0.027714373543858528,
-0.0803363248705864,
-0.016749555245041847,
0.006404341198503971,
0.014913564547896385,
-0.03907092288136482,
0.04178658872842789,
0.024151505902409554,
0.09109821915626526,
-0.04912431165575981,
0.041085485368967056,
0.013662029057741165,
0.007267592940479517,
0.008135628886520863,
0.0213626679033041,
0.021268954500555992,
-0.05814707279205322,
0.007126603275537491,
0.08346822112798691,
0.010350199416279793,
0.0049108765088021755,
-0.02362598478794098,
-0.05814658850431442,
0.004843619652092457,
0.0810813382267952,
-0.043406520038843155,
-0.08815811574459076,
-0.032633256167173386,
0.024791430681943893,
-0.011890681460499763,
0.06839486956596375,
1.4721769108488439e-33,
0.049054697155952454,
0.007416711188852787,
0.06574410945177078,
0.0646921768784523,
0.07642704248428345,
0.031394630670547485,
0.0041415938176214695,
0.04427214711904526,
-0.04849318414926529,
-0.08147785067558289,
0.030942687764763832,
-0.09559308737516403,
-0.0718950480222702,
0.08589742332696915,
0.05470968782901764,
-0.0119471475481987,
-0.09162133187055588,
0.01024739257991314,
0.02433576062321663,
0.12516459822654724,
0.051659297198057175,
0.004824627190828323,
0.0271757859736681,
0.01289473008364439,
-0.06732099503278732,
0.02173156477510929,
0.06286558508872986,
0.020061278715729713,
-0.0019576800987124443,
-0.018551774322986603,
-0.06951628625392914,
-0.04361383244395256,
-0.01173144206404686,
-0.015071481466293335,
0.11586771160364151,
0.02451365999877453,
-0.05702098831534386,
-0.04754403606057167,
0.05995482951402664,
0.0293642058968544,
-0.01061636209487915,
0.04642359912395477,
-0.1178397387266159,
0.014056592248380184,
-0.06665118783712387,
0.014274791814386845,
0.024746347218751907,
0.10750047117471695,
-0.11749286204576492,
0.03552507981657982,
0.038307953625917435,
0.030628925189375877,
0.044194575399160385,
-0.116883784532547,
-0.05023197829723358,
0.01210088562220335,
0.047770388424396515,
-0.0856366753578186,
-0.044897351413965225,
-0.05682101845741272,
0.05631715804338455,
-0.012178212404251099,
-0.018266592174768448,
0.013554507866501808,
0.030164198949933052,
0.009311293251812458,
0.03636067360639572,
0.08188235014677048,
0.051881421357393265,
0.02519463188946247,
-0.045355215668678284,
-0.004810453858226538,
-0.050495050847530365,
-0.028066065162420273,
0.038975849747657776,
0.03367913141846657,
-0.024752097204327583,
-0.03760295733809471,
-0.013690236024558544,
0.09782858937978745,
-0.08689917623996735,
-0.030403589829802513,
0.025399629026651382,
-0.003087927820160985,
-0.02944939211010933,
-0.07860531657934189,
0.003684234106913209,
-0.01390374731272459,
-0.04978913813829422,
-0.08160006254911423,
0.010063867084681988,
-0.0011145701864734292,
-0.00001597063965164125,
0.013507653959095478,
0.032276321202516556,
-6.9091457486430605e-34,
0.03237689658999443,
0.0773659497499466,
-0.026026735082268715,
0.049380842596292496,
-0.0402679368853569,
-0.061181943863630295,
0.03939647972583771,
0.054354842752218246,
0.10731511563062668,
-0.044436704367399216,
-0.04269194230437279,
-0.06198358163237572,
-0.08307844400405884,
-0.06958846002817154,
0.027350548654794693,
-0.0030100627336651087,
0.0057714576832950115,
-0.0528668612241745,
-0.08695008605718613,
-0.025832686573266983,
-0.010314675979316235,
0.03316289186477661,
0.017333198338747025,
0.0107090650126338,
-0.01869368739426136,
0.037572044879198074,
0.016894765198230743,
0.11037260293960571,
0.07632740586996078,
-0.05680600553750992,
-0.0012057643616572022,
-0.03127267211675644,
0.048140112310647964,
-0.02198675647377968,
-0.13626909255981445,
0.076592355966568,
0.03910699114203453,
-0.013780366629362106,
-0.003448223927989602,
0.10357527434825897,
-0.011189058423042297,
-0.04447649419307709,
-0.05760334059596062,
0.04508383944630623,
0.012803192250430584,
0.014632560312747955,
-0.024778414517641068,
0.02941657416522503,
-0.008824124932289124,
0.02534925378859043,
-0.07844357192516327,
0.08213990181684494,
-0.061716921627521515,
0.017599742859601974,
-0.015235877595841885,
-0.08588538318872452,
-0.08641958236694336,
0.0038528030272573233,
0.09493985027074814,
-0.02912084199488163,
-0.004187840037047863,
0.04903874546289444,
-0.01830224320292473,
-0.039189815521240234,
-0.020694712176918983,
-0.040658194571733475,
-0.058705613017082214,
-0.023044878616929054,
-0.016337303444743156,
0.03882361575961113,
0.0341121144592762,
0.0016062872018665075,
0.041824012994766235,
0.024905690923333168,
-0.008688847534358501,
-0.054179102182388306,
-0.010752227157354355,
-0.006219444330781698,
0.051677823066711426,
-0.07438509911298752,
0.03315633162856102,
-0.06670605391263962,
0.057448312640190125,
0.021403785794973373,
0.047119174152612686,
0.1019587442278862,
0.011978925205767155,
0.0341978594660759,
-0.01688002049922943,
0.0843428298830986,
-0.012377285398542881,
-0.0166944433003664,
0.03114502690732479,
0.09869526326656342,
-0.1041814386844635,
-3.707987517032052e-8,
-0.12523019313812256,
-0.0029997979290783405,
0.035352353006601334,
0.04932349920272827,
-0.079645536839962,
-0.052519217133522034,
0.11136519908905029,
0.06056685745716095,
0.0791521966457367,
-0.05703499913215637,
0.013252892531454563,
0.03581864386796951,
0.051502447575330734,
0.013082914054393768,
-0.0683034285902977,
0.03666763752698898,
0.0018190714763477445,
0.04514926299452782,
-0.036691900342702866,
0.07492713630199432,
0.0043329354375600815,
0.0012165383668616414,
-0.019312778487801552,
-0.04022514820098877,
0.011096013709902763,
0.0008276086300611496,
-0.043883033096790314,
0.024418117478489876,
0.09079382568597794,
-0.023806683719158173,
-0.005563669838011265,
0.01426231861114502,
0.019670579582452774,
-0.027138618752360344,
0.0014819939387962222,
0.06627387553453445,
-0.02195642702281475,
-0.029083868488669395,
-0.015695294365286827,
0.0032333028502762318,
-0.02835950441658497,
0.030503153800964355,
0.05943073704838753,
0.04801678657531738,
-0.03391081839799881,
0.10472526401281357,
0.058629393577575684,
-0.06801215559244156,
0.05044351518154144,
-0.005268698558211327,
-0.06710463017225266,
-0.009104963392019272,
-0.03184284269809723,
0.07688454538583755,
0.06138204410672188,
-0.016526538878679276,
0.07887641340494156,
-0.03986851125955582,
0.013299019075930119,
0.0412069596350193,
-0.042549360543489456,
0.015729520469903946,
-0.009429426863789558,
0.05334525927901268
] |
SSOLE: Rethinking Orthogonal Low-rank Embedding for Self-Supervised Learning | https://openreview.net/forum?id=zBgiCWCxJB | [
"Lun Huang",
"Qiang Qiu",
"Guillermo Sapiro"
] | Poster | Self-supervised learning (SSL) aims to learn meaningful representations from unlabeled data. Orthogonal Low-rank Embedding (OLE) shows promise for SSL by enhancing intra-class similarity in a low-rank subspace and promoting inter-class dissimilarity in a high-rank subspace, making it particularly suitable for multi-view learning tasks. However, directly applying OLE to SSL poses significant challenges: (1) the virtually infinite number of "classes" in SSL makes achieving the OLE objective impractical, leading to representational collapse; and (2) low-rank constraints may fail to distinguish between positively and negatively correlated features, further undermining learning. To address these issues, we propose SSOLE (Self-Supervised Orthogonal Low-rank Embedding), a novel framework that integrates OLE principles into SSL by (1) decoupling the low-rank and high-rank enforcement to align with SSL objectives; and (2) applying low-rank constraints to feature deviations from their mean, ensuring better alignment of positive pairs by accounting for the signs of cosine similarities. Our theoretical analysis and empirical results demonstrate that these adaptations are crucial to SSOLE’s effectiveness. Moreover, SSOLE achieves competitive performance across SSL benchmarks without relying on large batch sizes, memory banks, or dual-encoder architectures, making it an efficient and scalable solution for self-supervised tasks. Code is available at https://github.com/husthuaan/ssole. | self-supervised learning, orthogonal low-rank embedding | We address key challenges of applying orthogonal low-rank embedding to self-supervised learning. | 1,387 | null | [
-0.055127229541540146,
-0.05895661190152168,
0.006318884436041117,
0.019787926226854324,
0.02324821427464485,
0.08546581119298935,
-0.004374482668936253,
-0.04255767911672592,
0.05433131009340286,
-0.10837271064519882,
-0.007748507894575596,
-0.0005886857397854328,
0.05946283042430878,
0.07371872663497925,
-0.046079326421022415,
-0.05371936783194542,
0.10594020038843155,
0.0672348216176033,
-0.06238671392202377,
-0.014419959858059883,
0.0504973940551281,
-0.013302963227033615,
0.004010867793112993,
0.04383374750614166,
0.06751517206430435,
-0.018948419019579887,
0.048457205295562744,
0.018163083121180534,
0.0005898344097658992,
-0.06205436587333679,
0.08274488896131516,
0.03837084025144577,
-0.044083885848522186,
0.034272320568561554,
-0.07798553258180618,
-0.0612872913479805,
0.024091264232993126,
0.04380091279745102,
-0.1053646057844162,
-0.044539425522089005,
0.03680660203099251,
-0.0010663687717169523,
-0.07794669270515442,
0.045452382415533066,
-0.043629858642816544,
-0.014943910762667656,
-0.013088716194033623,
-0.10842656344175339,
-0.0485406219959259,
-0.03729666769504547,
-0.07983071357011795,
0.025507677346467972,
-0.12523342669010162,
0.03609771281480789,
-0.06576600670814514,
-0.012715515680611134,
0.054010942578315735,
-0.017869453877210617,
-0.0033253394067287445,
0.06875652074813843,
0.04280998930335045,
-0.06636803597211838,
0.04667765647172928,
0.018699390813708305,
-0.0042205071076750755,
0.02029263786971569,
0.022424116730690002,
0.03903714939951897,
-0.0599864162504673,
0.08201838284730911,
-0.05577731132507324,
0.06153743714094162,
-0.031856682151556015,
0.06745584309101105,
0.03401746600866318,
-0.038018204271793365,
-0.005285254679620266,
-0.013695835135877132,
0.11631719768047333,
-0.03915169835090637,
-0.01889253780245781,
0.06428548693656921,
-0.03439062461256981,
0.05849434435367584,
0.1386735886335373,
0.02723761834204197,
0.04641443118453026,
-0.01205992978066206,
-0.011903725564479828,
0.0474582202732563,
0.020083745941519737,
-0.009992585517466068,
-0.048145852982997894,
-0.014032425358891487,
0.018783841282129288,
-0.038531795144081116,
0.044711750000715256,
-0.04794316738843918,
-0.012671680189669132,
0.025182966142892838,
-0.026288289576768875,
0.06820017844438553,
-0.020148271694779396,
-0.03346794843673706,
0.013336671516299248,
-0.011003234423696995,
0.04468251019716263,
-0.007920986041426659,
0.04463253542780876,
-0.15009275078773499,
0.020065397024154663,
-0.028478212654590607,
-0.09663704037666321,
0.030482692644000053,
-0.05159813538193703,
0.04005221277475357,
0.024770431220531464,
-0.010968077927827835,
0.08682580292224884,
-0.039904382079839706,
-0.08238102495670319,
-0.030288342386484146,
0.02923855371773243,
0.10708166658878326,
-0.034927334636449814,
0.0022139300126582384,
-0.0858892872929573,
2.700214297949453e-33,
0.04766996204853058,
0.04913044348359108,
-0.039182957261800766,
-0.08158393204212189,
-0.014403780922293663,
-0.01189351174980402,
-0.053420886397361755,
0.06200258433818817,
-0.03429869934916496,
0.08662013709545135,
0.02012365125119686,
0.02779146283864975,
0.009860060177743435,
0.013540473766624928,
0.045924339443445206,
-0.0008603429305367172,
-0.07306768000125885,
0.050354115664958954,
-0.0006744102574884892,
-0.1318081170320511,
0.03954925388097763,
-0.0038394909352064133,
0.03152093291282654,
-0.060002975165843964,
0.009334486909210682,
0.009256062097847462,
0.02769169956445694,
-0.06443345546722412,
-0.05906284973025322,
0.0278499573469162,
-0.025289345532655716,
-0.04665743187069893,
0.053280677646398544,
0.005171954166144133,
0.07170168310403824,
-0.001377595355734229,
0.036701709032058716,
-0.015347480773925781,
0.0800795927643776,
-0.040055569261312485,
0.01509818620979786,
0.01680711656808853,
0.004919794388115406,
-0.03440369293093681,
-0.044662196189165115,
-0.0555901862680912,
0.04332023859024048,
0.023233791813254356,
-0.04303745925426483,
-0.06489238888025284,
-0.028704004362225533,
0.006723352707922459,
-0.016409246250987053,
-0.059648893773555756,
-0.07946469634771347,
0.0023922836408019066,
0.023406608030200005,
0.06427568942308426,
-0.03451399505138397,
-0.015759725123643875,
-0.10582457482814789,
0.006477778777480125,
-0.037364065647125244,
-0.0010197372175753117,
0.013098707422614098,
0.004506859928369522,
0.08722453564405441,
0.038450539112091064,
0.07426510006189346,
-0.0019687789026647806,
-0.008085581474006176,
0.046321023255586624,
-0.02814861200749874,
-0.09847194701433182,
0.06635361164808273,
0.031522974371910095,
0.07000357657670975,
0.0061955759301781654,
0.06709614396095276,
0.043130628764629364,
0.018903285264968872,
-0.021309297531843185,
-0.005209173075854778,
-0.045353710651397705,
-0.06914840638637543,
-0.03340502083301544,
0.0575709193944931,
-0.05852492153644562,
-0.01318514533340931,
-0.0006642480148002505,
0.020397407934069633,
-0.033839449286460876,
0.009248683229088783,
0.017180031165480614,
0.010344048961997032,
-2.7526543866365814e-33,
-0.04099230468273163,
0.054789453744888306,
-0.0027963314205408096,
-0.04792993515729904,
0.006790078245103359,
-0.0505046471953392,
0.005839387886226177,
0.04790690168738365,
-0.02988053858280182,
0.002327878726646304,
0.03260728344321251,
-0.10351613163948059,
0.03676079586148262,
0.009233894757926464,
0.0375116728246212,
0.03495273366570473,
-0.10428708046674728,
0.03392472118139267,
-0.029889900237321854,
0.03016209788620472,
-0.01746954582631588,
0.10202567279338837,
0.010272403247654438,
0.18413995206356049,
0.07298200577497482,
0.10064058750867844,
-0.016588915139436722,
0.014154903590679169,
-0.0012889824574813247,
-0.01436210423707962,
0.03629336878657341,
0.020222818478941917,
-0.03331606462597847,
-0.07941238582134247,
-0.114325612783432,
0.006285208277404308,
-0.021826865151524544,
-0.054324451833963394,
-0.0030254372395575047,
0.019475258886814117,
-0.0040636369958519936,
0.022183967754244804,
-0.06723135709762573,
-0.031575556844472885,
0.02735893800854683,
-0.10960343480110168,
-0.0421651192009449,
0.04787720367312431,
0.048124488443136215,
0.04730890691280365,
-0.10916736721992493,
0.08042696863412857,
-0.07141556590795517,
0.028429651632905006,
-0.07563389092683792,
-0.07447812706232071,
-0.02325991727411747,
0.02048136107623577,
0.03804030269384384,
0.015937302261590958,
0.05300468951463699,
-0.029282687231898308,
0.06829454749822617,
-0.018287815153598785,
-0.008083021268248558,
-0.09651315212249756,
-0.01273625809699297,
0.04673280566930771,
-0.01707029715180397,
-0.019183609634637833,
-0.035180941224098206,
-0.02036786638200283,
-0.04248819127678871,
-0.08637607842683792,
-0.025322431698441505,
-0.08906353265047073,
0.03062649630010128,
0.03591921553015709,
-0.015543040819466114,
0.001903777476400137,
0.06847193837165833,
0.040329184383153915,
0.004926056135445833,
0.05928608775138855,
0.07699137926101685,
0.040634024888277054,
0.03582233935594559,
0.05233537405729294,
0.049730703234672546,
-0.00942924339324236,
-0.004031123127788305,
0.04675331711769104,
-0.08222787082195282,
0.07562471926212311,
0.024260278791189194,
-3.9677860996789605e-8,
-0.1251087784767151,
0.03415749594569206,
0.04658139869570732,
0.04046081379055977,
0.05355467647314072,
-0.031337328255176544,
-0.03198709711432457,
0.0691676139831543,
-0.01982368528842926,
0.01578909531235695,
-0.00840388797223568,
-0.05040932819247246,
-0.07739477604627609,
-0.047838468104600906,
0.013818948529660702,
0.022091375663876534,
-0.03705999255180359,
-0.023950042203068733,
0.05354396626353264,
-0.026272794231772423,
0.031261272728443146,
-0.03965189680457115,
0.0540192686021328,
-0.02118368074297905,
0.002673482522368431,
-0.03808115795254707,
-0.0004676304233726114,
0.04262606427073479,
0.10922582447528839,
0.02379740960896015,
-0.031126365065574646,
0.042571619153022766,
0.041500672698020935,
-0.004643701482564211,
0.02608659490942955,
0.10979001224040985,
0.02723611705005169,
-0.0024054590612649918,
-0.08476032316684723,
0.013828180730342865,
-0.015330946072936058,
0.13220489025115967,
-0.03605516627430916,
0.03652658313512802,
0.049611322581768036,
0.0369228795170784,
0.06554432213306427,
-0.05225762724876404,
0.06565744429826736,
-0.021582970395684242,
0.004001149442046881,
-0.10447026789188385,
-0.01719903200864792,
0.06021004170179367,
-0.025506695732474327,
0.027271945029497147,
-0.015796592459082603,
-0.005211620591580868,
0.05071676895022392,
0.03972018137574196,
0.044509828090667725,
-0.002453282941132784,
-0.044957347214221954,
0.03171870484948158
] |
3DitScene: Editing Any Scene via Language-guided Disentangled Gaussian Splatting | https://openreview.net/forum?id=iKDbLpVgQc | [
"Qihang Zhang",
"Yinghao Xu",
"Chaoyang Wang",
"Hsin-Ying Lee",
"Gordon Wetzstein",
"Bolei Zhou",
"Ceyuan Yang"
] | Poster | Scene image editing is crucial for entertainment, photography, and advertising design. Existing methods solely focus on either 2D individual object or 3D global scene editing. This results in a lack of a unified approach to effectively control and manipulate scenes at the 3D level with different levels of granularity. In this work, we propose 3DitScene, a novel and unified scene editing framework leveraging language-guided disentangled Gaussian Splatting that enables seamless editing from 2D to 3D, allowing precise control over scene composition and individual objects. We first incorporate 3D Gaussians that are refined through generative priors and optimization techniques. Language features from CLIP then introduce semantics into 3D geometry for object disentanglement. With the disentangled Gaussians, 3DitScene allows for manipulation at both the global and individual levels, revolutionizing creative expression and empowering control over scenes and objects. Experimental results demonstrate the effectiveness and versatility of 3DitScene in scene image editing. | image editting, gaussian splatting, 3D | null | 1,386 | 2405.18424 | [
-0.04956397786736488,
-0.056096047163009644,
0.07125333696603775,
0.007309907115995884,
0.04280233755707741,
0.000670565408654511,
-0.02960377186536789,
-0.1034267246723175,
-0.016240054741501808,
-0.06771866977214813,
-0.03475451096892357,
-0.03527594730257988,
-0.025531290099024773,
0.0749155655503273,
0.01612228713929653,
0.027461307123303413,
0.08143552392721176,
0.08413117378950119,
-0.11322230845689774,
0.03548333793878555,
0.027113176882267,
-0.04869788885116577,
0.024322858080267906,
-0.05140846595168114,
0.09972153604030609,
-0.011303682811558247,
0.049264565110206604,
-0.06252417713403702,
0.04443679377436638,
0.03138167783617973,
0.05854714289307594,
0.0025990325957536697,
0.008773939684033394,
0.0432964488863945,
-0.08688659965991974,
0.07309486716985703,
-0.014781766571104527,
0.03203633427619934,
-0.011291849426925182,
-0.037154458463191986,
-0.0007271278300322592,
0.004316408187150955,
-0.03569383546710014,
0.029897741973400116,
0.034272387623786926,
0.05607662722468376,
0.0017977799288928509,
-0.021817710250616074,
0.03943559527397156,
0.0046552191488444805,
-0.02437838539481163,
-0.05044388771057129,
-0.016442695632576942,
-0.023400839418172836,
0.027607623487710953,
0.052281420677900314,
0.07930050045251846,
-0.04465124011039734,
0.026378128677606583,
-0.02617262862622738,
0.009731554426252842,
0.0022954815067350864,
0.03484926000237465,
-0.005422617774456739,
0.004710569512099028,
-0.07367269694805145,
0.02788815274834633,
-0.02922309748828411,
-0.0676373764872551,
-0.011922416277229786,
-0.004923103377223015,
-0.014292198233306408,
0.012033660896122456,
-0.011886077001690865,
0.032039888203144073,
-0.018958307802677155,
-0.07295706123113632,
-0.10293933749198914,
-0.044883180409669876,
-0.048876095563173294,
0.08732855319976807,
-0.03289823234081268,
0.02154395915567875,
0.059667665511369705,
0.00720931775867939,
0.016269147396087646,
0.009551326744258404,
0.06324471533298492,
-0.0271792933344841,
0.03896799311041832,
-0.06527453660964966,
-0.004132481291890144,
-0.009757188148796558,
-0.08140092343091965,
-0.02118513733148575,
-0.07030675560235977,
-0.025079939514398575,
-0.09564264118671417,
0.0052048806101083755,
0.00233456096611917,
0.02173634245991707,
-0.06204964965581894,
0.01686677895486355,
-0.04609028249979019,
0.07484178990125656,
-0.05790867656469345,
0.05220482125878334,
-0.028064021840691566,
-0.01809714175760746,
0.02509954571723938,
0.030255800113081932,
0.03857855871319771,
0.020851701498031616,
-0.034908074885606766,
-0.023634212091565132,
0.051575981080532074,
-0.00353538291528821,
0.003721412504091859,
0.01380088459700346,
0.020118707790970802,
0.05504381284117699,
0.019890544936060905,
-0.06651372462511063,
0.07472631335258484,
-0.03405928239226341,
-0.03700093552470207,
0.010141352191567421,
2.0712568619923722e-33,
0.09680148959159851,
-0.009391660802066326,
0.0056608798913657665,
0.10493725538253784,
0.0205620639026165,
0.06748691201210022,
-0.045401714742183685,
0.00849382858723402,
-0.10453180968761444,
-0.026213662698864937,
0.030570143833756447,
-0.1013612225651741,
-0.019897514954209328,
0.11970676481723785,
0.04353765770792961,
0.012356739491224289,
-0.032473258674144745,
0.026761222630739212,
0.037833862006664276,
0.09207561612129211,
0.0019663572311401367,
-0.028906187042593956,
-0.08242277055978775,
0.03977150842547417,
-0.06391597539186478,
-0.04771406203508377,
0.05967539921402931,
-0.027925118803977966,
0.054842859506607056,
-0.04377752169966698,
-0.09287001192569733,
0.012436854653060436,
0.023818418383598328,
-0.014989655464887619,
-0.04505794122815132,
-0.02474338747560978,
-0.0188628900796175,
-0.09172806143760681,
0.020225128158926964,
0.012186641804873943,
-0.062247540801763535,
0.03788537532091141,
-0.004285816103219986,
-0.02420824021100998,
-0.09663467109203339,
0.031264178454875946,
0.11429986357688904,
0.11043232679367065,
-0.04732980206608772,
0.06133800372481346,
0.09651482850313187,
0.0010032725986093283,
0.006417581811547279,
-0.013671890832483768,
-0.03647636994719505,
-0.05619601905345917,
0.020956596359610558,
-0.05443768575787544,
0.024219723418354988,
-0.00743733299896121,
0.03542407974600792,
0.07457391172647476,
-0.041195765137672424,
0.0645817369222641,
0.013579586520791054,
0.07550882548093796,
0.05293715372681618,
0.05505407229065895,
0.07053016871213913,
0.038940124213695526,
-0.1199839636683464,
-0.043079860508441925,
0.017681391909718513,
0.03058638423681259,
0.006478140130639076,
-0.01720595359802246,
0.0014805105747655034,
-0.09396551549434662,
0.08284856379032135,
0.03181770443916321,
-0.10211433470249176,
-0.021870842203497887,
0.0003154992882627994,
0.025710467249155045,
-0.09471755474805832,
-0.10793928802013397,
-0.0057052974589169025,
-0.012390805408358574,
-0.023012561723589897,
-0.044744379818439484,
-0.052030738443136215,
-0.07205603271722794,
-0.0023228907957673073,
-0.005716319661587477,
-0.006426147185266018,
-9.166545355330508e-34,
0.035589225590229034,
-0.0026696368586272,
-0.054178155958652496,
0.027840441092848778,
0.008695888333022594,
-0.04779910668730736,
-0.030229903757572174,
-0.0240092184394598,
0.052364788949489594,
-0.01587466336786747,
0.00330783030949533,
0.007345264777541161,
0.0048674545250833035,
0.003306731814518571,
-0.046321604400873184,
0.0009131300030276179,
-0.004330180585384369,
0.001120199216529727,
-0.07826385647058487,
-0.025371452793478966,
0.06845598667860031,
-0.04336092993617058,
-0.030980374664068222,
-0.022547027096152306,
-0.11864638328552246,
0.10194220393896103,
-0.05302796512842178,
0.0358668714761734,
0.003869029227644205,
0.02029508724808693,
-0.06895092129707336,
0.021720528602600098,
-0.002848114352673292,
-0.00521228089928627,
-0.07146994024515152,
-0.00784989818930626,
0.07029424607753754,
0.02188355103135109,
-0.05127877742052078,
-0.0564638152718544,
-0.01495532225817442,
-0.026943981647491455,
-0.0023723714984953403,
0.12960651516914368,
-0.015256194397807121,
0.05351245775818825,
-0.00836213305592537,
-0.007843602448701859,
-0.04707787185907364,
0.0986463725566864,
-0.15754978358745575,
0.08400126546621323,
-0.07190997153520584,
-0.06795383989810944,
-0.014976177364587784,
-0.09400810301303864,
-0.08572515845298767,
-0.040901269763708115,
0.13498298823833466,
-0.016185475513339043,
0.007771293167024851,
0.01686640828847885,
-0.04606609046459198,
-0.008953945711255074,
0.014167075045406818,
0.059248246252536774,
-0.02600562758743763,
0.014640686102211475,
0.0059328461065888405,
0.05989697575569153,
0.02417454682290554,
0.005776845384389162,
0.07340884208679199,
-0.023391708731651306,
-0.01617472618818283,
-0.016808250918984413,
0.03372844308614731,
-0.0076437899842858315,
0.058236800134181976,
-0.00909489206969738,
0.11136423796415329,
-0.04137909412384033,
0.04787043482065201,
0.10091084986925125,
0.02755003608763218,
0.08976203948259354,
-0.03513067215681076,
0.05482747405767441,
0.02584812231361866,
0.010174117051064968,
-0.06713274121284485,
0.0673443004488945,
0.0666075199842453,
0.03220951184630394,
-0.017591074109077454,
-3.345029497836549e-8,
-0.034082651138305664,
0.017423342913389206,
-0.011722555384039879,
-0.0518314391374588,
-0.06848384439945221,
-0.04660208523273468,
0.016848981380462646,
0.07796411961317062,
0.030560562387108803,
-0.08008966594934464,
0.0031848005019128323,
-0.019007423892617226,
-0.06548282504081726,
-0.05456904321908951,
0.009115325286984444,
0.047668974846601486,
-0.018237940967082977,
0.02855827286839485,
0.015986856073141098,
0.023643100634217262,
-0.027704283595085144,
0.05888376757502556,
-0.009463787078857422,
-0.02782071754336357,
-0.03442896157503128,
-0.0145138930529356,
0.003631321946159005,
0.10573915392160416,
0.11796244978904724,
-0.012418217025697231,
0.03082515485584736,
-0.005949679296463728,
0.028306467458605766,
0.11081020534038544,
-0.08317087590694427,
-0.005367065314203501,
-0.04662584885954857,
0.02665683999657631,
0.00450857263058424,
-0.06188223883509636,
0.0628737062215805,
0.07197044044733047,
0.008165265433490276,
0.037590306252241135,
0.028684193268418312,
0.02406552992761135,
0.10871072858572006,
-0.10705597698688507,
0.09276572614908218,
0.04643459618091583,
-0.02388293296098709,
0.016224203631281853,
0.0388590544462204,
0.08928655087947845,
0.03205135837197304,
0.0472310334444046,
-0.008470960892736912,
0.003475199919193983,
0.034816861152648926,
0.05407242104411125,
-0.08452768623828888,
0.023804958909749985,
-0.01899804174900055,
0.008693465031683445
] |
CoMotion: Concurrent Multi-person 3D Motion | https://openreview.net/forum?id=qKu6KWPgxt | [
"Alejandro Newell",
"Peiyun Hu",
"Lahav Lipson",
"Stephan Richter",
"Vladlen Koltun"
] | Poster | We introduce an approach for detecting and tracking detailed 3D poses of multiple people from a single monocular camera stream. Our system maintains temporally coherent predictions in crowded scenes filled with difficult poses and occlusions. Our model performs both strong per-frame detection and a learned pose update to track people from frame to frame. Rather than match detections across time, poses are updated directly from a new input image, which enables online tracking through occlusion. We train on numerous image and video datasets leveraging pseudo-labeled annotations to produce a model that matches state-of-the-art systems in 3D pose estimation accuracy while being faster and more accurate in tracking multiple people through time. | human pose estimation, 3d human pose, tracking | Online multi-person 3D pose estimation and tracking in video with recurrent cross-attention | 1,384 | null | [
-0.061882320791482925,
-0.06827188283205032,
-0.010242964141070843,
-0.056045930832624435,
0.14417803287506104,
0.0110926553606987,
0.05557846277952194,
-0.05196867883205414,
0.027468718588352203,
-0.06635907292366028,
-0.0016907879617065191,
-0.08061620593070984,
-0.07086247205734253,
0.10478819906711578,
-0.025927793234586716,
-0.04282660409808159,
0.06877046078443527,
0.10307764261960983,
-0.02553401328623295,
0.026043865829706192,
-0.05222241207957268,
-0.041849538683891296,
0.0882168561220169,
0.07967491447925568,
0.05336183309555054,
0.014155327342450619,
0.04447953402996063,
-0.0984528586268425,
-0.019425852224230766,
-0.0016374323749914765,
0.07634619623422623,
-0.009680177085101604,
0.0037216513883322477,
0.011518117971718311,
-0.0563817098736763,
0.02635396085679531,
-0.00546769006177783,
0.0012877454282715917,
-0.07164932787418365,
0.006816437467932701,
0.018704412505030632,
0.05887588858604431,
0.046842217445373535,
0.02708408050239086,
0.06046522036194801,
0.11212605983018875,
0.04572794958949089,
-0.07286044955253601,
0.07234571129083633,
0.006960708182305098,
-0.03382166102528572,
-0.03892005607485771,
0.06860862672328949,
0.0002523372822906822,
-0.03777554631233215,
-0.012837404385209084,
0.027749594300985336,
-0.09731698036193848,
0.025055890902876854,
-0.043691486120224,
0.07091483473777771,
-0.049603886902332306,
-0.01982632651925087,
0.04842881113290787,
0.015023373067378998,
-0.024046344682574272,
0.0055954488925635815,
0.0038258659187704325,
0.009476390667259693,
0.04605501517653465,
-0.040997982025146484,
0.036877620965242386,
0.03357155993580818,
-0.016364023089408875,
0.02080412209033966,
-0.022781897336244583,
0.028316786512732506,
-0.06654994934797287,
0.04416515678167343,
-0.06548595428466797,
0.02420365996658802,
-0.07723959535360336,
0.00018636553431861103,
-0.024912891909480095,
0.06091257929801941,
-0.009430172853171825,
-0.03832794725894928,
0.07586110383272171,
-0.049394309520721436,
-0.01564175821840763,
-0.13354715704917908,
0.007171583361923695,
0.060532212257385254,
-0.06566745042800903,
0.0012243549572303891,
0.029590943828225136,
-0.01338510774075985,
0.03552262857556343,
0.018624918535351753,
0.05479223281145096,
-0.046448007225990295,
0.06271450221538544,
0.029360894113779068,
0.025654662400484085,
0.03791921213269234,
0.04066760838031769,
0.026167934760451317,
0.027736829593777657,
0.046155184507369995,
-0.04532908275723457,
0.05210915207862854,
-0.0648554340004921,
0.019122455269098282,
0.03104887343943119,
0.06418829411268234,
0.0665392279624939,
-0.013839609920978546,
0.026023615151643753,
0.014285245910286903,
0.0040623401291668415,
0.007171528413891792,
-0.06654525548219681,
0.02083767019212246,
-0.012841946445405483,
0.00289361784234643,
0.027938751503825188,
0.01310963835567236,
1.6476031878262503e-34,
0.052566301077604294,
0.03830122947692871,
0.06539527326822281,
0.02613873966038227,
-0.04151342064142227,
0.01907418854534626,
-0.06013256311416626,
0.0011451110476627946,
-0.008227577432990074,
-0.013612824492156506,
0.041413046419620514,
-0.13235768675804138,
-0.031808819621801376,
0.10518039762973785,
-0.038038574159145355,
-0.08515258878469467,
-0.03784888610243797,
0.020458241924643517,
-0.041012242436409,
0.13056066632270813,
0.0595153383910656,
-0.018105557188391685,
-0.08607213944196701,
-0.00630188686773181,
0.008108611218631268,
-0.04047162085771561,
0.05620169639587402,
0.0658036321401596,
0.048179224133491516,
-0.022703852504491806,
0.03730805963277817,
-0.0005833631148561835,
0.007588906679302454,
0.042610008269548416,
0.11463876068592072,
-0.06938966363668442,
-0.023870788514614105,
-0.01999773643910885,
0.04373272508382797,
0.021641436964273453,
0.02994685061275959,
-0.0016363972099497914,
-0.047089554369449615,
-0.08694636076688766,
-0.06552543491125107,
-0.0014063817216083407,
0.003834635019302368,
0.12382204830646515,
-0.12865352630615234,
0.06998922675848007,
0.06034654751420021,
0.037770725786685944,
-0.10485787689685822,
-0.005876506678760052,
-0.11068745702505112,
0.01470969244837761,
-0.01749405264854431,
-0.0005212692776694894,
0.01212281733751297,
-0.05417759716510773,
-0.016135068610310555,
0.028165781870484352,
-0.03353423625230789,
-0.033398594707250595,
-0.035786766558885574,
0.03960627317428589,
0.08128184825181961,
0.08597246557474136,
0.1421467512845993,
0.06656132638454437,
-0.054998695850372314,
0.014483347535133362,
-0.013922099955379963,
-0.029100172221660614,
0.12429971247911453,
0.02270786464214325,
0.04572093486785889,
0.016026772558689117,
-0.0424957312643528,
0.09464533627033234,
-0.036231450736522675,
-0.06504378467798233,
0.02505568601191044,
0.042727865278720856,
-0.015267539769411087,
-0.027614479884505272,
-0.019722985103726387,
-0.015530363656580448,
-0.03866330161690712,
0.06069903448224068,
0.03448541462421417,
-0.05829615890979767,
-0.041579607874155045,
0.012657316401600838,
-0.021302996203303337,
9.448815528096517e-34,
0.09103485196828842,
0.041513022035360336,
-0.00588120287284255,
-0.048856109380722046,
0.07907593995332718,
-0.0920552909374237,
-0.0023649432696402073,
0.043767720460891724,
0.059913598001003265,
-0.06437323242425919,
-0.03752126544713974,
-0.052052050828933716,
0.054072678089141846,
-0.022599704563617706,
0.02546175941824913,
-0.0016747807385399938,
0.015198932029306889,
0.012475891038775444,
-0.04005769267678261,
0.012365814298391342,
0.017186755314469337,
-0.0284369308501482,
0.00455216271802783,
-0.0022973085287958384,
-0.08057387173175812,
0.10307314991950989,
0.01944759115576744,
-0.011676871217787266,
-0.04036377742886543,
-0.009057669900357723,
-0.005108661018311977,
-0.0549897775053978,
-0.01632099226117134,
-0.07432881742715836,
-0.07858818769454956,
0.038497183471918106,
-0.0000471787279821001,
0.017886048182845116,
-0.025811105966567993,
0.024761566892266273,
0.03607616201043129,
-0.04936426132917404,
-0.0835457295179367,
0.03680219501256943,
-0.04331549257040024,
0.01166264247149229,
-0.05541778355836868,
0.010072273202240467,
-0.021812893450260162,
0.06895308196544647,
-0.14109045267105103,
0.034042444080114365,
-0.05612147971987724,
-0.03060450404882431,
-0.004075217992067337,
-0.08757288008928299,
-0.02172238938510418,
-0.0719851478934288,
0.06668074429035187,
-0.04295450821518898,
-0.09381255507469177,
0.02015327848494053,
-0.028779489919543266,
0.08511866629123688,
0.0449250228703022,
-0.0012546784710139036,
-0.02403760887682438,
-0.03989062458276749,
-0.023432139307260513,
0.026391956955194473,
0.04609210789203644,
-0.04111146554350853,
-0.0327192060649395,
0.04099753871560097,
-0.03228776529431343,
-0.0666659027338028,
0.0013850046088919044,
0.007716528605669737,
0.09948765486478806,
0.0014081939589232206,
0.014312769286334515,
-0.07139766961336136,
0.021596558392047882,
0.008287711068987846,
0.01997474581003189,
0.07774040102958679,
0.03977791219949722,
-0.007276777643710375,
-0.0743831992149353,
0.049870893359184265,
-0.05163674056529999,
0.07165773957967758,
-0.0021609696559607983,
-0.008913233876228333,
-0.035013675689697266,
-2.9095954090507803e-8,
-0.1005011796951294,
0.06477626413106918,
0.005851786583662033,
-0.000930515700019896,
-0.05212489143013954,
0.0005771382711827755,
0.03810899704694748,
0.0661911889910698,
0.006994970142841339,
-0.009486488066613674,
0.033392585813999176,
0.027390245348215103,
0.051566168665885925,
-0.013474471867084503,
-0.01787429302930832,
0.013646243140101433,
-0.06610366702079773,
-0.047872528433799744,
-0.06323724240064621,
0.06770766526460648,
0.005304936319589615,
0.0049013313837349415,
0.034976691007614136,
0.023995008319616318,
-0.002711814595386386,
0.04927822947502136,
0.04126908257603645,
0.0776108130812645,
0.0580146349966526,
0.003061937401071191,
-0.0610436275601387,
0.029324183240532875,
0.015553826466202736,
-0.008652236312627792,
0.03484554588794708,
-0.02997114695608616,
-0.033023279160261154,
-0.031584132462739944,
0.050524286925792694,
-0.030852792784571648,
-0.033007875084877014,
0.079366035759449,
0.02130937948822975,
0.052755530923604965,
0.06873147934675217,
0.0619572289288044,
-0.01887454465031624,
-0.05254702270030975,
-0.028916209936141968,
-0.08874890208244324,
-0.03652004525065422,
-0.04714517667889595,
0.014552478678524494,
0.15941523015499115,
0.03159085288643837,
-0.010264670476317406,
-0.013678851537406445,
-0.039350077509880066,
0.07805431634187698,
0.017843397334218025,
-0.06408853828907013,
-0.06413717567920685,
-0.01741575077176094,
-0.01301045436412096
] |
Scaling Optimal LR Across Token Horizons | https://openreview.net/forum?id=WYL4eFLcxG | [
"Johan Bjorck",
"Alon Benhaim",
"Vishrav Chaudhary",
"Furu Wei",
"Xia Song"
] | Poster | State-of-the-art LLMs are powered by scaling -- scaling model size, training tokens, and cluster size. It is economically infeasible to extensively tune hyperparameters for the largest runs. Instead, approximately optimal hyperparameters must be inferred or transferred from smaller experiments. Hyperparameter transfer across model sizes has been studied in muP. However, hyperparameter transfer across training tokens -- or token horizon -- has not been studied yet. To remedy this we conduct a large-scale empirical study on how optimal learning rate (LR) depends on the token horizon in LLM training. We first demonstrate that the optimal LR changes significantly with token horizon -- longer training necessitates smaller LR. Secondly, we demonstrate that the optimal LR follows a scaling law and that the optimal LR for longer horizons can be accurately estimated from shorter horizons via such scaling laws. We also provide a rule-of-thumb for transferring LR across token horizons with zero overhead over current practices. Lastly, we provide evidence that LLama-1 used too high LR, and thus argue that hyperparameter transfer across data size is an overlooked component of LLM training. | LLMs, scaling laws, hyperparameters, mup | We show how the optimal LR for LLM pretraining depends on token horizon | 1,380 | 2409.19913 | [
-0.01805853471159935,
-0.07847648113965988,
0.053498923778533936,
0.02864406630396843,
0.10730259120464325,
-0.02667086012661457,
-0.09935979545116425,
0.012308205477893353,
0.0030370140448212624,
-0.03312128409743309,
-0.03668304905295372,
-0.018061766400933266,
-0.04097886383533478,
-0.0026605073362588882,
-0.017769847065210342,
-0.00195727008394897,
0.07838429510593414,
0.07934199273586273,
-0.11087605357170105,
-0.02102162502706051,
-0.010718005709350109,
-0.04637906700372696,
0.013342494145035744,
0.06186879426240921,
0.08302614092826843,
-0.03030170314013958,
-0.027937809005379677,
-0.04379144310951233,
0.003049014136195183,
-0.023303190246224403,
0.016246594488620758,
0.024910818785429,
-0.026718314737081528,
0.008113027550280094,
-0.05521923303604126,
0.047979094088077545,
-0.013133731670677662,
-0.012630051001906395,
0.048556990921497345,
-0.005839743185788393,
0.05058494582772255,
-0.05068136751651764,
0.01867111399769783,
0.015332715585827827,
0.015892712399363518,
0.05171254277229309,
0.019058851525187492,
0.003147420007735491,
-0.07745645195245743,
0.03396676108241081,
-0.02490808442234993,
-0.03934810683131218,
-0.07180638611316681,
0.05115630850195885,
-0.06125421077013016,
0.03691079840064049,
0.031296227127313614,
0.010901132598519325,
0.009106731042265892,
-0.002146719256415963,
-0.12352707982063293,
-0.08245091885328293,
-0.01431808527559042,
-0.013039265759289265,
-0.05475887656211853,
-0.06552194803953171,
0.03891514614224434,
0.028397921472787857,
-0.015527530573308468,
0.03323371335864067,
-0.003715932136401534,
0.0700535699725151,
-0.07800270617008209,
0.004033091943711042,
0.028747636824846268,
0.01757824420928955,
0.04535152018070221,
0.05402965843677521,
0.041925206780433655,
-0.05720937252044678,
0.053454622626304626,
0.06234017014503479,
0.012359357438981533,
-0.03992823511362076,
0.07053478807210922,
-0.07177823036909103,
0.05479789525270462,
-0.012713870033621788,
0.11100040376186371,
-0.04875246807932854,
0.0370170921087265,
-0.0384710468351841,
-0.05925767496228218,
0.007537703029811382,
0.0009192804573103786,
-0.01376072783023119,
-0.041172582656145096,
-0.047153037041425705,
0.010933621786534786,
0.0810558870434761,
0.0041767000220716,
0.055782709270715714,
0.058032311499118805,
0.029548279941082,
0.024871064350008965,
-0.01968584582209587,
0.025054240599274635,
0.03409565985202789,
0.09481846541166306,
-0.043117593973875046,
0.1035456508398056,
0.11864415556192398,
-0.0057079545222222805,
-0.005022313445806503,
0.014829390682280064,
0.0070993960835039616,
-0.026675065979361534,
-0.038681548088788986,
-0.0489630401134491,
0.07194799184799194,
-0.06631649285554886,
-0.0728919580578804,
0.01803811825811863,
0.03672206029295921,
-0.029398376122117043,
0.011623849160969257,
-0.0399855338037014,
4.0284228136878256e-33,
0.027712570503354073,
0.012930925004184246,
0.003825942985713482,
-0.08303037285804749,
-0.02749917283654213,
0.0028622422832995653,
0.06768780201673508,
0.05256759747862816,
-0.025297287851572037,
-0.05344276502728462,
-0.091010682284832,
0.009368028491735458,
-0.0020496544893831015,
0.04778819531202316,
0.06471345573663712,
-0.020433785393834114,
-0.0133786266669631,
0.053629837930202484,
0.00860036164522171,
-0.05719020217657089,
0.0688309445977211,
-0.07431475818157196,
-0.05098392441868782,
-0.11905039846897125,
-0.019276171922683716,
0.01912480965256691,
0.036141809076070786,
-0.025695577263832092,
-0.06180304288864136,
0.04705744609236717,
-0.06246126815676689,
-0.025340475142002106,
-0.012074767611920834,
0.08254271745681763,
0.006028052419424057,
0.06155005469918251,
-0.02381628192961216,
-0.03121861442923546,
0.02726949006319046,
-0.10148153454065323,
0.0063949148170650005,
0.05565473437309265,
0.023353179916739464,
-0.06703512370586395,
-0.08102978765964508,
-0.09072158485651016,
0.04126402363181114,
-0.039213620126247406,
-0.082712821662426,
-0.04623067378997803,
0.04434460401535034,
-0.06311145424842834,
-0.052894048392772675,
-0.07583737373352051,
-0.006229300517588854,
-0.0003910089435521513,
0.007361974101513624,
-0.0022529319394379854,
-0.033064767718315125,
0.09234581887722015,
-0.026667749509215355,
-0.019413556903600693,
-0.007992807775735855,
0.11585386842489243,
0.010945097543299198,
0.04644948989152908,
-0.013737455010414124,
0.016496362164616585,
0.010758128017187119,
-0.03730906546115875,
-0.0005565910832956433,
0.0071137831546366215,
0.005310814827680588,
-0.0628838762640953,
0.0714075043797493,
-0.057832270860672,
0.1248229369521141,
-0.004842130932956934,
-0.020254535600543022,
0.021635930985212326,
0.007857237011194229,
0.03221931308507919,
-0.0029114328790456057,
-0.04556133598089218,
0.009011595509946346,
-0.0607297308743,
-0.0005201718886382878,
-0.05992358550429344,
-0.03921191021800041,
-0.01526705827564001,
-0.046168189495801926,
-0.06286413967609406,
0.0010623286943882704,
0.013024529442191124,
-0.04194632172584534,
-3.566623613153551e-33,
-0.013133393600583076,
-0.03208261355757713,
0.05398543179035187,
0.1355949342250824,
-0.041233815252780914,
-0.014882766641676426,
0.003894743276759982,
0.08930043131113052,
-0.13804470002651215,
-0.04786629602313042,
-0.07768063247203827,
-0.014134353958070278,
-0.004144467879086733,
-0.06449209898710251,
0.024072512984275818,
-0.01540425792336464,
-0.006644461769610643,
0.04596364125609398,
0.05696212127804756,
-0.015981871634721756,
0.06523261964321136,
0.01894400268793106,
-0.05185522139072418,
0.07343783229589462,
-0.018800346180796623,
-0.010862087830901146,
-0.014425649307668209,
0.07792496681213379,
0.005137656815350056,
-0.02730831876397133,
-0.07925155758857727,
-0.022216785699129105,
-0.014863074757158756,
-0.008385241962969303,
-0.041561421006917953,
0.06500444561243057,
0.023917365819215775,
0.055828724056482315,
-0.03114837221801281,
0.1484043002128601,
0.0395183227956295,
-0.007313514593988657,
-0.019980616867542267,
0.024093355983495712,
0.032289668917655945,
-0.021981334313750267,
-0.020282354205846786,
-0.0014736693119630218,
0.04151582345366478,
0.01155149471014738,
-0.011898978613317013,
-0.02242770791053772,
-0.012554382905364037,
0.03586215525865555,
-0.07912912219762802,
-0.009701060131192207,
-0.0398511104285717,
0.013420387171208858,
-0.015591012313961983,
-0.038947705179452896,
0.003568700747564435,
0.01092435047030449,
-0.04898916929960251,
-0.008041033521294594,
-0.03183170408010483,
-0.038900312036275864,
0.021356210112571716,
-0.0021748091094195843,
-0.0228451918810606,
0.041978009045124054,
-0.018572861328721046,
-0.005127749405801296,
0.004447939805686474,
0.02750239707529545,
-0.06643868237733841,
-0.029532290995121002,
0.05821539834141731,
-0.09927733987569809,
-0.015470982529222965,
-0.009619127959012985,
-0.005257349461317062,
0.01617533154785633,
0.04237798973917961,
0.05534663796424866,
0.05963033437728882,
0.09794947504997253,
0.12367618083953857,
0.05828237161040306,
0.013848450966179371,
-0.009470745921134949,
-0.06753622740507126,
0.0021042879670858383,
0.03447428345680237,
0.005987484939396381,
-0.042886100709438324,
-4.418169652353754e-8,
-0.03986700624227524,
0.044690296053886414,
-0.006089739967137575,
0.06855002790689468,
0.03376859426498413,
0.10058559477329254,
-0.032912712544202805,
0.11605663597583771,
-0.047939423471689224,
0.03317803144454956,
0.1053292453289032,
0.01584361493587494,
-0.08379997313022614,
-0.0009671521838754416,
0.040488291531801224,
0.046241194009780884,
0.037174042314291,
-0.029985014349222183,
-0.039011094719171524,
0.023931842297315598,
0.0019855275750160217,
0.09324196726083755,
0.07784657180309296,
-0.05540091544389725,
0.10957348346710205,
-0.09327437728643417,
-0.023808516561985016,
0.1924244910478592,
0.029775656759738922,
-0.06699493527412415,
-0.060203660279512405,
0.040521394461393356,
-0.012189599685370922,
0.007505853194743395,
0.023519663140177727,
0.03904763236641884,
-0.04656769707798958,
0.013802769593894482,
0.03749324008822441,
0.019761962816119194,
0.00021615545847453177,
0.08888368308544159,
-0.057565439492464066,
0.020953133702278137,
0.007840178906917572,
0.021010231226682663,
-0.027042439207434654,
-0.0592578761279583,
-0.009413052350282669,
-0.021844105795025826,
0.10016666352748871,
-0.03557077795267105,
-0.04709034040570259,
0.07320203632116318,
0.06654424965381622,
0.012133806943893433,
-0.09271803498268127,
-0.04559994488954544,
0.009449693374335766,
0.017837369814515114,
0.06768936663866043,
-0.05074090510606766,
-0.17025217413902283,
0.0077911242842674255
] |
Order-aware Interactive Segmentation | https://openreview.net/forum?id=8ZLzw5pIrc | [
"Bin Wang",
"Anwesa Choudhuri",
"Meng Zheng",
"Zhongpai Gao",
"Benjamin Planche",
"Andong Deng",
"Qin Liu",
"Terrence Chen",
"Ulas Bagci",
"Ziyan Wu"
] | Poster | Interactive segmentation aims to accurately segment target objects with minimal user interactions. However, current methods often fail to accurately separate target objects from the background, due to a limited understanding of order, the relative depth between objects in a scene. To address this issue, we propose OIS: order-aware interactive segmentation, where we explicitly encode the relative depth between objects into order maps. We introduce a novel order-aware attention, where the order maps seamlessly guide the user interactions (in the form of clicks) to attend to the image features. We further present an object-aware attention module to incorporate a strong object-level understanding to better differentiate objects with similar order. Our approach allows both dense and sparse integration of user clicks, enhancing both accuracy and efficiency as compared to prior works. Experimental results demonstrate that OIS achieves state-of-the-art performance, improving mIoU after one click by 7.61 on the HQSeg44K dataset and 1.32 on the DAVIS dataset as compared to the previous state-of-the-art SegNext, while also doubling inference speed compared to current leading methods. | Interactive Segmentation, Image Segmentation | We propose OIS: order-aware interactive segmentation, to explicitly integrate relative depth infomation into 2D interactive segmentation, which improves the model's ability to distinguish objects based on their relative depths from each other. | 1,374 | 2410.12214 | [
0.009803974069654942,
-0.06369228661060333,
0.03200454264879227,
0.02073673903942108,
0.08352867513895035,
-0.0645361840724945,
-0.015401325188577175,
0.0004984766128472984,
0.0015631701098755002,
-0.08191215246915817,
-0.07634332776069641,
0.00913999229669571,
0.009727181866765022,
0.03787244111299515,
0.023586947470903397,
0.010119599290192127,
0.10783357173204422,
0.05967330560088158,
-0.11468896269798279,
0.004495456349104643,
0.0024735438637435436,
-0.06785270571708679,
0.06363833695650101,
-0.04417983070015907,
-0.0031233669724315405,
-0.041594862937927246,
-0.014483542181551456,
-0.055426713079214096,
0.022210344672203064,
-0.030687430873513222,
0.039521362632513046,
0.03178722411394119,
0.09424895793199539,
-0.0015549003146588802,
-0.02016586996614933,
0.013855365104973316,
0.0158200953155756,
-0.07101225107908249,
-0.02519986219704151,
-0.04643094539642334,
-0.02599569782614708,
0.02645687945187092,
0.03368021175265312,
0.0632171556353569,
0.007948464713990688,
-0.009898671880364418,
0.0442412830889225,
-0.04830637946724892,
0.08186047524213791,
0.03195673227310181,
-0.0495990552008152,
-0.08102792501449585,
-0.008020157925784588,
0.0646715834736824,
-0.09287010133266449,
0.05300736427307129,
0.036063238978385925,
-0.007248619571328163,
0.043087951838970184,
0.03697894886136055,
-0.0671481192111969,
-0.09808966517448425,
0.06853765994310379,
-0.007865529507398605,
-0.012347616255283356,
-0.1185721829533577,
0.05832894518971443,
-0.06772255152463913,
-0.04950932785868645,
-0.027204938232898712,
0.022261224687099457,
0.017496855929493904,
0.027421653270721436,
0.040652476251125336,
0.03782599791884422,
0.0382002592086792,
-0.0431758314371109,
-0.08619651198387146,
0.004804325755685568,
-0.1256023794412613,
-0.019135942682623863,
-0.08574653416872025,
-0.004897396545857191,
-0.007857679389417171,
0.05829630419611931,
-0.018350502476096153,
-0.0012042156886309385,
0.026970427483320236,
-0.07853744924068451,
0.050508175045251846,
-0.052586331963539124,
-0.07964000105857849,
0.014930465258657932,
-0.037223931401968,
0.09956800937652588,
-0.020550647750496864,
0.037415895611047745,
-0.04880286753177643,
-0.07261605560779572,
0.07132717967033386,
-0.02394135855138302,
0.03792585805058479,
-0.012765806168317795,
-0.0508025623857975,
0.054647598415613174,
0.002223996678367257,
0.002991247456520796,
0.01930565945804119,
0.014372703619301319,
0.010006307624280453,
0.014909184537827969,
-0.024907119572162628,
0.05494129657745361,
-0.014386279508471489,
-0.023066284134984016,
-0.02406657300889492,
-0.015069789253175259,
-0.005797755438834429,
0.07801727950572968,
-0.024086959660053253,
0.09386853873729706,
0.006090737879276276,
-0.06645652651786804,
0.032999731600284576,
0.057004015892744064,
-0.0018429787596687675,
-0.015520822256803513,
-1.2586799534326537e-33,
0.08195798844099045,
0.004151154775172472,
0.09775600582361221,
-0.04205268621444702,
0.01499119121581316,
-0.020316224545240402,
-0.007840540260076523,
-0.010428916662931442,
-0.08573591709136963,
-0.002741622505709529,
-0.032436080276966095,
-0.030469095334410667,
0.026234425604343414,
0.05837351456284523,
-0.008802831172943115,
-0.07876796275377274,
-0.003581418888643384,
0.05591435730457306,
-0.0935186967253685,
0.02633844129741192,
-0.04762783274054527,
0.007766661234200001,
-0.05028083547949791,
0.011567210778594017,
0.06461534649133682,
-0.004042435437440872,
0.02125416323542595,
-0.012603728100657463,
-0.0017498363740742207,
0.023616107180714607,
-0.08469576388597488,
0.004391959868371487,
-0.005184575915336609,
0.032849546521902084,
0.017078090459108353,
0.005527828354388475,
-0.03656536713242531,
-0.022258343175053596,
0.05512162670493126,
-0.02706075645983219,
0.01331624761223793,
-0.01694938912987709,
0.04647804796695709,
0.002544899471104145,
-0.12791569530963898,
-0.013675592839717865,
0.0016471993876621127,
0.13594791293144226,
-0.00982764083892107,
0.12390316277742386,
0.00360207655467093,
-0.011102361604571342,
0.02285662293434143,
-0.03484318032860756,
-0.043440449982881546,
-0.10109478235244751,
0.06554633378982544,
0.022963961586356163,
0.020085489377379417,
-0.023086225613951683,
0.003997369669377804,
0.0714593157172203,
-0.0373540073633194,
0.11582716554403305,
-0.015549001283943653,
0.010288408026099205,
0.000494845153298229,
0.049473583698272705,
0.03041321225464344,
0.054306380450725555,
-0.06355050951242447,
0.09659630805253983,
0.03189581632614136,
-0.029598575085401535,
0.06358397006988525,
0.023937612771987915,
0.015282426960766315,
-0.11738645285367966,
0.05891859158873558,
-0.03492160886526108,
-0.022199448198080063,
-0.029376868158578873,
0.03703221306204796,
-0.037661418318748474,
-0.046829402446746826,
-0.01909824274480343,
0.04511522501707077,
0.0009690074366517365,
-0.0450960136950016,
0.029306145384907722,
-0.022713806480169296,
-0.060622476041316986,
-0.016132371500134468,
0.027064265683293343,
0.020464705303311348,
3.191082555424601e-33,
0.09643474966287613,
0.08540425449609756,
0.030290216207504272,
0.02159353345632553,
-0.009323053993284702,
-0.0019833610858768225,
-0.0045323786325752735,
0.0069901770912110806,
-0.05887088179588318,
-0.036021795123815536,
0.0007043001824058592,
0.044853441417217255,
0.046073466539382935,
-0.03540298342704773,
0.0348428338766098,
0.06551094353199005,
-0.019131919369101524,
-0.004633128177374601,
-0.06351327896118164,
0.0068917423486709595,
0.023236844688653946,
-0.011872977018356323,
-0.009069278836250305,
-0.014742192812263966,
-0.13208481669425964,
0.04609904810786247,
0.01732397824525833,
0.007525030057877302,
-0.024205509573221207,
-0.05722762644290924,
-0.03248986229300499,
-0.0026807712856680155,
-0.008482346311211586,
-0.032729171216487885,
-0.0617574006319046,
0.05181103199720383,
0.041598740965127945,
0.005905249621719122,
0.01160902064293623,
0.06116004288196564,
0.015062548220157623,
-0.08867041021585464,
-0.01581432670354843,
0.0319240503013134,
-0.06551413238048553,
-0.02962602488696575,
-0.0539899505674839,
0.09155929833650589,
-0.11135218292474747,
0.08727070689201355,
-0.1328887939453125,
0.12511718273162842,
-0.09756731241941452,
-0.06820207834243774,
-0.016494639217853546,
-0.0012155137956142426,
-0.0917501300573349,
0.03943749517202377,
0.008470655418932438,
-0.05280487239360809,
0.03214804828166962,
0.03582943603396416,
-0.045347198843955994,
0.042987972497940063,
0.032496463507413864,
0.016418971121311188,
-0.038189876824617386,
-0.012002838775515556,
0.05065608024597168,
0.01359524019062519,
0.03760071471333504,
-0.040522780269384384,
0.027716703712940216,
-0.03873078525066376,
-0.011529861949384212,
0.01082567311823368,
0.05019387602806091,
-0.005312168505042791,
0.05154092237353325,
-0.04329261928796768,
0.0287748035043478,
-0.004693258088082075,
0.019996626302599907,
0.058284692466259,
0.07602086663246155,
0.14365287125110626,
0.019424235448241234,
0.026794133707880974,
0.0408623106777668,
0.05050387233495712,
-0.06004824861884117,
-0.01613995060324669,
0.04327672719955444,
0.08190049976110458,
-0.04780324548482895,
-3.912812118755937e-8,
-0.07439317554235458,
0.009142349474132061,
0.04142657667398453,
0.02602636069059372,
0.04494946077466011,
-0.05392354726791382,
0.008272764272987843,
0.18571729958057404,
-0.027575252577662468,
0.04876674339175224,
0.03801117464900017,
0.04733308404684067,
-0.02223987877368927,
-0.09541281312704086,
0.02918834425508976,
0.04233848303556442,
-0.017595931887626648,
-0.015380376018583775,
-0.0417085625231266,
0.026626184582710266,
-0.0026659672148525715,
0.0048318421468138695,
0.06905372440814972,
-0.007369935512542725,
0.03430153802037239,
-0.05015002191066742,
0.027459800243377686,
0.1364881694316864,
0.11765339225530624,
-0.0428420715034008,
0.0046324823051691055,
0.015412427484989166,
-0.06221780553460121,
0.014574293978512287,
-0.02319946698844433,
0.028495153412222862,
0.0038476570043712854,
0.033444229513406754,
-0.048345185816287994,
-0.09511218965053558,
0.03871244937181473,
-0.020153097808361053,
0.05747544392943382,
0.033589623868465424,
-0.022762589156627655,
0.08530224859714508,
-0.010162515565752983,
-0.07637551426887512,
0.05610954016447067,
-0.0576690137386322,
-0.056744929403066635,
-0.011068743653595448,
0.014048947021365166,
0.10189243406057358,
0.017777977511286736,
0.0390738844871521,
0.014211263507604599,
0.039757583290338516,
0.05009975656867027,
0.10718352347612381,
-0.004231179598718882,
0.03596673533320427,
-0.05644771456718445,
0.013127455487847328
] |
Re-Imagining Multimodal Instruction Tuning: A Representation View | https://openreview.net/forum?id=zxg6601zoc | [
"Yiyang Liu",
"James Chenhao Liang",
"Ruixiang Tang",
"Yugyung Lee",
"MAJID RABBANI",
"Sohail Dianat",
"Raghuveer Rao",
"Lifu Huang",
"Dongfang Liu",
"Qifan Wang",
"Cheng Han"
] | Poster | Multimodal instruction tuning has proven to be an effective strategy for achieving zero-shot generalization by fine-tuning pre-trained Large Multimodal Models (LMMs) with instruction-following data. However, as the scale of LMMs continues to grow, fully fine-tuning these models has become highly parameter-intensive. Although Parameter-Efficient Fine-Tuning (PEFT) methods have been introduced to reduce the number of tunable parameters, a significant performance gap remains compared to full fine-tuning. Furthermore, existing PEFT approaches are often highly parameterized, making them difficult to interpret and control. In light of this, we introduce Multimodal Representation Tuning (MRT), a novel approach that focuses on directly editing semantically rich multimodal representations to achieve strong performance and provide intuitive control over LMMs. Empirical results show that our method surpasses current state-of-the-art baselines with significant performance gains (e.g., 1580.40 MME score) while requiring substantially fewer tunable parameters (e.g., 0.03% parameters). Additionally, we conduct experiments on editing instrumental tokens within multimodal representations, demonstrating that direct manipulation of these representations enables simple yet effective control over network behavior. | Representation Tuning, Large Multimodal Models, Parameter-efficient Fine-tuning | Multimodal Representation Tuning for Zero-shot Multimodal Instruction Learning | 1,371 | 2503.00723 | [
0.07868334650993347,
-0.11722609400749207,
0.04734316095709801,
0.006862487178295851,
0.058402687311172485,
0.0037979301996529102,
-0.02387157455086708,
0.003943303599953651,
0.04973699897527695,
-0.061497852206230164,
-0.04388334974646568,
-0.046789783984422684,
0.010376072488725185,
0.02217799983918667,
0.02136404626071453,
-0.01566016860306263,
0.02333107218146324,
0.0677107721567154,
-0.11912928521633148,
-0.02711707353591919,
0.06811293214559555,
-0.012061720713973045,
0.06764001399278641,
-0.06588016450405121,
0.04532982036471367,
0.011746184900403023,
-0.053290728479623795,
-0.03621188551187515,
0.061723221093416214,
-0.06699173897504807,
-0.042906153947114944,
0.053146444261074066,
-0.03585824370384216,
0.06686654686927795,
-0.07236979901790619,
0.08629998564720154,
-0.055661335587501526,
0.05072962865233421,
-0.047443147748708725,
-0.06029338762164116,
-0.030094781890511513,
0.04635489359498024,
0.022365674376487732,
0.043259624391794205,
0.0384848490357399,
-0.026387415826320648,
-0.0221254862844944,
-0.10786593705415726,
-0.015347739681601524,
0.013234738260507584,
-0.10366105288267136,
-0.06595318764448166,
-0.019314555451273918,
0.029679477214813232,
-0.019938863813877106,
-0.011344623751938343,
0.051598843187093735,
0.006751371081918478,
0.03513440489768982,
-0.05016561597585678,
-0.10129552334547043,
-0.03036521188914776,
0.0053838337771594524,
-0.02896014228463173,
0.03724215552210808,
-0.033420901745557785,
0.03355107083916664,
-0.018102461472153664,
0.016502195969223976,
-0.010170064866542816,
-0.06951650977134705,
-0.004095855168998241,
-0.03744421899318695,
0.015744755044579506,
-0.015580281615257263,
0.06027693673968315,
-0.018534336239099503,
-0.0434248261153698,
0.02262355387210846,
-0.08000356703996658,
0.07208080589771271,
-0.059225719422101974,
-0.030450278893113136,
-0.017864299938082695,
0.09859473258256912,
0.02931775152683258,
0.08127649873495102,
-0.00960578303784132,
0.07410449534654617,
0.06211739405989647,
-0.05813340097665787,
-0.015424893237650394,
-0.014130122028291225,
-0.018059736117720604,
0.0013028646353632212,
0.06893472373485565,
0.037616219371557236,
-0.08946701139211655,
-0.03632892668247223,
0.06832446157932281,
0.0857153907418251,
0.032283902168273926,
-0.016717862337827682,
0.01776222325861454,
0.004511809442192316,
-0.02504844218492508,
0.1319977343082428,
0.05377555266022682,
-0.026136059314012527,
-0.06290409713983536,
0.0400765985250473,
0.057996444404125214,
-0.0016758375568315387,
0.009638377465307713,
0.0124257393181324,
0.01313552726060152,
0.028188250958919525,
-0.043032608926296234,
0.050683531910181046,
0.043475281447172165,
-0.018253624439239502,
-0.004055331461131573,
-0.0798409953713417,
-0.0030966217163950205,
0.036820389330387115,
-0.02512413263320923,
-0.09976940602064133,
1.2431784890231694e-33,
0.061094604432582855,
0.03063952922821045,
0.036136217415332794,
0.03607608750462532,
0.04587345942854881,
-0.03684750199317932,
-0.013367107138037682,
0.08030089735984802,
-0.06203658506274223,
0.021984122693538666,
0.0395115464925766,
-0.01723862625658512,
-0.042692702263593674,
0.15278120338916779,
-0.013837446458637714,
-0.01728001795709133,
-0.09332912415266037,
0.021229678764939308,
-0.00003295300484751351,
0.027376342564821243,
-0.008111819624900818,
0.006432827562093735,
-0.047807030379772186,
-0.015204641036689281,
-0.03507603704929352,
0.04196707904338837,
0.041500523686409,
-0.09472024440765381,
-0.0673128291964531,
-0.010324982926249504,
-0.16057775914669037,
-0.0030177952721714973,
-0.004450555890798569,
-0.021984154358506203,
0.008649774827063084,
-0.12610018253326416,
-0.0007056482136249542,
-0.01860264129936695,
0.04991377145051956,
-0.019849658012390137,
-0.017704911530017853,
-0.00929627288132906,
-0.009328258223831654,
-0.017898263409733772,
-0.06792201101779938,
-0.04289477691054344,
-0.02038581296801567,
0.07549839466810226,
-0.05325675383210182,
-0.004675493575632572,
0.01185136940330267,
0.07290574908256531,
0.013300984166562557,
-0.04728462174534798,
-0.023692427203059196,
-0.033576350659132004,
0.07016785442829132,
0.013202410191297531,
-0.029537411406636238,
0.009955814108252525,
-0.014961132779717445,
0.09765299409627914,
0.00183996616397053,
0.10592293739318848,
-0.0033674289006739855,
0.054280836135149,
0.026975644752383232,
0.13157722353935242,
0.0973009243607521,
0.012081312015652657,
-0.12663790583610535,
-0.020755145698785782,
0.04061948508024216,
-0.054169662296772,
-0.04178310185670853,
-0.012740815058350563,
0.057195644825696945,
-0.03556569293141365,
0.05937022343277931,
0.07810168713331223,
0.05265804007649422,
0.03388327732682228,
-0.021905720233917236,
-0.04422932490706444,
0.006023053545504808,
-0.12430285662412643,
0.02465387061238289,
-0.09498070925474167,
0.010995838791131973,
-0.06268046796321869,
-0.02049509808421135,
0.010345470160245895,
-0.0690319612622261,
-0.04206442832946777,
-0.06887096166610718,
-6.326101928167648e-34,
0.07219817489385605,
0.03460784628987312,
-0.0032312709372490644,
0.07092635333538055,
-0.02485346421599388,
-0.054139185696840286,
0.05277758464217186,
0.01821243390440941,
0.040814369916915894,
-0.051987987011671066,
-0.03314295783638954,
-0.04902449995279312,
-0.0397367961704731,
-0.05610131844878197,
0.029851200059056282,
0.007490417454391718,
-0.06624431908130646,
0.03382105752825737,
0.011963329277932644,
0.013964669778943062,
-0.02594071999192238,
0.05006423965096474,
0.017059970647096634,
0.07902200520038605,
-0.021389447152614594,
0.0621386282145977,
-0.04882160201668739,
0.02636084333062172,
-0.0281908567994833,
-0.03802639618515968,
0.009233388118445873,
0.0021945959888398647,
0.09511710703372955,
0.011892862617969513,
-0.0036197761073708534,
0.005201377440243959,
0.07458636164665222,
-0.031542230397462845,
0.026498323306441307,
0.09918204694986343,
0.09840218722820282,
0.015023401007056236,
-0.03487078845500946,
0.007120423950254917,
-0.004564352799206972,
-0.005922906566411257,
-0.08547807484865189,
-0.04188978299498558,
0.012721942737698555,
0.05404720455408096,
-0.09311066567897797,
0.03958965837955475,
-0.08568442612886429,
-0.03695038706064224,
-0.026975879445672035,
-0.14659063518047333,
-0.014023070223629475,
-0.07250102609395981,
0.09002121537923813,
0.009465680457651615,
-0.015524527058005333,
-0.048525337129831314,
-0.023187870159745216,
-0.1261519193649292,
0.001564283506013453,
-0.005997123662382364,
0.0040041073225438595,
-0.08993198722600937,
0.012306543067097664,
0.050482384860515594,
-0.00637799222022295,
-0.01728476583957672,
0.02224528230726719,
-0.004243194125592709,
-0.010522928088903427,
0.026913607493042946,
0.04913584515452385,
-0.026084493845701218,
-0.013693800196051598,
-0.046665795147418976,
0.024619564414024353,
-0.031440429389476776,
-0.01931520365178585,
0.05906585976481438,
0.021381083875894547,
0.1358136683702469,
-0.015398584306240082,
0.14284484088420868,
0.03997018560767174,
-0.0038593304343521595,
-0.029967550188302994,
0.06834746897220612,
0.0928099974989891,
0.07031217962503433,
-0.03805273771286011,
-4.223644367584711e-8,
-0.052076343446969986,
0.018590103834867477,
0.010049564763903618,
0.08727903664112091,
-0.0007605181308463216,
-0.05105569213628769,
-0.008729358203709126,
0.06115955114364624,
0.01285043265670538,
-0.004168371204286814,
0.08614858984947205,
-0.05041429027915001,
-0.0193884689360857,
-0.038540929555892944,
0.036186546087265015,
0.039953816682100296,
-0.028169166296720505,
0.07619769126176834,
-0.014270585030317307,
0.00992165319621563,
0.01675238087773323,
0.06075717508792877,
-0.022422999143600464,
-0.047824129462242126,
0.008529480546712875,
-0.029012013226747513,
0.024387551471590996,
0.09005644172430038,
0.0496242493391037,
-0.0143349040299654,
0.009581842459738255,
0.01404581218957901,
0.04273315146565437,
-0.07117994129657745,
-0.03373456373810768,
0.009583313949406147,
0.07956653833389282,
-0.0065070753917098045,
-0.007721331901848316,
-0.035540420562028885,
0.017586424946784973,
0.04102643206715584,
-0.05226321890950203,
0.03096342645585537,
-0.016133448109030724,
-0.029318513348698616,
-0.035739999264478683,
-0.06569943577051163,
0.02078006975352764,
-0.03608297184109688,
0.012444018386304379,
0.018552886322140694,
-0.028316151350736618,
0.08950331807136536,
0.03886759653687477,
0.0013360782759264112,
0.00316485739313066,
0.06761663407087326,
0.025895802304148674,
0.06467646360397339,
0.01727679930627346,
0.04496286064386368,
-0.05038056522607803,
-0.020168907940387726
] |
Looped Transformers for Length Generalization | https://openreview.net/forum?id=2edigk8yoU | [
"Ying Fan",
"Yilun Du",
"Kannan Ramchandran",
"Kangwook Lee"
] | Poster | Recent work has shown that Transformers trained from scratch can successfully solve various arithmetic and algorithmic tasks, such as adding numbers and computing parity. While these Transformers generalize well on unseen inputs of the same length, they struggle with length generalization, i.e., handling inputs of unseen lengths. In this work, we demonstrate that looped Transformers with an adaptive number of steps significantly improve length generalization. We focus on tasks with a known iterative solution, involving multiple iterations of a RASP-L operation—a length-generalizable operation that can be expressed by a finite-sized Transformer. We train looped Transformers using our proposed learning algorithm and observe that they learn highly length-generalizable solutions for various tasks. | Transformers | null | 1,367 | 2409.15647 | [
-0.08732528984546661,
0.032165415585041046,
0.023653050884604454,
-0.02548837661743164,
-0.029723649844527245,
-0.04468002915382385,
-0.016931749880313873,
0.006434273906052113,
-0.06580962985754013,
-0.06738083809614182,
0.0024332846514880657,
0.02264234609901905,
-0.02263178676366806,
0.06632910668849945,
-0.014764398336410522,
0.05150386691093445,
-0.10258350521326065,
0.11834825575351715,
-0.08805596828460693,
-0.08208848536014557,
-0.01230393536388874,
0.0072183506563305855,
-0.049740035086870193,
0.026989081874489784,
0.08888307958841324,
0.048713333904743195,
-0.00805877149105072,
-0.09799541532993317,
0.031602054834365845,
-0.001167250913567841,
0.029704881832003593,
-0.03185180574655533,
0.047329358756542206,
0.047521330416202545,
-0.01010170578956604,
0.04343758895993233,
-0.09681957215070724,
0.05882734805345535,
0.004595634061843157,
-0.015519571490585804,
0.07674191147089005,
0.01335168443620205,
0.0807860791683197,
-0.00690716877579689,
-0.0029062984976917505,
0.02781756781041622,
0.02827531471848488,
-0.09958426654338837,
-0.008362938649952412,
-0.010528932325541973,
0.0595841184258461,
-0.008580820634961128,
-0.08533120900392532,
0.10945643484592438,
-0.010177815333008766,
-0.05398063361644745,
0.024841060861945152,
-0.0076669929549098015,
-0.05468844994902611,
-0.031457871198654175,
-0.02824496291577816,
-0.08351146429777145,
-0.03719020262360573,
-0.061937082558870316,
-0.08186466991901398,
-0.03182528167963028,
0.05813014134764671,
-0.0016847931547090411,
-0.06849440932273865,
0.032701555639505386,
-0.06633252650499344,
0.061864376068115234,
-0.029832052066922188,
0.0986759141087532,
0.0915173590183258,
0.042805664241313934,
0.1126924604177475,
-0.04139209911227226,
-0.01986328698694706,
-0.01500454731285572,
-0.012161141261458397,
-0.012173548340797424,
0.03711925074458122,
0.08543922752141953,
0.06960495561361313,
-0.02204485796391964,
-0.02571927197277546,
0.05380258336663246,
-0.012276227585971355,
0.003970532212406397,
-0.006292121484875679,
-0.08797752857208252,
-0.00108817289583385,
-0.02764047682285309,
-0.015979282557964325,
0.04291411489248276,
0.01239969115704298,
-0.07720496505498886,
-0.08586179465055466,
0.0604683980345726,
-0.03436344489455223,
-0.05809348449110985,
0.03939605504274368,
-0.03404195234179497,
-0.04668008163571358,
0.018573833629488945,
-0.02938614971935749,
0.012228607200086117,
0.009492181241512299,
-0.12267451733350754,
0.04877400025725365,
0.022697197273373604,
0.023210560902953148,
-0.017590973526239395,
0.05750120431184769,
-0.032663535326719284,
0.00351921608671546,
-0.017850473523139954,
-0.05404883623123169,
0.034625403583049774,
0.029746271669864655,
0.037547990679740906,
-0.057668644934892654,
0.12525621056556702,
-0.0021218634210526943,
0.017842335626482964,
-0.02259369008243084,
1.862142110781849e-33,
-0.05210650712251663,
0.10092538595199585,
0.07725515961647034,
0.028503326699137688,
0.01852966472506523,
0.04805181547999382,
0.024010326713323593,
0.03444555401802063,
0.04099701717495918,
0.09190534055233002,
-0.03765581548213959,
0.062265507876873016,
-0.034150801599025726,
0.035045918077230453,
0.005694510880857706,
-0.00851291324943304,
0.012928202748298645,
0.05611006170511246,
0.006105899810791016,
-0.0798618346452713,
0.03644127398729324,
-0.04029195383191109,
0.022452542558312416,
-0.02179049141705036,
-0.005813521333038807,
-0.07421933114528656,
0.050618257373571396,
-0.02295895852148533,
-0.003197661368176341,
-0.052646514028310776,
-0.08401314914226532,
0.03281330317258835,
0.00962833408266306,
0.022032788023352623,
0.0592266209423542,
-0.08861910551786423,
0.06906462460756302,
-0.05241864547133446,
0.006554516032338142,
0.02947457879781723,
0.01967739872634411,
-0.015976659953594208,
-0.01429419219493866,
0.04208460450172424,
-0.05112511292099953,
0.007381992880254984,
0.007630562409758568,
0.01663438230752945,
-0.05523860454559326,
0.015545416623353958,
-0.03273874893784523,
-0.002792246174067259,
-0.0875139981508255,
-0.013864429667592049,
0.13583417236804962,
0.055383533239364624,
-0.001531558227725327,
0.06529495865106583,
0.0623016431927681,
0.011484876275062561,
0.011882963590323925,
0.07250705361366272,
-0.049147725105285645,
0.04216989129781723,
-0.08170384168624878,
0.024537498131394386,
0.04713383689522743,
0.03476718068122864,
0.03826531767845154,
-0.03491431847214699,
-0.07092750817537308,
-0.030014745891094208,
-0.05106852576136589,
-0.0013853141572326422,
0.086344875395298,
-0.07863322645425797,
0.08216764032840729,
-0.0729069784283638,
-0.0019395191920921206,
-0.014660682529211044,
-0.025462234392762184,
0.05765129625797272,
0.05152472108602524,
-0.0133138969540596,
0.022196652367711067,
-0.05014186352491379,
0.032788682729005814,
-0.08839229494333267,
-0.032158903777599335,
-0.10292535275220871,
-0.12701494991779327,
0.022134102880954742,
-0.011471275240182877,
0.011988854967057705,
0.008332430385053158,
-2.2039183789803754e-33,
-0.002478264272212982,
0.011247343383729458,
-0.08060554414987564,
0.03135533630847931,
0.0321231372654438,
-0.04201047122478485,
-0.03536219149827957,
0.029889220371842384,
0.005724696442484856,
-0.01871640235185623,
-0.023116527125239372,
-0.0010888668475672603,
-0.04375644773244858,
0.00035705167101696134,
0.04895928502082825,
0.032054442912340164,
-0.04908324405550957,
-0.007491776254028082,
-0.010069776326417923,
0.00931573286652565,
-0.005489748902618885,
0.11418700218200684,
-0.04587997496128082,
-0.019868794828653336,
-0.09590122848749161,
0.01826615072786808,
-0.07024270296096802,
0.07174801081418991,
0.06360185891389847,
0.05886359140276909,
-0.011854174546897411,
-0.02838028222322464,
-0.004464771132916212,
0.008820579387247562,
-0.0366831012070179,
0.013332477770745754,
0.034674253314733505,
0.05030474811792374,
0.04152873158454895,
0.03597366437315941,
-0.0099685313180089,
-0.13720181584358215,
0.04711312800645828,
0.06800159066915512,
-0.059809695929288864,
-0.04779211804270744,
0.023880988359451294,
-0.024237360805273056,
-0.022840337827801704,
0.031233172863721848,
-0.04148014634847641,
-0.007606415078043938,
-0.10869987308979034,
-0.05418987572193146,
0.024198785424232483,
-0.03565918654203415,
-0.019035913050174713,
-0.023881463333964348,
0.11761486530303955,
-0.03486563637852669,
-0.07028874009847641,
-0.060689568519592285,
0.08311370015144348,
-0.05946507304906845,
0.011684183031320572,
0.05935727059841156,
-0.0016173928743228316,
-0.03721245005726814,
0.01496659591794014,
0.06285693496465683,
0.039868421852588654,
-0.007640808820724487,
-0.03433983400464058,
-0.034479688853025436,
-0.09200741350650787,
-0.041693754494190216,
-0.004003163427114487,
-0.030719595029950142,
0.04098940268158913,
-0.03473609313368797,
-0.00394787872210145,
0.018879467621445656,
0.009371974505484104,
0.0180758498609066,
-0.011058810167014599,
0.02528478018939495,
0.08699212968349457,
0.08499426394701004,
0.05621163547039032,
0.0369623564183712,
-0.04028002917766571,
0.11392797529697418,
0.034666504710912704,
0.05466528609395027,
-0.03628646954894066,
-3.680513671611152e-8,
-0.060613133013248444,
0.030326707288622856,
-0.07296382635831833,
-0.013017548248171806,
0.061921391636133194,
-0.018577560782432556,
0.02934805117547512,
0.09516388177871704,
0.008202861994504929,
-0.031282659620046616,
0.1010391116142273,
-0.029374759644269943,
0.018446698784828186,
0.029309570789337158,
-0.0009490061202086508,
0.06844738125801086,
-0.012776561081409454,
-0.03989928588271141,
-0.015571828000247478,
0.04836466908454895,
-0.04924137145280838,
0.06249651312828064,
-0.020111462101340294,
-0.04636894538998604,
-0.002574261976405978,
-0.09449342638254166,
-0.04249973222613335,
0.04295598715543747,
0.06841831654310226,
0.05702383071184158,
0.013994641602039337,
0.03506357967853546,
0.0029223235324025154,
0.01509101502597332,
0.02356865629553795,
0.07107172161340714,
0.06943179666996002,
-0.034298792481422424,
0.01949373446404934,
-0.06027373671531677,
-0.011532559990882874,
0.06393491476774216,
-0.029836783185601234,
0.0362187922000885,
-0.051249049603939056,
-0.03494361788034439,
0.03287786990404129,
-0.11041576415300369,
-0.028601504862308502,
0.012864391319453716,
0.06765894591808319,
0.0024187234230339527,
0.016333509236574173,
-0.01307964138686657,
0.08875419944524765,
-0.010887705720961094,
0.007366883102804422,
-0.10154981166124344,
0.03108133003115654,
0.11320426315069199,
-0.12486779689788818,
0.0073129585944116116,
-0.02832021564245224,
0.028712479397654533
] |
Exploring channel distinguishability in local neighborhoods of the model space in quantum neural networks | https://openreview.net/forum?id=gDcL7cgZBt | [
"Sabrina Herbst",
"Sandeep Suresh Cranganore",
"Vincenzo De Maio",
"Ivona Brandić"
] | Poster | With the increasing interest in Quantum Machine Learning, Quantum Neural Networks (QNNs) have emerged and gained significant attention. These models have, however, been shown to be notoriously difficult to train, which we hypothesize is partially due to the architectures, called ansatzes, that are hardly studied at this point. Therefore, in this paper, we take a step back and analyze ansatzes. We initially consider their expressivity, i.e., the space of operations they are able to express, and show that the closeness to being a 2-design, the primarily used measure, fails at capturing this property. Hence, we look for alternative ways to characterize ansatzes, unrelated to expressivity, by considering the local neighborhood of the model space, in particular, analyzing model distinguishability upon small perturbation of parameters. We derive an upper bound on their distinguishability, showcasing that QNNs using the Hardware Efficient Ansatz with few parameters are hardly discriminable upon update. Our numerical experiments support our bounds and further indicate that there is a significant degree of variability, which stresses the need for warm-starting or clever initialization. Altogether, our work provides an ansatz-centric perspective on training dynamics and difficulties in QNNs, ultimately suggesting that iterative training of small quantum models may not be effective, which contrasts their initial motivation. | Quantum Machine Learning, Quantum Neural Network | null | 1,364 | 2410.09470 | [
-0.05295011028647423,
-0.09161689877510071,
0.003783503780141473,
0.05485432595014572,
-0.05584649369120598,
0.06275728344917297,
-0.0032934502232819796,
-0.05232726410031319,
0.05102282762527466,
-0.07792146503925323,
-0.04114942625164986,
-0.03132278844714165,
-0.017710832878947258,
0.04458407685160637,
-0.10886852443218231,
-0.019908953458070755,
0.08237156271934509,
-0.005722924135625362,
-0.08377111703157425,
-0.009682298637926579,
0.007572002708911896,
0.03853454813361168,
0.02192898839712143,
0.019960742443799973,
0.07364165782928467,
-0.047137029469013214,
-0.011672439984977245,
0.08073516935110092,
0.02438056468963623,
-0.015373480506241322,
0.05628499761223793,
-0.006734889000654221,
-0.021984988823533058,
0.008905294351279736,
0.02718420699238777,
0.044724009931087494,
-0.0011110081104561687,
-0.03607993200421333,
0.034212302416563034,
-0.07026667892932892,
0.0667322501540184,
0.022433243691921234,
-0.0033518425188958645,
0.06428491324186325,
0.10235347598791122,
0.058955151587724686,
0.0710100457072258,
0.0251675546169281,
-0.08633892983198166,
-0.16866353154182434,
-0.011796414852142334,
0.00031765669700689614,
-0.003163520246744156,
0.05094648152589798,
0.01990051381289959,
-0.005598903633654118,
0.0003284169069956988,
0.04978393018245697,
-0.08971412479877472,
-0.00958920270204544,
-0.05242418497800827,
-0.03918362408876419,
-0.004255180712789297,
0.012659984640777111,
0.020440945401787758,
0.026946958154439926,
-0.03092137910425663,
0.03919921815395355,
0.012763099744915962,
0.07620638608932495,
0.016907932236790657,
0.02904677763581276,
-0.04861653968691826,
0.011721753515303135,
0.044527385383844376,
-0.010606938973069191,
0.09991505742073059,
0.05111722648143768,
0.018600372597575188,
-0.061463892459869385,
-0.029557473957538605,
-0.1122831478714943,
-0.02091830037534237,
0.005862713325768709,
0.07452914863824844,
-0.001444835215806961,
-0.04930480942130089,
-0.003915513399988413,
0.023234663531184196,
-0.09097398817539215,
-0.09253427386283875,
-0.0012210496934130788,
-0.014845868572592735,
-0.07867155969142914,
0.06419412046670914,
0.02921888418495655,
0.04141536355018616,
-0.029492473229765892,
0.014751357957720757,
0.09105392545461655,
0.016253001987934113,
0.05456554889678955,
0.06499874591827393,
-0.014024917967617512,
0.0655900239944458,
0.04482651874423027,
-0.01815507374703884,
0.009719456546008587,
0.05197085440158844,
-0.12425245344638824,
-0.0006156194140203297,
0.02351555787026882,
0.006521128583699465,
0.02501710131764412,
0.059613920748233795,
0.026192961260676384,
0.01135926228016615,
0.08315998315811157,
0.0397607758641243,
0.07071590423583984,
-0.05292484909296036,
-0.020392201840877533,
-0.06186231970787048,
0.05194760113954544,
-0.009971562772989273,
0.0051363068632781506,
-0.14443425834178925,
4.184860907973117e-34,
0.04754914343357086,
0.04397805780172348,
-0.015075603500008583,
0.04187046363949776,
0.09216626733541489,
-0.030441628769040108,
0.11101056635379791,
-0.013154256157577038,
0.040562015026807785,
0.047288764268159866,
-0.024499991908669472,
0.004145399201661348,
-0.03190915659070015,
0.0030572237446904182,
0.07324712723493576,
0.06491896510124207,
-0.07847724854946136,
-0.05646786093711853,
-0.046565063297748566,
-0.06961969286203384,
0.040481068193912506,
-0.06773011386394501,
-0.07049480825662613,
0.02576640248298645,
-0.03704363852739334,
-0.09563534706830978,
0.07227973639965057,
0.03834352269768715,
-0.11668683588504791,
-0.01153846736997366,
-0.06634929776191711,
0.05805172398686409,
-0.014883887022733688,
0.007148046046495438,
0.045131485909223557,
-0.0128254946321249,
0.035434138029813766,
0.02853691205382347,
0.01741456426680088,
-0.059306588023900986,
0.02546127513051033,
0.006320029962807894,
0.03088025376200676,
0.009591259993612766,
0.0024954702239483595,
-0.02768782526254654,
-0.023859400302171707,
-0.0361473374068737,
0.0044981688261032104,
-0.07612384110689163,
0.045871783047914505,
0.014038692228496075,
0.018008554354310036,
-0.0637751892209053,
0.018766045570373535,
-0.07220359891653061,
0.05538994446396828,
0.07973985373973846,
0.0012879475252702832,
0.07386711984872818,
-0.031261883676052094,
-0.0024443306028842926,
-0.024929551407694817,
0.010909155942499638,
-0.014567295089364052,
0.06092425435781479,
-0.02787735126912594,
0.06846322119235992,
-0.0011392671149224043,
0.04888518899679184,
-0.04022059962153435,
0.024494199082255363,
-0.033013395965099335,
-0.08253821730613708,
0.11581962555646896,
-0.022028792649507523,
0.005731699988245964,
-0.06563674658536911,
0.008275453932583332,
-0.02298949658870697,
0.0253435205668211,
0.008590520359575748,
-0.036116138100624084,
0.0928407683968544,
-0.07085815817117691,
-0.11624285578727722,
0.03232515975832939,
-0.04617784172296524,
-0.018777897581458092,
-0.031611476093530655,
-0.01631022058427334,
-0.014442086219787598,
0.07261843979358673,
-0.03741391375660896,
-0.044709060341119766,
-1.6106432577134975e-33,
-0.0904395580291748,
0.05360334366559982,
-0.07604973763227463,
0.008784927427768707,
-0.07722805440425873,
-0.026612555608153343,
0.004481700249016285,
-0.005290356930345297,
-0.05623055621981621,
-0.03306780010461807,
-0.025061339139938354,
0.04696492478251457,
0.04624813422560692,
0.007086677011102438,
-0.02072044089436531,
0.054836876690387726,
-0.004795711021870375,
-0.029289329424500465,
0.09241002798080444,
0.006756328046321869,
0.07496856153011322,
0.07958637177944183,
-0.06669730693101883,
0.0008005890995264053,
-0.055866699665784836,
0.0806213915348053,
-0.042194873094558716,
0.03887432441115379,
0.043622326105833054,
-0.03741598129272461,
-0.0385323129594326,
-0.09744758903980255,
-0.055739834904670715,
-0.032870274037122726,
-0.015111240558326244,
0.05387574061751366,
0.043707337230443954,
0.0022737623658031225,
-0.01335203368216753,
0.023600591346621513,
0.009155548177659512,
-0.07420411705970764,
-0.0776832327246666,
0.0560440868139267,
0.07557689398527145,
-0.025603432208299637,
0.009208343923091888,
-0.04875429347157478,
-0.09569770842790604,
0.0313926599919796,
0.01517297513782978,
0.09435184299945831,
0.0016359984874725342,
0.00122129509691149,
-0.12657631933689117,
0.014180513098835945,
-0.009593001566827297,
0.08636067062616348,
0.1346898078918457,
0.009092104621231556,
-0.005449555814266205,
-0.06688719987869263,
-0.03827117010951042,
0.02124840021133423,
-0.07493960857391357,
-0.03162834048271179,
-0.021915489807724953,
0.03204305097460747,
0.08760422468185425,
0.0022515221498906612,
-0.07176884263753891,
-0.06972767412662506,
-0.019767437130212784,
-0.010635138489305973,
-0.064847432076931,
0.02641148492693901,
0.04887595772743225,
-0.04269680380821228,
-0.02786651812493801,
-0.005907351616770029,
0.000567883369512856,
0.038648076355457306,
-0.028174228966236115,
0.005159834399819374,
0.07563081383705139,
0.06336724013090134,
0.028985241428017616,
0.04254154860973358,
0.031147686764597893,
-0.0409863255918026,
0.0636897012591362,
0.08896521478891373,
-0.028623931109905243,
-0.0053755599074065685,
-0.0388396792113781,
-4.586688717722609e-8,
-0.03552742674946785,
0.024162881076335907,
-0.006316693499684334,
0.0262744203209877,
0.0677061676979065,
-0.06703203171491623,
0.051151495426893234,
0.04556412622332573,
-0.06927739828824997,
-0.02499876730144024,
0.024389920756220818,
0.047811754047870636,
-0.0439947284758091,
0.0008472545305266976,
-0.035438284277915955,
0.06536650657653809,
-0.05247357860207558,
-0.03501031920313835,
-0.009682660922408104,
0.0199250727891922,
0.0038333141710609198,
0.06853363662958145,
-0.020234229043126106,
-0.03661877289414406,
-0.01701531745493412,
-0.13106384873390198,
-0.06274182349443436,
0.023040704429149628,
0.04722205549478531,
0.03903287276625633,
0.040205929428339005,
0.019844353199005127,
0.07915985584259033,
0.03716352954506874,
-0.026068774983286858,
0.07358533889055252,
-0.047337085008621216,
-0.03969031199812889,
-0.024580130353569984,
0.0007096381159499288,
-0.04112684354186058,
0.011848974972963333,
-0.0026807370595633984,
0.06589999049901962,
0.026474671438336372,
-0.06721338629722595,
0.036911942064762115,
-0.07980530709028244,
0.05446233972907066,
0.047901712357997894,
0.03150276467204094,
0.013321579433977604,
-0.03239737078547478,
-0.062050770968198776,
-0.05873168259859085,
-0.003869358915835619,
-0.01342456042766571,
-0.09507075697183609,
-0.04518644139170647,
0.05624959617853165,
0.016795719042420387,
0.05988944694399834,
-0.06748805195093155,
0.010437002405524254
] |
Does Spatial Cognition Emerge in Frontier Models? | https://openreview.net/forum?id=WK6K1FMEQ1 | [
"Santhosh Kumar Ramakrishnan",
"Erik Wijmans",
"Philipp Kraehenbuehl",
"Vladlen Koltun"
] | Poster | Not yet. We present SPACE, a benchmark that systematically evaluates spatial cognition in frontier models. Our benchmark builds on decades of research in cognitive science. It evaluates large-scale mapping abilities that are brought to bear when an organism traverses physical environments, smaller-scale reasoning about object shapes and layouts, and cognitive infrastructure such as spatial attention and memory. For many tasks, we instantiate parallel presentations via text and images, allowing us to benchmark both large language models and large multimodal models. Results suggest that contemporary frontier models fall short of the spatial intelligence of animals, performing near chance level on a number of classic tests of animal cognition. | Frontier models, spatial cognition | We present SPACE, a benchmark that systematically evaluates spatial cognition in frontier models. | 1,357 | 2410.06468 | [
0.13715806603431702,
-0.13829369843006134,
0.06378599256277084,
-0.015930544584989548,
0.02545781619846821,
-0.021818511188030243,
-0.06904490292072296,
-0.0031542477663606405,
0.058524202555418015,
0.11680769920349121,
-0.03504878282546997,
-0.0020298222079873085,
0.01125355251133442,
0.1501305103302002,
0.00841858983039856,
-0.05147205665707588,
-0.004687965847551823,
0.012070951983332634,
-0.0322650671005249,
-0.014623685739934444,
-0.0034045937936753035,
0.02039174735546112,
0.08438337594270706,
0.00004674710726249032,
0.061905741691589355,
-0.01201925240457058,
-0.02349293977022171,
-0.04106734320521355,
-0.019508520141243935,
0.022377224639058113,
0.033812012523412704,
0.0076478589326143265,
0.008912920951843262,
0.03883831575512886,
-0.021144453436136246,
0.015552097000181675,
-0.00023245577176567167,
0.008443575352430344,
0.04341082647442818,
-0.04148875176906586,
-0.04067844897508621,
-0.008957047946751118,
0.04061379283666611,
0.007039356045424938,
0.06270261853933334,
-0.006278162822127342,
-0.0220391433686018,
-0.059233006089925766,
0.023776696994900703,
-0.06017692759633064,
-0.09817355126142502,
-0.027193425223231316,
-0.07629154622554779,
-0.014429024420678616,
-0.036364249885082245,
0.0539298914372921,
0.06572267413139343,
-0.06526575982570648,
0.027974557131528854,
-0.032914940267801285,
0.003925956320017576,
-0.07936310023069382,
0.038099609315395355,
0.03247473016381264,
0.029617585241794586,
0.011166660115122795,
-0.09291627258062363,
-0.007389097474515438,
-0.050188545137643814,
-0.05894734710454941,
0.026820030063390732,
0.009934486821293831,
0.028053022921085358,
-0.057796914130449295,
0.09344018995761871,
0.015534620732069016,
-0.03993067145347595,
-0.01707293465733528,
0.04793216660618782,
-0.1389658898115158,
0.050372667610645294,
0.025659969076514244,
-0.02139422670006752,
0.003847569925710559,
0.11594609916210175,
0.007954989559948444,
-0.03555617108941078,
0.05889621004462242,
-0.04096715524792671,
-0.07311797142028809,
0.0593295581638813,
-0.13609285652637482,
0.008705132640898228,
-0.012577365152537823,
-0.043540362268686295,
-0.04262121021747589,
0.006322004366666079,
-0.11598559468984604,
-0.050305504351854324,
0.010027769021689892,
0.01939419098198414,
-0.012902387417852879,
0.04033314809203148,
-0.058128081262111664,
0.08607097715139389,
-0.054899848997592926,
0.05995895713567734,
0.01598222367465496,
0.014255632646381855,
-0.0432475209236145,
-0.007269886787980795,
-0.015142267569899559,
-0.05010758712887764,
0.16901051998138428,
-0.05282523110508919,
-0.04670383036136627,
0.011501284316182137,
-0.04394805431365967,
0.07828572392463684,
-0.04412249103188515,
-0.0004784294287674129,
0.02809891104698181,
0.04164626821875572,
-0.015239016152918339,
0.009983600117266178,
-0.034358978271484375,
-0.04317875951528549,
1.1971720048960111e-34,
-0.016556473448872566,
-0.051822394132614136,
0.06884891539812088,
0.03919316083192825,
0.09184128046035767,
-0.023769795894622803,
-0.01633339561522007,
-0.05920488014817238,
0.0013325178297236562,
-0.006217833608388901,
-0.048241086304187775,
0.07993138581514359,
0.019706042483448982,
0.10740868747234344,
0.017394978553056717,
0.04124539718031883,
-0.036288097500801086,
0.07512109726667404,
-0.036640655249357224,
0.025634640827775,
-0.009004437364637852,
0.00821810681372881,
-0.006109076086431742,
-0.04907405376434326,
0.04221723973751068,
0.004000459797680378,
-0.07982560992240906,
-0.06518561393022537,
-0.09349195659160614,
0.020173247903585434,
-0.14915628731250763,
0.017581181600689888,
-0.059709273278713226,
0.04958436265587807,
-0.008653069846332073,
-0.07398447394371033,
0.026322614401578903,
-0.07789977639913559,
0.022612232714891434,
-0.03952392563223839,
-0.009913064539432526,
0.00991174764931202,
-0.0341157503426075,
0.01028650812804699,
-0.008830456994473934,
0.0282965786755085,
0.037860967218875885,
0.07321079820394516,
-0.022440219298005104,
0.09166525304317474,
0.039325810968875885,
0.03788061812520027,
0.00870408583432436,
-0.06542662531137466,
0.031523775309324265,
0.03719514608383179,
0.0133823798969388,
-0.002231005812063813,
0.08485192060470581,
0.027929337695240974,
0.005905591882765293,
0.05560891330242157,
-0.01915820688009262,
0.009256665594875813,
0.09569339454174042,
-0.02931489795446396,
-0.03095429204404354,
0.06303377449512482,
0.04290863499045372,
0.08635739982128143,
0.017467379570007324,
0.009887455962598324,
-0.06473294645547867,
-0.042916178703308105,
0.010375653393566608,
-0.03606046736240387,
0.058804549276828766,
-0.024998199194669724,
-0.0570950023829937,
-0.02902921475470066,
0.001927312114275992,
-0.05915975198149681,
-0.06953100115060806,
-0.004807811696082354,
0.00955614447593689,
0.016763439401984215,
0.0742892473936081,
-0.08105073124170303,
-0.04953940585255623,
-0.03777209669351578,
-0.011224843561649323,
-0.03877978399395943,
-0.015508159063756466,
-0.047451648861169815,
-0.01332144159823656,
-1.1728006320489813e-33,
-0.058326419442892075,
-0.061321500688791275,
-0.020432015880942345,
0.008725878782570362,
-0.0752771645784378,
-0.06397125124931335,
0.012677174992859364,
-0.026050381362438202,
-0.07409282773733139,
-0.0845676138997078,
-0.08292704820632935,
0.04456206411123276,
0.05969756841659546,
0.002689522923901677,
0.055332984775304794,
-0.06220156326889992,
0.07167579233646393,
0.05027037113904953,
0.05786513909697533,
-0.05701299011707306,
0.018276356160640717,
0.023736370727419853,
-0.015577995218336582,
-0.004686411004513502,
-0.09719632565975189,
0.07236608862876892,
-0.058430060744285583,
-0.030464453622698784,
-0.10142001509666443,
-0.015994912013411522,
-0.04534512013196945,
0.005307680461555719,
0.05172371491789818,
0.005590733140707016,
-0.02390243299305439,
0.020787935703992844,
0.05938154086470604,
-0.08814936876296997,
-0.09032910317182541,
0.023068580776453018,
0.015302120707929134,
-0.02593490108847618,
-0.02005302906036377,
0.01138810720294714,
-0.0406641885638237,
0.032724298536777496,
-0.02673993445932865,
0.03956252336502075,
-0.028638752177357674,
0.1113552525639534,
0.028483157977461815,
0.05211286619305611,
-0.045252639800310135,
-0.07744484394788742,
0.011926875449717045,
-0.03958549350500107,
-0.030239099636673927,
-0.029012411832809448,
0.020113972947001457,
-0.027823250740766525,
-0.033036377280950546,
-0.012405220419168472,
-0.05382755026221275,
0.029549764469265938,
0.04277120158076286,
-0.0027871797792613506,
-0.05636031553149223,
0.009683927521109581,
0.035058047622442245,
0.027291040867567062,
0.02346918359398842,
0.009756629355251789,
-0.04778324440121651,
-0.001927920151501894,
0.03609776869416237,
0.10389332473278046,
0.07858984917402267,
-0.030902443453669548,
0.04773646220564842,
-0.014980291947722435,
0.02024407684803009,
-0.04763633757829666,
0.013396048918366432,
0.0861542597413063,
0.036101605743169785,
0.11362817138433456,
-0.04747426137328148,
-0.008578375913202763,
0.04847997426986694,
0.02156190015375614,
-0.01896063983440399,
0.048028070479631424,
-0.03423754498362541,
0.024495622143149376,
-0.057153284549713135,
-3.3896657924969986e-8,
-0.09913654625415802,
0.11021128296852112,
0.07508710026741028,
-0.007255226839333773,
-0.010280276648700237,
0.023479124531149864,
-0.014015206135809422,
0.07679541409015656,
-0.07077133655548096,
0.03224753960967064,
0.04516324773430824,
-0.02955455332994461,
0.02146357297897339,
0.0759720429778099,
0.07101056724786758,
0.02613215707242489,
0.03750917315483093,
0.05058928206562996,
-0.010360628366470337,
0.060384225100278854,
-0.019758358597755432,
0.051973067224025726,
0.0036650572437793016,
-0.023860536515712738,
-0.004360836464911699,
-0.044194478541612625,
-0.0329362191259861,
0.08367669582366943,
-0.00012934736150782555,
-0.06705111265182495,
0.04394552484154701,
-0.01769649051129818,
-0.08052695542573929,
0.024795571342110634,
0.05460422858595848,
-0.06667357683181763,
-0.01914859376847744,
-0.022966474294662476,
-0.004098749253898859,
-0.016452522948384285,
0.006339529529213905,
0.04034499078989029,
0.004912712145596743,
0.056738048791885376,
0.07720833271741867,
0.09553959965705872,
0.011732647195458412,
-0.0642424151301384,
-0.00834559090435505,
-0.01649031601846218,
-0.09237345308065414,
0.0478229820728302,
-0.05961955338716507,
0.038977935910224915,
-0.002176921349018812,
0.028733350336551666,
0.00021092452516313642,
-0.03866536542773247,
0.00731875654309988,
0.0798271968960762,
0.053712066262960434,
0.0988900363445282,
-0.09523775428533554,
0.015280942432582378
] |
Adaptive teachers for amortized samplers | https://openreview.net/forum?id=BdmVgLMvaf | [
"Minsu Kim",
"Sanghyeok Choi",
"Taeyoung Yun",
"Emmanuel Bengio",
"Leo Feng",
"Jarrid Rector-Brooks",
"Sungsoo Ahn",
"Jinkyoo Park",
"Nikolay Malkin",
"Yoshua Bengio"
] | Poster | Amortized inference is the task of training a parametric model, such as a neural network, to approximate a distribution with a given unnormalized density where exact sampling is intractable. When sampling is modeled as a sequential decision-making process, reinforcement learning (RL) methods, such as generative flow networks, can be used to train the sampling policy. Off-policy RL training facilitates the discovery of diverse, high-reward candidates, but existing methods still face challenges in efficient exploration. We propose to use an adaptive training distribution (the Teacher) to guide the training of the primary amortized sampler (the Student). The Teacher, an auxiliary behavior model, is trained to sample high-loss regions of the Student and can generalize across unexplored modes, thereby enhancing mode coverage by providing an efficient training curriculum. We validate the effectiveness of this approach in a synthetic environment designed to present an exploration challenge, two diffusion-based sampling tasks, and four biochemical discovery tasks demonstrating its ability to improve sample efficiency and mode coverage. Source code is available at https://github.com/alstn12088/adaptive-teacher. | amortized inference, generative models, reinforcement learning, GFlowNets | We guide the training or amortized sequential samplers with a adaptive teacher model, leading to better mode coverage on a wide range of problems. | 1,343 | 2410.01432 | [
-0.10942983627319336,
-0.04715844988822937,
0.04548001289367676,
0.04333208128809929,
0.025924231857061386,
-0.007436000742018223,
-0.022025026381015778,
0.0018544256454333663,
0.031164029613137245,
-0.002710677683353424,
-0.07841768860816956,
-0.026644622907042503,
-0.024958830326795578,
0.07068249583244324,
-0.011998025700449944,
-0.009727767668664455,
0.04167972505092621,
0.07381651550531387,
-0.008277860470116138,
0.011570594273507595,
0.024296866729855537,
0.032374393194913864,
0.07746120542287827,
0.0454767569899559,
0.02798428013920784,
0.03815975785255432,
0.051220204681158066,
-0.11066914349794388,
-0.0036569125950336456,
-0.023252593353390694,
0.1267050951719284,
0.034338269382715225,
0.029093438759446144,
-0.07140213251113892,
0.0057172346860170364,
0.005849080625921488,
-0.08657043427228928,
0.04203755408525467,
-0.01174259651452303,
-0.0021072733215987682,
0.041175320744514465,
0.007335877977311611,
-0.025508426129817963,
0.022973082959651947,
0.07907813042402267,
-0.05518394336104393,
0.04394597187638283,
-0.06437159329652786,
0.08049660921096802,
0.06331844627857208,
0.009624402970075607,
-0.0734180137515068,
-0.03123176097869873,
-0.01821012794971466,
0.033066023141145706,
0.01919323392212391,
0.05793236196041107,
0.01634039729833603,
-0.004848575685173273,
-0.04153461754322052,
0.004877016879618168,
-0.09319976717233658,
-0.04560808837413788,
-0.005321071483194828,
-0.08230820298194885,
-0.030769817531108856,
0.013352073729038239,
0.14109188318252563,
0.08992629498243332,
-0.026739822700619698,
0.016239428892731667,
0.07875552773475647,
-0.04290878400206566,
-0.018910005688667297,
0.07722447067499161,
-0.02006799355149269,
-0.025340190157294273,
-0.04419504478573799,
0.046815212815999985,
-0.040198151022195816,
-0.032013118267059326,
-0.033889174461364746,
-0.006864830851554871,
-0.03192936256527901,
0.05205685645341873,
0.000447581842308864,
-0.0031846340280026197,
0.07261835038661957,
0.06165854632854462,
-0.01193928997963667,
0.010684243403375149,
0.10476105660200119,
-0.08042335510253906,
-0.02098601497709751,
-0.04303490370512009,
0.0008679988677613437,
0.007760418578982353,
-0.04891087859869003,
0.020170418545603752,
0.07115499675273895,
0.026324711740016937,
0.08542534708976746,
-0.004054951015859842,
-0.025211993604898453,
-0.029186557978391647,
-0.1094239205121994,
0.04037370905280113,
0.02062540501356125,
0.06952062994241714,
0.009418100118637085,
0.04696398228406906,
0.000892446783836931,
0.06523904204368591,
0.05237998440861702,
-0.0027665896341204643,
0.004087367560714483,
0.010966840200126171,
-0.026246754452586174,
-0.052249740809202194,
0.0957086905837059,
-0.023511508479714394,
-0.04467099532485008,
-0.03617151454091072,
-0.02276061289012432,
0.08757863193750381,
-0.0591864250600338,
-0.03417952358722687,
3.831566080833479e-34,
0.07925204932689667,
-0.012976537458598614,
0.03823816031217575,
0.024756664410233498,
0.014750923961400986,
0.009648015722632408,
0.0926460474729538,
0.021962415426969528,
0.04526335373520851,
0.009552491828799248,
-0.03232273831963539,
0.020748216658830643,
-0.08550223708152771,
0.1576642096042633,
0.03496597334742546,
-0.037472471594810486,
-0.08552379161119461,
0.0367315448820591,
-0.00282855867408216,
-0.040441155433654785,
-0.01976357214152813,
-0.06103880703449249,
-0.03880714997649193,
-0.06781990826129913,
-0.01476891990751028,
0.027229465544223785,
0.04808914661407471,
0.02587411180138588,
-0.008301676250994205,
0.020160526037216187,
-0.03687328100204468,
-0.006708825472742319,
-0.07960502803325653,
0.017983127385377884,
0.04224206507205963,
0.07282738387584686,
0.005329403094947338,
-0.0001948190329130739,
-0.022207166999578476,
-0.0039240350015461445,
-0.003474377328529954,
0.009888377971947193,
0.1217552199959755,
0.030474521219730377,
-0.11902156472206116,
-0.06505768746137619,
0.03538239002227783,
-0.0029269063379615545,
0.011450149118900299,
-0.015438546426594257,
-0.01924288645386696,
-0.0658070519566536,
0.004848690237849951,
-0.11380144208669662,
-0.025829603895545006,
0.05149190500378609,
-0.03661925718188286,
0.0372246578335762,
-0.017148515209555626,
0.050492700189352036,
0.018084384500980377,
0.03416107967495918,
-0.03513070568442345,
0.11218637973070145,
0.03915521129965782,
0.07947167754173279,
-0.10209160298109055,
-0.008342097513377666,
0.08467084169387817,
-0.008838878944516182,
-0.011974303983151913,
0.005094799678772688,
-0.02414740063250065,
-0.08544345945119858,
0.07070008665323257,
-0.012595883570611477,
0.05225250869989395,
-0.01644578017294407,
0.006558787077665329,
-0.001007073326036334,
0.004272721242159605,
-0.09905081242322922,
-0.08263251930475235,
-0.017112672328948975,
-0.07372656464576721,
0.033720433712005615,
0.026711292564868927,
-0.0521387979388237,
0.050807323306798935,
-0.02491813711822033,
-0.08349686861038208,
-0.0885830670595169,
-0.049880217760801315,
0.030698800459504128,
0.030540388077497482,
-5.6524302483087224e-34,
0.033739250153303146,
0.002461428986862302,
0.04029412567615509,
0.016246946528553963,
0.030990585684776306,
0.024725163355469704,
-0.015985937789082527,
-0.009377527981996536,
0.03571144491434097,
-0.1059299185872078,
-0.09478002041578293,
0.05305176228284836,
0.0281313955783844,
-0.017498139292001724,
0.012223806232213974,
0.01070948876440525,
-0.07606783509254456,
-0.01925278827548027,
-0.004360483959317207,
-0.0002073160867439583,
-0.05101456493139267,
0.07236211001873016,
-0.09792855381965637,
-0.0016313142841681838,
-0.04953926429152489,
0.023028956726193428,
-0.07322759181261063,
0.08817292749881744,
0.08326583355665207,
0.018495934084057808,
-0.017906859517097473,
-0.09106975793838501,
-0.014679224230349064,
-0.016465865075588226,
-0.057673629373311996,
0.10127609968185425,
0.1130339503288269,
0.04244792088866234,
-0.07751355320215225,
0.09643543511629105,
0.03725360333919525,
-0.06454058736562729,
-0.06944011151790619,
0.02638663351535797,
0.06870225071907043,
-0.02729238197207451,
-0.026193343102931976,
-0.0023044152185320854,
-0.0007428186945617199,
0.07064801454544067,
-0.022365061566233635,
0.05982191860675812,
-0.06478473544120789,
0.009004436433315277,
0.010094997473061085,
-0.08525941520929337,
-0.015573211945593357,
-0.024477148428559303,
0.07199309766292572,
-0.040906213223934174,
-0.05985120311379433,
-0.03205936774611473,
0.013879832811653614,
0.004161044955253601,
-0.07396742701530457,
-0.0016953970771282911,
0.009377295151352882,
0.04749548062682152,
0.03163675218820572,
0.03241175413131714,
0.011807197704911232,
0.03730539605021477,
0.06609192490577698,
-0.05662926286458969,
0.0771031528711319,
-0.10112658143043518,
-0.03303369879722595,
-0.043486785143613815,
-0.05991438403725624,
-0.04010117053985596,
0.019628001376986504,
0.015358710661530495,
-0.014482080936431885,
0.0371866375207901,
0.08311735838651657,
0.07517801970243454,
0.011167490854859352,
0.004106673412024975,
0.0007654455257579684,
-0.029316537082195282,
-0.007183790672570467,
-0.02680226042866707,
0.02479923889040947,
0.01431701984256506,
-0.05898481607437134,
-4.4685453559623056e-8,
-0.04449234902858734,
0.07013587653636932,
0.08372737467288971,
0.044874999672174454,
0.09722519665956497,
-0.02031378448009491,
-0.021943392232060432,
0.02547946758568287,
-0.026609499007463455,
-0.013020742684602737,
0.10177217423915863,
0.0045615993440151215,
0.035923074930906296,
-0.10950937867164612,
0.08509702980518341,
0.015392531640827656,
0.07256415486335754,
0.00011689898383338004,
-0.04393152892589569,
-0.008038751780986786,
0.027496326714754105,
-0.01062746811658144,
-0.007432005368173122,
-0.054863907396793365,
0.053499676287174225,
-0.060583774000406265,
0.005676377099007368,
0.016611073166131973,
0.07044952362775803,
-0.008532625623047352,
0.04643728956580162,
0.02007061056792736,
0.06681261956691742,
0.04606349393725395,
0.01698905974626541,
0.02639325149357319,
0.03693443909287453,
-0.04882476478815079,
-0.041317135095596313,
-0.04317723587155342,
-0.1284630447626114,
0.01428068708628416,
-0.037845831364393234,
-0.0026446382980793715,
-0.017729295417666435,
-0.05019727349281311,
-0.03646430745720863,
-0.07239217311143875,
0.05755702033638954,
-0.011474693194031715,
0.02569158934056759,
-0.031911641359329224,
-0.033674534410238266,
0.011771188117563725,
0.08894208818674088,
0.01729818992316723,
-0.02746088057756424,
-0.04237201809883118,
-0.005669022910296917,
0.05035154148936272,
0.059265997260808945,
-0.034475184977054596,
-0.12513911724090576,
-0.0203645471483469
] |
MLPs Learn In-Context on Regression and Classification Tasks | https://openreview.net/forum?id=MbX0t1rUlp | [
"William Lingxiao Tong",
"Cengiz Pehlevan"
] | Poster | In-context learning (ICL), the remarkable ability to solve a task from only input exemplars, is often assumed to be a unique hallmark of Transformer models. By examining commonly employed synthetic ICL tasks, we demonstrate that multi-layer perceptrons (MLPs) can also learn in-context. Moreover, MLPs, and the closely related MLP-Mixer models, learn in-context comparably with Transformers under the same compute budget in this setting. We further show that MLPs outperform Transformers on a series of classical tasks from psychology designed to test relational reasoning, which are closely related to in-context classification. These results underscore a need for studying in-context learning beyond attention-based architectures, while also challenging prior arguments against MLPs' ability to solve relational tasks. Altogether, our results highlight the unexpected competence of MLPs in a synthetic setting, and support the growing interest in all-MLP alternatives to Transformer architectures. It remains unclear how MLPs perform against Transformers at scale on real-world tasks, and where a performance gap may originate. We encourage further exploration of these architectures in more complex settings to better understand the potential comparative advantage of attention-based schemes. | In-context learning, relational reasoning, synthetic tasks, MLP, MLP-Mixer, Transformer | On a range of widely studied synthetic in-context learning tasks, we find that MLPs perform comparably with Transformers under the same compute budget. | 1,336 | 2405.15618 | [
-0.08079597353935242,
-0.07424823939800262,
0.010252426378428936,
0.013843867927789688,
0.07353422045707703,
-0.051248013973236084,
0.017760463058948517,
0.027827560901641846,
0.039255715906620026,
-0.01911298744380474,
-0.0800756886601448,
-0.057474590837955475,
0.009341861121356487,
0.02511526644229889,
0.028595533221960068,
0.02779259718954563,
0.07995150238275528,
0.04556912183761597,
-0.0960087701678276,
-0.025370130315423012,
-0.024656306952238083,
-0.03907044604420662,
0.0026414019521325827,
0.010903634130954742,
0.026987427845597267,
0.04562786966562271,
0.022978685796260834,
-0.024198096245527267,
0.008147154934704304,
0.030216066166758537,
0.049817752093076706,
0.011914720758795738,
-0.03583487868309021,
0.05743363872170448,
-0.03226699307560921,
0.10561634600162506,
-0.01693362556397915,
0.0379115529358387,
0.003164799651131034,
-0.10382372885942459,
0.004991042427718639,
-0.07444863021373749,
0.011740218847990036,
0.009882476180791855,
0.08209379017353058,
0.015578123740851879,
0.04382140561938286,
-0.019646234810352325,
-0.01613575965166092,
-0.06430879235267639,
-0.09716380387544632,
-0.02054002694785595,
-0.027714502066373825,
0.034085098654031754,
-0.05774499848484993,
0.0897473469376564,
0.08423500508069992,
0.07548883557319641,
-0.026332836598157883,
0.03314043954014778,
-0.09524501115083694,
-0.09531795233488083,
-0.03214850276708603,
0.006703964900225401,
-0.09171540290117264,
0.04604244977235794,
0.02098163589835167,
0.029885495081543922,
-0.026331834495067596,
-0.044548436999320984,
-0.0048253703862428665,
0.0730401873588562,
-0.029560737311840057,
0.05975040793418884,
0.09098855406045914,
0.011715964414179325,
0.07390101253986359,
-0.07064400613307953,
0.07758210599422455,
-0.043023135513067245,
0.021761294454336166,
0.03809036687016487,
-0.03932155296206474,
-0.0010407768422737718,
0.12728607654571533,
0.005706800147891045,
-0.0029495488852262497,
0.01695932075381279,
-0.03684357553720474,
0.015872372314333916,
-0.0565032958984375,
-0.14195406436920166,
-0.010324658825993538,
0.010533399879932404,
0.06948207318782806,
0.04393168166279793,
0.017618421465158463,
-0.03219446539878845,
-0.031729258596897125,
0.08340134471654892,
-0.014340400695800781,
0.037008460611104965,
-0.015196751803159714,
-0.009314881637692451,
-0.04925974830985069,
-0.025017576292157173,
-0.033435747027397156,
0.0434798039495945,
0.01633182354271412,
-0.12696915864944458,
0.012751798145473003,
0.014716051518917084,
0.03210219368338585,
-0.044747624546289444,
0.019222941249608994,
-0.05067962035536766,
0.035894472151994705,
-0.06712790578603745,
0.03678636625409126,
0.07624655216932297,
-0.07406562566757202,
0.04358969256281853,
-0.07905508577823639,
0.10325965285301208,
-0.021744826808571815,
-0.0652465894818306,
-0.07897757738828659,
1.6023224107486065e-33,
-0.005768024828284979,
-0.0056582908146083355,
0.09833879768848419,
0.037985920906066895,
0.01665685512125492,
0.024483365938067436,
0.014713393524289131,
0.05454573407769203,
0.05843360722064972,
0.08000260591506958,
0.033187419176101685,
0.07394074648618698,
-0.04102504998445511,
0.05107150599360466,
0.030840741470456123,
0.02427227422595024,
-0.04944233223795891,
0.03727482259273529,
-0.06417340785264969,
-0.06352857500314713,
-0.011271199211478233,
-0.034553784877061844,
-0.031032675877213478,
-0.0235077403485775,
-0.014702820219099522,
0.07764923572540283,
0.020285172387957573,
-0.023681070655584335,
0.0407986119389534,
-0.008536376990377903,
-0.08623671531677246,
0.06812981516122818,
-0.07310903817415237,
0.09365762770175934,
0.003692538943141699,
-0.019051002338528633,
-0.007826294749975204,
-0.07948935031890869,
0.014282076619565487,
-0.0026467598509043455,
-0.051692914217710495,
0.06427210569381714,
-0.020949585363268852,
0.04597342014312744,
-0.07381156086921692,
-0.01911771111190319,
0.017203962430357933,
-0.043862711638212204,
-0.03935433551669121,
-0.054258543998003006,
0.019324762746691704,
0.04326843097805977,
-0.027197064831852913,
-0.06380822509527206,
0.03365574777126312,
-0.004116762895137072,
0.06068813055753708,
0.020494813099503517,
0.06532768160104752,
0.06988140195608139,
0.00706395972520113,
0.01722847670316696,
-0.04723265394568443,
0.08509175479412079,
0.05643730238080025,
0.025349942967295647,
-0.01946026273071766,
0.08818834275007248,
0.05500820651650429,
-0.028152991086244583,
-0.025191662833094597,
0.06595870107412338,
-0.01751677133142948,
-0.05920233577489853,
0.10125923156738281,
0.0037850323133170605,
-0.018611980602145195,
-0.052389077842235565,
-0.0073812417685985565,
0.024678945541381836,
0.06574984639883041,
0.03185977414250374,
-0.02912655845284462,
0.02452027052640915,
-0.03048998862504959,
0.004277914296835661,
-0.04054231569170952,
-0.09578181058168411,
-0.011385191231966019,
-0.037676166743040085,
-0.08637554198503494,
-0.01998799853026867,
-0.05594492331147194,
0.0005835764459334314,
-0.06320485472679138,
-2.649825263233517e-33,
-0.0036550594959408045,
0.013680953532457352,
-0.0771404355764389,
0.01096493098884821,
-0.0008410098380409181,
-0.03748314827680588,
0.017356514930725098,
-0.015023326501250267,
-0.04406284913420677,
0.022509675472974777,
-0.00755365239456296,
-0.08289793878793716,
0.05688360333442688,
-0.00965972151607275,
0.021976958960294724,
-0.058831922709941864,
-0.06736686080694199,
-0.02229425497353077,
-0.02356424368917942,
0.05815192684531212,
0.03718118369579315,
0.12398646771907806,
-0.05998766049742699,
0.008632931858301163,
-0.12946096062660217,
-0.009511329233646393,
-0.10672450810670853,
0.06972574442625046,
0.03482610359787941,
-0.014050337485969067,
-0.03543854504823685,
-0.0010125564876943827,
0.030664782971143723,
-0.06305494904518127,
-0.02915450744330883,
0.0410393625497818,
-0.0026629529893398285,
-0.023963846266269684,
-0.0391092374920845,
0.016976721584796906,
0.055775780230760574,
-0.03278645500540733,
-0.04833583906292915,
0.02593587525188923,
-0.037270646542310715,
0.002720870776101947,
-0.02520996332168579,
-0.040673255920410156,
0.013618563301861286,
0.013217071071267128,
-0.0530020073056221,
0.022268831729888916,
-0.07538057118654251,
-0.060892101377248764,
-0.03792939335107803,
-0.045977964997291565,
-0.024764612317085266,
-0.027068687602877617,
0.05837854743003845,
-0.024343539029359818,
-0.052240461111068726,
-0.02046201378107071,
0.07002624869346619,
-0.025497259572148323,
-0.018813857808709145,
0.02093241922557354,
0.01668490655720234,
-0.00010049720003735274,
0.05404110997915268,
0.019132811576128006,
-0.013944094069302082,
0.002844127593562007,
0.0057565681636333466,
-0.02853091061115265,
-0.007048348430544138,
-0.027515530586242676,
-0.026140499860048294,
-0.09030336141586304,
0.020782699808478355,
-0.038099505007267,
-0.04550953581929207,
-0.0948154553771019,
0.04170987382531166,
0.062196169048547745,
0.04361917823553085,
0.03394529968500137,
0.06421338766813278,
0.0673745721578598,
0.05447966977953911,
-0.03376924619078636,
-0.023712534457445145,
0.090488001704216,
-0.03468456119298935,
0.074278324842453,
-0.02755868434906006,
-4.598721758952706e-8,
-0.08375971019268036,
0.10119248181581497,
0.0010651189368218184,
0.012536201626062393,
0.031570881605148315,
-0.00197288254275918,
-0.011844417080283165,
0.11477191746234894,
-0.03863992542028427,
0.09232895076274872,
0.03778015449643135,
-0.05552787333726883,
0.006506238132715225,
-0.06088116765022278,
0.06952334195375443,
0.061251312494277954,
0.11747925728559494,
-0.04511752352118492,
0.014175120741128922,
-0.01726892776787281,
0.07186237722635269,
0.043150775134563446,
-0.0013561372179538012,
-0.028270475566387177,
0.05375922471284866,
-0.09573613107204437,
-0.05937331169843674,
0.0972575768828392,
0.05659006908535957,
0.04888315126299858,
-0.023706315085291862,
0.007793603930622339,
-0.003638161811977625,
-0.07930872589349747,
0.05453047156333923,
0.0678909569978714,
0.07618161290884018,
-0.048473015427589417,
0.018509840592741966,
-0.07012847810983658,
-0.024601249024271965,
0.023524085059762,
-0.07199195772409439,
0.07346834242343903,
0.01171396765857935,
0.06369395554065704,
-0.0013842687476426363,
-0.12221312522888184,
0.003189150942489505,
0.030111068859696388,
0.048358745872974396,
-0.016138942912220955,
-0.02988692745566368,
0.04110018536448479,
0.07022430002689362,
0.0006234548054635525,
-0.01809133030474186,
-0.064654640853405,
-0.03710886090993881,
0.09213629364967346,
-0.011141804046928883,
0.09001712501049042,
-0.06932736933231354,
0.022466089576482773
] |
Restyling Unsupervised Concept Based Interpretable Networks with Generative Models | https://openreview.net/forum?id=CexatBp6rx | [
"Jayneel Parekh",
"Quentin Bouniot",
"Pavlo Mozharovskyi",
"Alasdair Newson",
"Florence d'Alché-Buc"
] | Poster | Developing inherently interpretable models for prediction has gained prominence in recent years. A subclass of these models, wherein the interpretable network relies on learning high-level concepts, are valued because of closeness of concept representations to human communication. However, the visualization and understanding of the learnt unsupervised dictionary of concepts encounters major limitations, especially for large-scale images. We propose here a novel method that relies on mapping the concept features to the latent space of a pretrained generative model. The use of a generative model enables high quality visualization, and lays out an intuitive and interactive procedure for better interpretation of the learnt concepts by imputing concept activations and visualizing generated modifications. Furthermore, leveraging pretrained generative models has the additional advantage of making the training of the system more efficient. We quantitatively ascertain the efficacy of our method in terms of accuracy of the interpretable prediction network, fidelity of reconstruction, as well as faithfulness and consistency of learnt concepts. The experiments are conducted on multiple image recognition benchmarks for large-scale images. Project page available at https://jayneelparekh.github.io/VisCoIN_project_page/ | explainability, generative models, concepts | null | 1,335 | 2407.01331 | [
0.030343303456902504,
-0.12959232926368713,
0.01920459233224392,
0.05927080661058426,
0.05983696132898331,
-0.009566640481352806,
-0.06410665810108185,
-0.03490610420703888,
0.011104720644652843,
-0.022425679489970207,
-0.054085623472929,
-0.03192727267742157,
0.06156151741743088,
0.047542303800582886,
-0.07000395655632019,
-0.022683607414364815,
0.13609082996845245,
0.1527450829744339,
-0.09045346081256866,
-0.08673389256000519,
0.036891233175992966,
-0.002843171590939164,
0.06142902374267578,
-0.01868927851319313,
0.10161479562520981,
-0.0006935842102393508,
0.0212373286485672,
0.04364265128970146,
0.00657507823780179,
-0.08157436549663544,
0.0021867183968424797,
-0.014142841100692749,
-0.06772462278604507,
0.07152563333511353,
-0.08548738062381744,
0.13286066055297852,
-0.02378729172050953,
0.07963638007640839,
0.004177505616098642,
-0.01774681732058525,
0.030810631811618805,
0.02720915898680687,
-0.013021694496273994,
0.010419320315122604,
0.08348409086465836,
-0.028942586854100227,
-0.025652237236499786,
-0.08942122757434845,
0.00668510003015399,
-0.014128012582659721,
-0.05529243126511574,
0.004539409186691046,
-0.08854624629020691,
0.03054325468838215,
0.01973694935441017,
0.019763456657528877,
0.06158659607172012,
-0.013025084510445595,
0.017808396369218826,
-0.021958284080028534,
-0.03871772438287735,
-0.06437967717647552,
0.036063842475414276,
-0.014775950461626053,
0.0002493461361154914,
-0.06917276233434677,
0.052740100771188736,
0.021774601191282272,
0.05250886455178261,
-0.044981859624385834,
0.03315368667244911,
0.0492863766849041,
-0.05413009598851204,
-0.002349067712202668,
0.029007116332650185,
0.06495881825685501,
0.04836335405707359,
0.016969239339232445,
0.036256104707717896,
-0.07145325094461441,
0.022212637588381767,
0.03753204271197319,
0.03994816914200783,
0.005788334645330906,
0.07466533780097961,
0.08272027224302292,
0.006570839323103428,
0.06916990876197815,
-0.023294245824217796,
0.018282553181052208,
-0.06673283129930496,
-0.03498952090740204,
0.00977686420083046,
-0.0075572794303298,
0.05656454339623451,
-0.001362431445159018,
0.04871808737516403,
-0.08843789249658585,
-0.06962709873914719,
0.0399450957775116,
0.0006268948200158775,
0.07509108632802963,
0.03247230872511864,
-0.06401832401752472,
0.0671302080154419,
-0.030083943158388138,
0.0959235429763794,
0.02791478857398033,
0.028488028794527054,
-0.04468998685479164,
0.01678021438419819,
0.0401977077126503,
-0.07714302837848663,
0.0037165582180023193,
0.010705623775720596,
-0.01558788400143385,
0.048092760145664215,
-0.03535788506269455,
-0.026309436187148094,
0.04340565577149391,
-0.010504942387342453,
0.03800464794039726,
0.02893296629190445,
-0.006236405577510595,
0.03006879985332489,
-0.09456438571214676,
-0.08405978977680206,
1.9988919603878137e-33,
0.03387576341629028,
0.003686324693262577,
0.08322683721780777,
0.07699950039386749,
0.10123562812805176,
-0.010536575689911842,
-0.07196246832609177,
0.042149122804403305,
-0.027998032048344612,
-0.003414715873077512,
-0.04986502230167389,
0.07999899983406067,
-0.09141195565462112,
0.14465376734733582,
0.039196718484163284,
-0.04161089286208153,
-0.0667494386434555,
-0.0033810327295213938,
-0.0005595526308752596,
-0.029584141448140144,
-0.04410413280129433,
-0.047575123608112335,
-0.020041298121213913,
0.019397491589188576,
0.001165333786047995,
0.008817004971206188,
0.07572147995233536,
-0.032755594700574875,
-0.0057040005922317505,
-0.019527176395058632,
-0.13178148865699768,
0.02332424744963646,
0.10133016854524612,
-0.023059507831931114,
-0.0012749196030199528,
-0.09927086532115936,
0.012494536116719246,
-0.04659592732787132,
0.045664384961128235,
-0.02923477441072464,
-0.008402297273278236,
-0.019771091639995575,
-0.043085042387247086,
-0.039083901792764664,
-0.0763636901974678,
-0.01718117482960224,
0.030062615871429443,
0.05715202912688255,
-0.04911470040678978,
-0.00415528379380703,
0.03265028074383736,
0.024836016818881035,
-0.01870950497686863,
-0.042381513863801956,
-0.08385705202817917,
0.0005859149969182909,
0.04385931044816971,
0.003147423267364502,
0.02399411052465439,
-0.038352396339178085,
0.002290277974680066,
0.06736776977777481,
0.03753882646560669,
0.12456262111663818,
0.019079895690083504,
-0.001909982063807547,
0.0013861092738807201,
0.06588230282068253,
-0.006017976440489292,
0.042394641786813736,
-0.08963403105735779,
0.05409306660294533,
-0.047882210463285446,
-0.07650714367628098,
0.04402931407094002,
0.01774686947464943,
-0.023759013041853905,
-0.01992991380393505,
0.01552570890635252,
0.04314082860946655,
-0.01076321117579937,
-0.055196259170770645,
-0.010725347325205803,
-0.029392531141638756,
-0.017758818343281746,
0.013004308566451073,
0.052291736006736755,
-0.08362450450658798,
-0.005435760132968426,
0.03191316872835159,
-0.04638676345348358,
-0.026339231058955193,
0.010504930280148983,
-0.015590322203934193,
0.0056083472445607185,
-1.6076095639334144e-33,
-0.013536745682358742,
0.04573552682995796,
-0.009557110257446766,
0.047340027987957,
0.003651374252513051,
-0.047405343502759933,
-0.038976021111011505,
0.11131132394075394,
-0.06961335986852646,
-0.05813740938901901,
0.03108539991080761,
-0.03251155838370323,
-0.02578573301434517,
-0.05206822231411934,
-0.025397609919309616,
-0.03320479393005371,
0.004358747508376837,
0.023767946287989616,
-0.03464287519454956,
0.0074796793051064014,
0.022612392902374268,
0.06289596855640411,
-0.08395421504974365,
0.024663161486387253,
-0.0527862086892128,
0.11311481893062592,
-0.030651845037937164,
0.025498459115624428,
-0.040938183665275574,
-0.01761690154671669,
0.024720314890146255,
-0.02642783150076866,
0.029521232470870018,
-0.003307155566290021,
-0.074233777821064,
0.06931491941213608,
0.09475705772638321,
-0.04081830754876137,
-0.05973169207572937,
0.06627032905817032,
0.06265123188495636,
-0.0610954612493515,
-0.10607940703630447,
-0.04673328250646591,
-0.03769772872328758,
-0.09091716259717941,
-0.011910508386790752,
-0.001522057456895709,
0.07622271031141281,
0.12501981854438782,
-0.017627758905291557,
0.016059812158346176,
-0.10130694508552551,
0.0016466672532260418,
-0.01288660243153572,
-0.08243036270141602,
0.016411473974585533,
-0.03290950506925583,
0.15216703712940216,
-0.029803162440657616,
-0.045425329357385635,
-0.08445363491773605,
-0.005690010264515877,
-0.030153870582580566,
-0.05887977033853531,
-0.0572909414768219,
-0.08151717483997345,
-0.009739757515490055,
0.03270789235830307,
0.04098162800073624,
-0.05204197019338608,
-0.028431417420506477,
0.011425472795963287,
-0.003969893325120211,
0.05656418576836586,
-0.049783848226070404,
0.030567370355129242,
0.04878648370504379,
0.020693369209766388,
-0.08850856870412827,
0.043708935379981995,
-0.049091510474681854,
-0.014870496466755867,
0.051508430391550064,
0.11428533494472504,
0.055617351084947586,
-0.01754601113498211,
0.015353532508015633,
0.10584010928869247,
0.014443306252360344,
-0.020608527585864067,
0.01762598380446434,
-0.08841075003147125,
0.048784613609313965,
-0.06614138931035995,
-4.185769952869123e-8,
-0.07796579599380493,
0.03590833395719528,
0.04304034635424614,
-0.0027318147476762533,
0.010731136426329613,
-0.0701945498585701,
-0.006481718271970749,
0.11046556383371353,
-0.056108780205249786,
-0.029052944853901863,
0.09396984428167343,
-0.03415104001760483,
-0.013506811112165451,
-0.02487807162106037,
0.04054531455039978,
0.042011141777038574,
0.040958963334560394,
0.052705321460962296,
0.045214127749204636,
-0.0066376738250255585,
0.00011106729652965441,
-0.050585098564624786,
-0.003338766749948263,
0.022221000865101814,
0.0513148233294487,
-0.06362742185592651,
-0.07042881846427917,
0.09678543359041214,
-0.03239172324538231,
-0.012806897051632404,
0.027047866955399513,
-0.003476619254797697,
0.04606335610151291,
-0.017806295305490494,
-0.011418187990784645,
0.057970404624938965,
0.0056532686576247215,
-0.0004703619924839586,
-0.007036149967461824,
-0.03765350952744484,
0.02957374043762684,
0.023553594946861267,
0.005959074012935162,
0.05271584913134575,
0.002706450643017888,
0.008526457473635674,
0.01953929476439953,
0.021675631403923035,
-0.014340189285576344,
0.015685290098190308,
-0.05607407167553902,
0.006706824526190758,
0.03258072957396507,
0.04786546900868416,
0.03553332760930061,
-0.02877911366522312,
0.03277851641178131,
-0.007005743682384491,
0.060755934566259384,
0.05743759125471115,
0.0056449901312589645,
0.016916178166866302,
-0.05489930883049965,
-0.020482877269387245
] |
Active Learning for Neural PDE Solvers | https://openreview.net/forum?id=x4ZmQaumRg | [
"Daniel Musekamp",
"Marimuthu Kalimuthu",
"David Holzmüller",
"Makoto Takamoto",
"Mathias Niepert"
] | Poster | Solving partial differential equations (PDEs) is a fundamental problem in engineering and science. While neural PDE solvers can be more efficient than established numerical solvers, they often require large amounts of training data that is costly to obtain. Active learning (AL) could help surrogate models reach the same accuracy with smaller training sets by querying classical solvers with more informative initial conditions and PDE parameters. While AL is more common in other domains, it has yet to be studied extensively for neural PDE solvers. To bridge this gap, we introduce AL4PDE, a modular and extensible active learning benchmark. It provides multiple parametric PDEs and state-of-the-art surrogate models for the solver-in-the-loop setting, enabling the evaluation of existing and the development of new AL methods for PDE solving. We use the benchmark to evaluate batch active learning algorithms such as uncertainty- and feature-based methods. We show that AL reduces the average error by up to 71\% compared to random sampling and significantly reduces worst-case errors. Moreover, AL generates similar datasets across repeated runs, with consistent distributions over the PDE parameters and initial conditions. The acquired datasets are reusable, providing benefits for surrogate models not involved in the data generation. | Active Learning, Neural PDE Solvers, Scientific Machine Learning, Benchmark, Framework, Neural Operators | A extensible benchmark to evaluate pool-based active learning for neural PDE solvers. | 1,332 | 2408.01536 | [
-0.07908768206834793,
-0.07851046323776245,
0.03785790130496025,
0.09303098917007446,
0.023775631561875343,
-0.02791653200984001,
-0.028649574145674706,
0.01579545997083187,
-0.044028837233781815,
-0.02192586101591587,
0.008942068554461002,
-0.01599852740764618,
-0.05903884395956993,
-0.016332432627677917,
-0.040553782135248184,
0.06356588006019592,
-0.02235870622098446,
0.053809333592653275,
-0.04797299578785896,
-0.004452482331544161,
-0.009269182570278645,
0.0289328470826149,
-0.0261822696775198,
-0.016727613285183907,
0.004826684482395649,
-0.006983959581702948,
0.08048496395349503,
-0.016648676246404648,
-0.05333952605724335,
0.03131205961108208,
0.08912334591150284,
-0.0467049665749073,
-0.048592329025268555,
-0.06440294533967972,
-0.014312871731817722,
0.03519972413778305,
-0.0984487533569336,
-0.0011411692248657346,
0.010764769278466702,
-0.1088409498333931,
0.06591972708702087,
-0.014715270139276981,
0.020299270749092102,
0.011093189008533955,
0.11638068407773972,
-0.027698783203959465,
-0.022287767380475998,
-0.035770732909440994,
0.05606013536453247,
0.025611069053411484,
-0.03864171355962753,
-0.02564030885696411,
-0.04350907355546951,
0.006724564358592033,
0.025408409535884857,
-0.0016144698020070791,
0.05981172248721123,
-0.01704254560172558,
0.044086482375860214,
0.05732820928096771,
-0.03692062199115753,
-0.06450038403272629,
-0.04082685336470604,
-0.045112039893865585,
-0.028439730405807495,
-0.009571794420480728,
0.08957593888044357,
0.03662211447954178,
0.023453567177057266,
0.07039334625005722,
-0.055464036762714386,
0.09306955337524414,
-0.07026660442352295,
0.0243989210575819,
0.018648112192749977,
0.0005581993027590215,
0.10307478904724121,
-0.044075239449739456,
0.00457343365997076,
-0.029029112309217453,
-0.046100128442049026,
0.043644532561302185,
0.0002736444876063615,
-0.08238822966814041,
0.11653801053762436,
-0.010888759046792984,
-0.017269551753997803,
0.03052823431789875,
0.01104614045470953,
-0.018335485830903053,
0.00210614874958992,
-0.00015603062638547271,
0.02157423086464405,
-0.06801562756299973,
-0.014928474090993404,
-0.0011259306920692325,
0.031633410602808,
-0.012552669271826744,
-0.028561539947986603,
0.08103480190038681,
-0.017260652035474777,
0.06749750673770905,
0.009009155444800854,
0.001353273168206215,
0.00455885473638773,
-0.02211286686360836,
0.10768502950668335,
0.04411178082227707,
-0.010736772790551186,
-0.05421094596385956,
0.0404316745698452,
-0.0006596151506528258,
0.04085050895810127,
-0.011383487842977047,
0.059465643018484116,
-0.0829092338681221,
0.029251905158162117,
-0.13712212443351746,
0.00002720230804698076,
0.025021888315677643,
-0.06135432794690132,
-0.01130416989326477,
-0.019591787829995155,
0.039576128125190735,
0.05160268023610115,
-0.07459485530853271,
-0.026790818199515343,
2.2653151836077927e-33,
0.03047671541571617,
-0.0384216271340847,
0.005102882161736488,
0.045902352780103683,
-0.0063834660686552525,
-0.021233126521110535,
0.09323269873857498,
0.08137787133455276,
0.006470301188528538,
0.08642666786909103,
-0.0730513408780098,
0.008050668984651566,
-0.05149668827652931,
0.0797826424241066,
-0.020383773371577263,
-0.08846081048250198,
-0.0534711591899395,
0.038731370121240616,
-0.02926723100244999,
-0.0023490111343562603,
0.005902885925024748,
-0.09026655554771423,
0.010199261829257011,
-0.0012954625999554992,
-0.018850646913051605,
-0.0020541949197649956,
-0.013917091302573681,
0.01334407739341259,
-0.035119760781526566,
0.009646479971706867,
-0.08515693992376328,
-0.0336533784866333,
-0.08337054401636124,
0.11607998609542847,
0.019354967400431633,
0.024557122960686684,
0.001363355084322393,
-0.0060343062505126,
-0.028343984857201576,
0.012779348529875278,
-0.04801490157842636,
0.05892990157008171,
0.12305013090372086,
0.11252710223197937,
-0.04579463228583336,
0.005014792084693909,
-0.03203519806265831,
-0.03942578285932541,
0.03316551446914673,
-0.04268287867307663,
-0.06489463150501251,
-0.032073285430669785,
0.018231207504868507,
-0.04604363068938255,
-0.03289935737848282,
-0.008683676831424236,
0.012667532078921795,
0.010713321156799793,
-0.005554674193263054,
0.05365610122680664,
-0.027183538302779198,
0.030329247936606407,
-0.048112113028764725,
0.005591047462075949,
-0.020482558757066727,
0.0921584963798523,
0.044552385807037354,
0.013757586479187012,
-0.005821462720632553,
-0.050234273076057434,
0.04735206440091133,
0.0503835566341877,
0.09197406470775604,
-0.033510446548461914,
0.13331539928913116,
-0.00041342503391206264,
0.10303592681884766,
-0.04322149604558945,
-0.011461697518825531,
-0.037694547325372696,
0.014726727269589901,
-0.0007852127309888601,
-0.004189459607005119,
0.0277072936296463,
-0.0053737573325634,
-0.04342687129974365,
0.012801066972315311,
-0.03510221466422081,
-0.10148978978395462,
-0.061929646879434586,
-0.10601460188627243,
0.02754443697631359,
-0.030868632718920708,
-0.02642371505498886,
0.059270262718200684,
-1.5859450193767673e-33,
0.031968411058187485,
0.05112984776496887,
-0.013148415833711624,
0.029558002948760986,
-0.01387783233076334,
-0.038042787462472916,
-0.0833684429526329,
0.025546731427311897,
-0.00888901948928833,
-0.10642830282449722,
-0.025899576023221016,
-0.0591999888420105,
0.040641479194164276,
-0.050166260451078415,
0.05484955385327339,
0.0010604967828840017,
-0.10683654993772507,
-0.03185981884598732,
0.05085931345820427,
0.048109687864780426,
-0.014548339881002903,
0.07164420187473297,
-0.021315982565283775,
0.019615745171904564,
-0.07855818420648575,
-0.03025592304766178,
-0.015229745768010616,
0.061117466539144516,
0.0049699461087584496,
-0.016327373683452606,
-0.013554089702665806,
-0.030224699527025223,
-0.05084393918514252,
0.0355294905602932,
-0.0765460953116417,
0.16327323019504547,
0.06231890618801117,
-0.006884362548589706,
0.05633467808365822,
0.04827107861638069,
0.0915088877081871,
-0.05834652855992317,
-0.07104352861642838,
0.006347897928208113,
0.01985144056379795,
0.07749518752098083,
0.0013081432553008199,
-0.0012224053498357534,
-0.021139031276106834,
0.04394999518990517,
0.056250765919685364,
0.026379162445664406,
-0.10788872092962265,
-0.023289354518055916,
-0.06230027601122856,
-0.02096397429704666,
-0.06013965234160423,
-0.049630820751190186,
0.03129202872514725,
-0.0038706448394805193,
0.0007018314790911973,
0.03729157894849777,
0.08816337585449219,
0.05471936613321304,
-0.07271125912666321,
0.0181287694722414,
-0.010716911405324936,
0.07113603502511978,
0.003941845148801804,
0.01008254662156105,
-0.007688387297093868,
-0.03106120601296425,
-0.01104540005326271,
0.0020803865045309067,
-0.01981149986386299,
-0.0564345046877861,
-0.016821740195155144,
-0.06428322196006775,
-0.02862849459052086,
0.002452025655657053,
0.061396028846502304,
-0.01594509929418564,
0.045958325266838074,
0.05039685219526291,
-0.009503818117082119,
0.11719722300767899,
0.07414894551038742,
0.004783198237419128,
0.015819501131772995,
-0.003370819380506873,
-0.022424012422561646,
0.0859733298420906,
0.02572225220501423,
0.04106247052550316,
-0.029495202004909515,
-3.597719455683546e-8,
0.011170810088515282,
0.12798699736595154,
0.07725763320922852,
-0.023517489433288574,
0.06561411917209625,
0.007583527360111475,
-0.030777042731642723,
0.030553804710507393,
-0.04535289108753204,
-0.01754131354391575,
0.0727895200252533,
0.00841887854039669,
-0.040185656398534775,
-0.09466049820184708,
-0.07564286887645721,
0.05451035499572754,
-0.050327956676483154,
0.007951431907713413,
-0.0671406164765358,
0.018324026837944984,
0.031204428523778915,
0.08319227397441864,
-0.025910675525665283,
-0.09560263901948929,
0.062446337193250656,
-0.02188626118004322,
-0.031629402190446854,
0.035616252571344376,
0.11471613496541977,
-0.010939974337816238,
0.009823708795011044,
0.0019201450049877167,
0.018424397334456444,
0.015325929969549179,
0.08181578665971756,
0.09907136857509613,
0.06916974484920502,
-0.11383146047592163,
-0.08446717262268066,
-0.054458919912576675,
-0.05327082425355911,
0.04256156086921692,
-0.028831161558628082,
-0.010429766960442066,
-0.004767597187310457,
0.045266762375831604,
-0.0015498402062803507,
-0.08918273448944092,
0.13777004182338715,
-0.021791426464915276,
0.028412330895662308,
0.015466567128896713,
-0.028555454686284065,
0.04422217980027199,
0.04076920077204704,
-0.03221318498253822,
-0.030509261414408684,
-0.003901401534676552,
0.08868682384490967,
0.016823865473270416,
0.00012958627485204488,
0.04807489737868309,
-0.05667366832494736,
0.05404915660619736
] |
Contextual Self-paced Learning for Weakly Supervised Spatio-Temporal Video Grounding | https://openreview.net/forum?id=yHj6EunfVQ | [
"Akash Kumar",
"Zsolt Kira",
"Yogesh S Rawat"
] | Poster | In this work, we focus on Weakly Supervised Spatio-Temporal Video Grounding (WSTVG). It is a multimodal task aimed at localizing specific subjects spatio-temporally based on textual queries without bounding box supervision. Motivated by recent advancements in multi-modal foundation models for grounding tasks, we first explore the potential of state-of-the-art object detection models for WSTVG. Despite their robust zero-shot capabilities, our adaptation reveals significant limitations, including inconsistent temporal predictions, inadequate understanding of complex queries, and challenges in adapting to difficult scenarios. We propose CoSPaL (Contextual Self-Paced Learning), a novel approach which is designed to overcome these limitations. CoSPaL integrates three core components: (1) Tubelet Phrase Grounding (TPG), which introduces spatio-temporal prediction by linking textual queries to tubelets; (2) Contextual Referral Grounding (CRG), which improves comprehension of complex queries by extracting contextual information to refine object identification over time; and (3) Self-Paced Scene Understanding (SPS), a training paradigm that progressively increases task difficulty, enabling the model to adapt to complex scenarios by transitioning from coarse to fine-grained understanding. | spatio-temporal video grounding, weakly supervised learning, multimodal foundation models, dense video tasks | First adaptation of multimodal foundation models for dense video detection task without any labels. | 1,331 | 2501.17053 | [
0.048469532281160355,
-0.09422250837087631,
0.05655846744775772,
0.04338424280285835,
0.14121559262275696,
0.037394873797893524,
0.037781987339258194,
0.006082376930862665,
0.023483015596866608,
-0.09203606843948364,
-0.07961101084947586,
-0.09771198779344559,
-0.09690459072589874,
0.10610149055719376,
-0.017106803134083748,
-0.0883864313364029,
0.09642691910266876,
0.04216287285089493,
-0.07594259828329086,
0.012411016970872879,
0.03947910666465759,
0.014871128834784031,
0.07083209604024887,
0.011161498725414276,
0.06477627903223038,
0.031414758414030075,
-0.014684787020087242,
-0.05790083855390549,
0.06345149129629135,
-0.015347539447247982,
0.036531299352645874,
0.009066413156688213,
-0.026713620871305466,
0.06958997994661331,
-0.06664611399173737,
0.00839937012642622,
0.03368506580591202,
-0.017168469727039337,
-0.027459565550088882,
0.007653729058802128,
0.010867642238736153,
0.030648699030280113,
0.058071400970220566,
0.03400997072458267,
0.043552737683057785,
0.04766831919550896,
0.018361037597060204,
-0.070333331823349,
0.007852103561162949,
-0.008822975680232048,
-0.057830266654491425,
-0.04057569429278374,
0.0006718214135617018,
0.04083609580993652,
-0.0640900656580925,
0.04809783399105072,
0.034153904765844345,
0.015542273409664631,
0.11215610802173615,
-0.007698332890868187,
0.0375099778175354,
-0.10724654048681259,
0.030428938567638397,
0.029208572581410408,
-0.012376750819385052,
-0.036466460675001144,
0.06959988921880722,
-0.056316226720809937,
0.035891834646463394,
0.0010655816877260804,
-0.027503367513418198,
0.04792405292391777,
-0.0286055039614439,
0.026005126535892487,
-0.019731933251023293,
0.044901732355356216,
-0.018845072016119957,
-0.08430039137601852,
0.06491389125585556,
-0.06427141278982162,
0.07758034020662308,
-0.01453693863004446,
0.06733816862106323,
0.005387084558606148,
0.1026756539940834,
0.031759124249219894,
0.03796173632144928,
0.08673255145549774,
-0.04401647672057152,
-0.005201082676649094,
-0.06846895813941956,
-0.02427119016647339,
0.04894168674945831,
-0.02079187147319317,
0.004670578055083752,
0.018428221344947815,
-0.016466889530420303,
-0.05687488988041878,
-0.024514131247997284,
0.04970073699951172,
-0.000123799909488298,
0.018561404198408127,
-0.03455319628119469,
0.03273741528391838,
0.0175409447401762,
0.04090738296508789,
-0.0022913403809070587,
0.009247549809515476,
0.025888735428452492,
0.019245371222496033,
0.0006966384244151413,
0.010687453672289848,
0.052977919578552246,
-0.04303346946835518,
-0.003508165245875716,
0.03505095839500427,
0.015429560095071793,
0.024553915485739708,
0.023254187777638435,
0.009210171177983284,
0.08016997575759888,
-0.05812012031674385,
-0.07015345990657806,
-0.06999613344669342,
0.04444253072142601,
-0.10322549194097519,
-0.006204663310199976,
1.4137004749992506e-33,
0.0996062159538269,
0.012614979408681393,
0.10374122112989426,
-0.0012317593209445477,
-0.0034557459875941277,
-0.0010488141560927033,
-0.05288474261760712,
0.06230909377336502,
-0.00006452744855778292,
0.0474371612071991,
0.009848176501691341,
-0.008205465972423553,
-0.07404626160860062,
0.02800663746893406,
0.004992910195142031,
-0.015215719118714333,
-0.08936671167612076,
-0.057967059314250946,
0.01467573270201683,
0.04769333451986313,
-0.00027836248045787215,
-0.019808296114206314,
-0.07032297551631927,
0.0022418363951146603,
0.05507184937596321,
0.009691673330962658,
0.09616713225841522,
-0.03617307171225548,
0.02797495573759079,
-0.013654694892466068,
-0.11894401907920837,
-0.01666807010769844,
-0.08521272987127304,
0.030548835173249245,
0.04634607955813408,
-0.10940377414226532,
-0.08888853341341019,
-0.02650403417646885,
0.017896296456456184,
-0.016735613346099854,
-0.016857866197824478,
-0.04763615503907204,
-0.027279658243060112,
-0.020034700632095337,
-0.06436526030302048,
-0.07948394864797592,
0.021797096356749535,
0.04433576390147209,
-0.10414134711027145,
0.03134587034583092,
0.020778270438313484,
0.06508873403072357,
-0.0075437831692397594,
-0.09429898113012314,
0.005317033734172583,
0.031032251194119453,
0.0528494156897068,
0.0686570405960083,
0.04119313880801201,
0.0029777262825518847,
0.051721472293138504,
0.023804128170013428,
-0.007896500639617443,
0.0758732482790947,
-0.01135764829814434,
0.04887521639466286,
-0.02077953703701496,
0.08251501619815826,
0.07769886404275894,
-0.0006546580116264522,
-0.037407081574201584,
0.0831669345498085,
0.008955774828791618,
-0.046206604689359665,
0.06716280430555344,
0.002227949444204569,
0.07738664746284485,
-0.06953232735395432,
0.003684708848595619,
0.05842096731066704,
0.023615248501300812,
-0.09406459331512451,
0.04499954357743263,
-0.024763163179159164,
-0.01674029789865017,
-0.02366647683084011,
0.03184927627444267,
-0.08549334108829498,
-0.007562158163636923,
-0.027617497369647026,
-0.01495095994323492,
-0.030358420684933662,
-0.07120366394519806,
0.022132253274321556,
-0.009462644346058369,
7.644188794696968e-34,
0.06430966407060623,
-0.0012477404670789838,
-0.11058112233877182,
0.004776638932526112,
0.0067960903979837894,
-0.02835976891219616,
0.017754752188920975,
-0.030822614207863808,
0.059857845306396484,
-0.06678984314203262,
-0.05027354881167412,
-0.04248601198196411,
0.02401544898748398,
0.025638142600655556,
-0.006947777699679136,
0.034899648278951645,
-0.03427119180560112,
-0.009769434109330177,
-0.10029256343841553,
0.027779890224337578,
0.042050983756780624,
0.030141664668917656,
-0.0748959630727768,
-0.023294290527701378,
-0.07317423075437546,
0.02601424977183342,
-0.008879157714545727,
0.0437246635556221,
-0.04043332487344742,
0.01968839019536972,
0.025303207337856293,
-0.044700171798467636,
0.0556853711605072,
-0.13367675244808197,
-0.12608258426189423,
0.038019489496946335,
0.03860284015536308,
-0.07777763158082962,
-0.040679190307855606,
0.014608382247388363,
0.09823355823755264,
0.00279632187448442,
-0.007632037159055471,
0.007069265935570002,
-0.024405907839536667,
-0.0015083131147548556,
-0.09387315064668655,
0.029426882043480873,
-0.016368720680475235,
0.08293846249580383,
-0.1026591956615448,
0.06238727644085884,
-0.13435131311416626,
0.011549805290997028,
0.00260811485350132,
-0.04952922835946083,
-0.039519086480140686,
-0.029804091900587082,
0.044331952929496765,
-0.0085041094571352,
-0.047739848494529724,
-0.001186202629469335,
-0.023370342329144478,
-0.010461158119142056,
0.06006745994091034,
-0.025898192077875137,
-0.037949543446302414,
-0.08475048094987869,
0.03795279189944267,
0.04973791912198067,
0.014135347679257393,
-0.003721625078469515,
0.017104104161262512,
-0.016796836629509926,
0.029094524681568146,
0.05217312276363373,
0.029510753229260445,
0.003024730132892728,
0.03472436964511871,
-0.04634793847799301,
0.04124772176146507,
-0.07239485532045364,
0.01749146543443203,
0.08913960307836533,
0.07317672669887543,
0.13271240890026093,
0.006971981376409531,
0.07206599414348602,
0.05257225036621094,
0.04112948104739189,
-0.02421860210597515,
0.017011798918247223,
0.011356066912412643,
0.0634530708193779,
-0.007597501389682293,
-3.404874249213208e-8,
-0.09489046037197113,
0.08515888452529907,
-0.03958166763186455,
0.003988566808402538,
0.03528440743684769,
-0.000966136809438467,
0.006849547848105431,
0.03723818436264992,
0.0477234348654747,
0.011815700680017471,
0.00742915365844965,
-0.041242118924856186,
0.03165925294160843,
-0.04402215778827667,
-0.0010979747166857123,
0.07284893095493317,
-0.003529345616698265,
0.018013646826148033,
-0.00013387470971792936,
0.03457396477460861,
0.06279091536998749,
-0.019446874037384987,
0.008444291539490223,
0.021435242146253586,
-0.03967384621500969,
0.004164735786616802,
0.038680534809827805,
0.04840979352593422,
0.0857040286064148,
-0.018082765862345695,
-0.05655914917588234,
0.05183091387152672,
-0.014803511090576649,
-0.08642268925905228,
-0.10882139950990677,
0.07448182255029678,
0.04348969832062721,
-0.009564735926687717,
0.016146505251526833,
-0.021417057141661644,
-0.003829988418146968,
0.0119552630931139,
-0.03138551861047745,
0.03720371425151825,
-0.038832589983940125,
0.09822084754705429,
0.030068404972553253,
-0.1206323429942131,
0.031128311529755592,
-0.023694584146142006,
0.04274284094572067,
-0.05972083657979965,
-0.0012112305266782641,
0.10956048965454102,
0.048729728907346725,
0.04345306009054184,
0.032021913677453995,
0.025000352412462234,
0.02544005773961544,
0.05704498291015625,
-0.10373561829328537,
0.008230926468968391,
-0.05292053520679474,
0.01952267810702324
] |
Multimodal Lego: Model Merging and Fine-Tuning Across Topologies and Modalities in Biomedicine | https://openreview.net/forum?id=pH543jrbe8 | [
"Konstantin Hemker",
"Nikola Simidjievski",
"Mateja Jamnik"
] | Poster | Learning holistic computational representations in physical, chemical or biological systems requires the ability to process information from different distributions and modalities within the same model. Thus, the demand for multimodal machine learning models has sharply risen for modalities that go beyond vision and language, such as sequences, graphs, time series, or tabular data. While there are many available multimodal fusion and alignment approaches, most of them require end-to-end training, scale quadratically with the number of modalities, cannot handle cases of high modality imbalance in the training set, or are highly topology-specific, making them too restrictive for many biomedical learning tasks. This paper presents Multimodal Lego (MM-Lego), a general-purpose fusion framework to turn any set of encoders into a competitive multimodal model with no or minimal fine-tuning. We achieve this by introducing a wrapper for any unimodal encoder that enforces shape consistency between modality representations. It harmonises these representations by learning features in the frequency domain to enable model merging with little signal interference. We show that MM-Lego 1) can be used as a model merging method which achieves competitive performance with end-to-end fusion models without any fine-tuning, 2) can operate on any unimodal encoder, and 3) is a model fusion method that, with minimal fine-tuning, surpasses all benchmarks in five out of seven datasets. | multimodal, deep learning, fusion, biomedicine, model merging | Multimodal fusion and model merging approach for structurally heterogeneous modalities | 1,326 | null | [
-0.020676877349615097,
-0.14049416780471802,
0.009408305399119854,
-0.03662063553929329,
0.038531213998794556,
-0.04907720908522606,
-0.05339880660176277,
0.009938504546880722,
-0.012252158485352993,
-0.06427887827157974,
-0.010589166544377804,
-0.11334861069917679,
0.00578760402277112,
0.10397442430257797,
-0.00006697347271256149,
-0.0101155461743474,
-0.05102447420358658,
0.11970510333776474,
-0.10385160893201828,
-0.00022924818040337414,
0.02831306867301464,
0.001466394867748022,
0.018081076443195343,
0.028086276724934578,
0.028206609189510345,
0.09520482271909714,
0.012449759989976883,
-0.03235083818435669,
-0.016076019033789635,
-0.055964983999729156,
0.04051317647099495,
0.015463772229850292,
-0.12005842477083206,
0.021674856543540955,
-0.012457015924155712,
0.06252995133399963,
-0.0966164767742157,
0.05399569869041443,
-0.018437540158629417,
-0.07382091879844666,
0.014453260228037834,
0.08304349333047867,
-0.019924938678741455,
0.010483775287866592,
0.06978179514408112,
-0.0042663393542170525,
-0.04621737822890282,
-0.10569676756858826,
0.06519966572523117,
0.016871090978384018,
-0.11550088971853256,
-0.020087534561753273,
-0.005930151790380478,
0.06832049787044525,
0.06528665870428085,
-0.03273807466030121,
0.017063003033399582,
-0.07079681754112244,
-0.04067741334438324,
-0.056069716811180115,
0.013074423186480999,
-0.059973422437906265,
-0.02875664457678795,
-0.035630133002996445,
0.014875062741339207,
-0.0924161747097969,
0.061450328677892685,
0.06897792965173721,
-0.010585881769657135,
0.02515011467039585,
-0.018463395535945892,
0.023467011749744415,
-0.005857537966221571,
0.02391536720097065,
0.028358520939946175,
0.007940714247524738,
0.06917165219783783,
0.054653145372867584,
0.07389446347951889,
-0.004315435886383057,
0.038743793964385986,
0.06708446890115738,
0.02291417121887207,
-0.04290876165032387,
0.11045574396848679,
0.010951710864901543,
-0.05951744318008423,
0.06888040155172348,
-0.01393828820437193,
0.011450028978288174,
-0.03332218900322914,
0.007808682043105364,
0.0612935833632946,
-0.027710797265172005,
-0.03445868939161301,
0.016583174467086792,
0.06389374285936356,
-0.02088204398751259,
0.02866434119641781,
0.0901680588722229,
-0.03413600102066994,
0.015312598086893559,
-0.03962470963597298,
0.014832442626357079,
0.0959000363945961,
-0.0834222361445427,
0.054257672280073166,
0.07420170307159424,
0.07327111810445786,
-0.04080802947282791,
0.03603911027312279,
0.006347151007503271,
0.0378393717110157,
0.057566575706005096,
0.011402810923755169,
0.03849009796977043,
0.029024431481957436,
-0.016141951084136963,
0.10393093526363373,
0.055341172963380814,
-0.06790528446435928,
-0.04125719517469406,
-0.04579586163163185,
-0.010799627751111984,
0.08378047496080399,
0.015507606789469719,
-0.08056281507015228,
1.086752241516882e-33,
0.04385914281010628,
0.04752511903643608,
0.06883943825960159,
0.041807424277067184,
0.0817645713686943,
-0.019197963178157806,
-0.047198839485645294,
0.012636976316571236,
-0.01678551733493805,
0.042477235198020935,
-0.10107463598251343,
0.04216175153851509,
-0.09350830316543579,
0.1393112689256668,
-0.028711112216114998,
-0.028062833473086357,
-0.07404941320419312,
-0.005737181752920151,
0.01116328127682209,
-0.0418647937476635,
0.016397854313254356,
-0.0549151636660099,
-0.03837131708860397,
0.06136047840118408,
-0.0569254569709301,
0.019356897100806236,
0.01619540899991989,
-0.011372272856533527,
-0.005391972605139017,
0.013202516362071037,
-0.1316862404346466,
0.0027944764588028193,
0.021674633026123047,
0.004987476393580437,
0.07143642753362656,
-0.05657733231782913,
-0.013874556869268417,
-0.041834838688373566,
-0.015580050647258759,
-0.032149773091077805,
0.03404778242111206,
0.036605872213840485,
-0.050680551677942276,
-0.07099512219429016,
0.02260650135576725,
-0.0024725005496293306,
0.053809188306331635,
0.021958814933896065,
-0.06397457420825958,
-0.1000051349401474,
0.014821706339716911,
-0.0020536493975669146,
0.04123428836464882,
-0.1283556967973709,
-0.015224548056721687,
0.030495470389723778,
-0.007546762004494667,
0.008721204474568367,
0.09109535813331604,
0.015917198732495308,
-0.046700604259967804,
0.006749696563929319,
-0.0268127229064703,
0.10807058960199356,
0.0444302074611187,
0.012490695342421532,
-0.038672205060720444,
0.040918488055467606,
0.06662634760141373,
0.037971220910549164,
-0.04496380686759949,
0.047510258853435516,
-0.04832834750413895,
-0.045216966420412064,
0.0006026411429047585,
0.030420539900660515,
0.043122414499521255,
-0.06164737045764923,
-0.02735297940671444,
0.08135198801755905,
0.04687179997563362,
0.07897038757801056,
-0.07064589858055115,
-0.03356529772281647,
-0.020686710253357887,
-0.06323295831680298,
0.03535594791173935,
-0.04667491838335991,
-0.05822885408997536,
-0.03926708921790123,
-0.07089763879776001,
-0.002251945436000824,
0.01817071996629238,
-0.05599518492817879,
-0.04603263735771179,
-1.9296294455144644e-33,
0.0078019993379712105,
0.028786659240722656,
0.03160490095615387,
0.02394406497478485,
0.05286382511258125,
0.03623267635703087,
-0.015954142436385155,
-0.0011721911141648889,
0.023043448105454445,
-0.021606137976050377,
0.05760250240564346,
-0.045390740036964417,
0.0477566234767437,
-0.03324692323803902,
-0.03650103881955147,
0.030566047877073288,
-0.06646258383989334,
0.06246543303132057,
0.0359422042965889,
-0.0006723622791469097,
-0.0012686385307461023,
0.07016965001821518,
-0.08776339888572693,
0.030395962297916412,
-0.014397086575627327,
0.07774277031421661,
-0.09777513146400452,
0.007067832164466381,
0.009654952213168144,
0.018115492537617683,
-0.01183757558465004,
-0.07521125674247742,
0.023678438737988472,
-0.017642129212617874,
-0.00046577316243201494,
0.05686533451080322,
0.007723647635430098,
-0.05431898683309555,
0.030785415321588516,
0.06298463046550751,
0.006328497547656298,
-0.04642732813954353,
-0.13264213502407074,
0.042187463492155075,
0.052144140005111694,
-0.07190902531147003,
-0.0656738355755806,
-0.0656854659318924,
0.03239492326974869,
0.022962776944041252,
-0.010256772860884666,
0.012890482321381569,
-0.07083619385957718,
-0.08177319914102554,
-0.017796307802200317,
-0.09166929870843887,
0.00843043439090252,
-0.02035616524517536,
0.09281464666128159,
-0.03244629129767418,
-0.04412084072828293,
-0.11738211661577225,
0.010080498643219471,
0.0032228720374405384,
-0.044156596064567566,
0.012186717242002487,
0.06072143092751503,
-0.04112149029970169,
0.002268881071358919,
0.019857652485370636,
-0.042685121297836304,
0.055370233952999115,
0.04424629360437393,
0.014122703112661839,
0.04330216348171234,
-0.015825659036636353,
-0.0022126471158117056,
-0.03078017756342888,
-0.03428671509027481,
-0.011053062044084072,
-0.005202412139624357,
-0.11846869438886642,
-0.012966866604983807,
0.08112260699272156,
0.013105008751153946,
0.07201622426509857,
0.0676761344075203,
0.050535302609205246,
0.03774753957986832,
-0.03967072069644928,
0.004686694126576185,
0.027516113594174385,
0.0009705707780085504,
0.09854110330343246,
0.00011015744530595839,
-4.4139294885781055e-8,
-0.002877575345337391,
0.017791664227843285,
0.001718951971270144,
-0.07302960753440857,
-0.017956573516130447,
-0.03978850319981575,
-0.000884894747287035,
0.010086381807923317,
0.00836891494691372,
0.011075761169195175,
0.0922677218914032,
0.0019871830008924007,
-0.05560372769832611,
-0.00501649035140872,
0.04981590434908867,
0.05314783751964569,
-0.01927310973405838,
0.03248398378491402,
-0.009845231659710407,
-0.001658588764257729,
-0.022005382925271988,
-0.0031105964444577694,
0.03519674390554428,
0.010201350785791874,
0.054329290986061096,
-0.11165110766887665,
-0.10005922615528107,
0.06380479782819748,
0.050417277961969376,
-0.00841268990188837,
-0.025414463132619858,
-0.006900561507791281,
0.07928479462862015,
-0.01421146746724844,
-0.08332090824842453,
0.002899435581639409,
0.014035053551197052,
-0.016477379947900772,
-0.09987644851207733,
0.005038166418671608,
0.007744447328150272,
0.06512896716594696,
-0.1274784952402115,
0.01933044008910656,
-0.0005385276745073497,
0.01423789095133543,
0.08263087272644043,
-0.036376845091581345,
-0.019160926342010498,
-0.028097212314605713,
0.03179365023970604,
0.042520664632320404,
0.06145191565155983,
0.03714370355010033,
0.00853269174695015,
0.08638694137334824,
0.014431742019951344,
-0.008653840981423855,
0.07388225197792053,
-0.02229403518140316,
-0.012084380723536015,
-0.02179190143942833,
-0.06723146885633469,
-0.0059508695267140865
] |
Depth Pro: Sharp Monocular Metric Depth in Less Than a Second | https://openreview.net/forum?id=aueXfY0Clv | [
"Alexey Bochkovskiy",
"Amaël Delaunoy",
"Hugo Germain",
"Marcel Santos",
"Yichao Zhou",
"Stephan Richter",
"Vladlen Koltun"
] | Poster | We present a foundation model for zero-shot metric monocular depth estimation. Our model, Depth Pro, synthesizes high-resolution depth maps with unparalleled sharpness and high-frequency details. The predictions are metric, with absolute scale, without relying on the availability of metadata such as camera intrinsics. And the model is fast, producing a 2.25-megapixel depth map in 0.3 seconds on a standard GPU. These characteristics are enabled by a number of technical contributions, including an efficient multi-scale vision transformer for dense prediction, a training protocol that combines real and synthetic datasets to achieve high metric accuracy alongside fine boundary tracing, dedicated evaluation metrics for boundary accuracy in estimated depth maps, and state-of-the-art focal length estimation from a single image. Extensive experiments analyze specific design choices and demonstrate that Depth Pro outperforms prior work along multiple dimensions. We release code & weights at https://github.com/apple/ml-depth-pro | depth estimation, computer vision | A foundation model for zero-shot metric monocular depth estimation that synthesizes high-resolution depth maps with unparalleled sharpness and high-frequency detail | 1,325 | 2410.02073 | [
-0.054539043456315994,
-0.07306235283613205,
0.057594019919633865,
-0.044039029628038406,
0.04504670202732086,
-0.09988512098789215,
-0.059456292539834976,
0.040279872715473175,
-0.029535192996263504,
-0.0154141616076231,
-0.06194862350821495,
-0.10695908963680267,
-0.011266312561929226,
0.03483053669333458,
-0.018801288679242134,
0.022528192028403282,
0.05790096893906593,
0.0565296970307827,
-0.08601493388414383,
-0.02061561495065689,
0.06477605551481247,
-0.010826118290424347,
0.06829613447189331,
0.009445426054298878,
0.03229396790266037,
-0.04291664436459541,
0.058219246566295624,
-0.05884573608636856,
-0.048982180655002594,
-0.03573267534375191,
0.020517615601420403,
0.003446167102083564,
-0.012975337915122509,
-0.015489664860069752,
0.026240220293402672,
0.0035581672564148903,
0.054488781839609146,
-0.0009057060233317316,
-0.03151025250554085,
-0.040239982306957245,
-0.035628367215394974,
0.07179603725671768,
0.03293384611606598,
0.11367785185575485,
0.04765969514846802,
0.07287421822547913,
0.06466451287269592,
-0.06574124097824097,
0.06280592828989029,
0.005300234537571669,
-0.03408375009894371,
-0.03737182915210724,
-0.01849888637661934,
0.029905110597610474,
0.00042432371992617846,
0.09849125891923904,
0.0298234224319458,
-0.09149853885173798,
-0.014620660804212093,
-0.0012599334586411715,
-0.022170601412653923,
-0.06577933579683304,
-0.03370356932282448,
-0.001174607197754085,
-0.04821358621120453,
-0.0376972071826458,
0.007171186152845621,
-0.06353337317705154,
-0.034090060740709305,
-0.009695773012936115,
-0.06558668613433838,
0.062189459800720215,
-0.02733135223388672,
0.004784113727509975,
-0.02537086419761181,
-0.008681882172822952,
0.02886456623673439,
-0.038295991718769073,
0.03349221125245094,
-0.059028204530477524,
0.0706324651837349,
-0.029213786125183105,
0.015334744937717915,
0.03943664953112602,
-0.004527835175395012,
0.0032906474079936743,
0.017436053603887558,
0.089542455971241,
0.007680728565901518,
-0.010529291816055775,
-0.016450487077236176,
-0.08160758763551712,
-0.07696589827537537,
-0.021436909213662148,
-0.01707317680120468,
-0.03236766159534454,
0.00960069615393877,
-0.12336679548025131,
-0.05713508650660515,
0.08356837183237076,
-0.030710559338331223,
-0.028681792318820953,
-0.023755235597491264,
-0.022741444408893585,
0.07729644328355789,
0.09198866039514542,
0.04350270330905914,
0.08262307196855545,
0.036844633519649506,
0.014516566880047321,
0.09925585985183716,
-0.009528553113341331,
0.017889179289340973,
-0.0036000856198370457,
0.06612426042556763,
0.0842149630188942,
0.007025667931884527,
0.017731312662363052,
-0.012287973426282406,
0.02055295556783676,
-0.09624370187520981,
-0.009230228140950203,
-0.05292634293437004,
0.03110922873020172,
-0.02369624190032482,
0.0002889450406655669,
-0.017032012343406677,
2.1793696488313904e-33,
0.025432169437408447,
0.11061230301856995,
0.09468094259500504,
0.02168949879705906,
0.013706437312066555,
-0.018120229244232178,
0.0682794451713562,
0.040569007396698,
-0.005370337516069412,
0.03456563875079155,
-0.054484643042087555,
-0.022502241656184196,
-0.042354267090559006,
0.08313573151826859,
-0.03521014750003815,
-0.003628208301961422,
0.006581765133887529,
0.08099320530891418,
-0.01269566174596548,
0.10811638832092285,
0.008891255594789982,
-0.017168190330266953,
0.0347532220184803,
0.00020477409998420626,
0.02696578949689865,
-0.019315090030431747,
-0.04432058334350586,
0.025034969672560692,
0.003545676590874791,
0.0031996374018490314,
-0.08707980066537857,
0.022295420989394188,
0.008662045001983643,
0.012788953259587288,
-0.014661455526947975,
-0.03565424308180809,
-0.012865385971963406,
-0.033613450825214386,
-0.028661543503403664,
0.00533420080319047,
-0.0023739344906061888,
0.06197057664394379,
-0.007282602600753307,
-0.07356519997119904,
-0.10974393039941788,
-0.056968409568071365,
-0.03158748149871826,
0.05551544949412346,
-0.09487414360046387,
0.08906982839107513,
0.10814128071069717,
0.009885035455226898,
-0.05845892056822777,
-0.06698772311210632,
-0.10517606139183044,
-0.024125538766384125,
0.04079296067357063,
-0.0391458198428154,
0.026466459035873413,
0.018554920330643654,
0.029552076011896133,
-0.00878183264285326,
-0.08942105621099472,
0.055155545473098755,
-0.08099525421857834,
-0.017511988058686256,
0.0028729878831654787,
0.02798069827258587,
-0.01832575909793377,
0.03259972110390663,
-0.062447380274534225,
0.002735527465119958,
0.02142934314906597,
-0.05940505862236023,
0.03984838351607323,
0.07987746596336365,
0.009686971083283424,
-0.06286496669054031,
0.06756805628538132,
0.10823080688714981,
-0.07751275599002838,
0.101566381752491,
0.052953511476516724,
-0.017654912546277046,
-0.017187686637043953,
0.021397192031145096,
0.010087407194077969,
-0.0037239431403577328,
-0.04783853515982628,
-0.04081900045275688,
-0.053118158131837845,
-0.07523979246616364,
-0.02769557572901249,
-0.04069902002811432,
-0.03755519539117813,
-8.39864907888104e-34,
0.07921522855758667,
0.027724692597985268,
-0.019002655521035194,
0.06715507805347443,
-0.0036159527953714132,
-0.05130891874432564,
0.0217638798058033,
0.07971513271331787,
0.05820755288004875,
-0.06640862673521042,
0.009583570994436741,
0.08084332197904587,
0.057457055896520615,
-0.020601468160748482,
0.07240291684865952,
-0.017632363364100456,
0.0019026192603632808,
-0.006374379154294729,
-0.0982976108789444,
-0.024416694417595863,
0.08859427273273468,
0.041970737278461456,
-0.00941827055066824,
0.038366708904504776,
-0.10612818598747253,
0.07564651221036911,
-0.0584196113049984,
-0.006272182799875736,
0.007961668074131012,
-0.040813133120536804,
0.057494357228279114,
-0.06103545054793358,
0.062293220311403275,
-0.05219157412648201,
-0.03192778304219246,
0.02518652006983757,
0.046208787709474564,
-0.01718299277126789,
-0.06366096436977386,
0.027014674618840218,
0.0005738285253755748,
-0.01337859220802784,
0.0040862285532057285,
0.05768505856394768,
-0.047881171107292175,
0.04497651383280754,
0.010209809057414532,
-0.05132380500435829,
0.049603745341300964,
0.03144621104001999,
-0.13008002936840057,
0.04170001298189163,
-0.05198600888252258,
-0.0022774217650294304,
0.030681798234581947,
-0.08959055691957474,
-0.10758756846189499,
0.006947813555598259,
0.10492319613695145,
-0.0005147831398062408,
-0.02049296721816063,
-0.025870123878121376,
-0.11042571067810059,
0.0384606197476387,
0.03153092786669731,
0.029105108231306076,
-0.07730785012245178,
-0.020382383838295937,
0.034848764538764954,
0.10523844510316849,
0.019272414967417717,
-0.003209881717339158,
0.01607782579958439,
-0.0060530840419232845,
-0.02591109275817871,
-0.0001182517153210938,
0.042128000408411026,
-0.05419265106320381,
0.05851585417985916,
0.054307278245687485,
0.05206168815493584,
-0.050525836646556854,
0.02140423096716404,
0.12923985719680786,
0.13477471470832825,
0.07142533361911774,
0.029247676953673363,
-0.016408901661634445,
0.07694820314645767,
-0.0038721428718417883,
-0.0653717964887619,
0.0026104594580829144,
-0.05585674196481705,
-0.033634334802627563,
-0.01469830609858036,
-4.146118470771398e-8,
0.017509348690509796,
0.09232132881879807,
0.03339675813913345,
-0.009669418446719646,
-0.039154887199401855,
-0.06340090930461884,
0.08532550185918808,
0.09981376677751541,
0.014370756223797798,
0.019089560955762863,
0.04014594852924347,
-0.05057153105735779,
-0.01623694598674774,
0.008902939967811108,
0.020313464105129242,
0.05221099033951759,
0.011996718123555183,
0.018846051767468452,
0.0032815781887620687,
0.06050737574696541,
-0.09184600412845612,
0.03285297751426697,
0.013260220177471638,
-0.07334034889936447,
0.03253177925944328,
-0.09273232519626617,
-0.032647084444761276,
0.06157442554831505,
0.08592120558023453,
-0.040992867201566696,
0.03435175120830536,
0.013192463666200638,
0.015201671980321407,
-0.022006088867783546,
-0.03618387505412102,
-0.01722835749387741,
0.01350710541009903,
0.059745002537965775,
0.01910787634551525,
-0.02056501992046833,
-0.05606678128242493,
0.06784738600254059,
0.01334993913769722,
0.01820727065205574,
0.009139010682702065,
0.05158001929521561,
0.07128196209669113,
-0.03197845444083214,
-0.07254859805107117,
-0.04429778456687927,
0.006703164894133806,
0.044820647686719894,
-0.026105403900146484,
0.08589308708906174,
0.06044123321771622,
0.051480285823345184,
-0.010783368721604347,
-0.0545598641037941,
-0.02342722937464714,
0.01822635531425476,
0.006594257894903421,
-0.061745401471853256,
-0.08511523902416229,
-0.034343767911195755
] |
Zero-Shot Natural Language Explanations | https://openreview.net/forum?id=X6VVK8pIzZ | [
"Fawaz Sammani",
"Nikos Deligiannis"
] | Poster | Natural Language Explanations (NLEs) interpret the decision-making process of a given model through textual sentences. Current NLEs suffer from a severe limitation; they are unfaithful to the model’s actual reasoning process, as a separate textual decoder is explicitly trained to generate those explanations using annotated datasets for a specific task, leading them to reflect what annotators desire. In this work, we take the first step towards generating faithful NLEs for any visual classification model without any training data. Our approach models the relationship between class embeddings from the classifier of the vision model and their corresponding class names via a simple MLP which trains in seconds. After training, we can map any new text to the classifier space and measure its association with the visual features. We conduct experiments on 38 vision models, including both CNNs and Transformers. In addition to NLEs, our method offers other advantages such as zero-shot image classification and fine-grained concept discovery. | Natural Language Explanations, interpretability, explainability | null | 1,324 | null | [
-0.05282125622034073,
-0.08559712022542953,
0.07348988205194473,
0.038491521030664444,
0.15591977536678314,
-0.041627541184425354,
-0.026722345501184464,
0.0012939788866788149,
0.015501756221055984,
-0.028125464916229248,
-0.023615442216396332,
-0.11875864863395691,
0.016395891085267067,
0.03942011669278145,
-0.06491374969482422,
0.03583098575472832,
0.09878719598054886,
0.029683062806725502,
-0.09489846229553223,
-0.016283048316836357,
0.029578829184174538,
-0.01878596656024456,
0.08359356224536896,
0.024022094905376434,
0.018157703801989555,
0.011079925112426281,
0.022738823667168617,
-0.03477083146572113,
0.020607803016901016,
-0.03679688647389412,
-0.013229592703282833,
-0.055467639118433,
-0.04208814725279808,
0.10564251989126205,
-0.07840032130479813,
0.06286866217851639,
0.03259849548339844,
0.08482825011014938,
-0.0395849347114563,
-0.04889467731118202,
-0.039951141923666,
0.03332006558775902,
0.021014656871557236,
0.05807841941714287,
0.11404319852590561,
0.021636011078953743,
-0.012130357325077057,
-0.04617806151509285,
0.00418397132307291,
-0.016800377517938614,
-0.09790512919425964,
-0.04190092533826828,
-0.03127409517765045,
0.012821181677281857,
-0.06071348488330841,
0.010987116023898125,
0.06376701593399048,
0.007721514906734228,
0.02474568970501423,
0.01194437500089407,
-0.000516346248332411,
-0.07148604094982147,
0.03692839667201042,
0.04506474360823631,
-0.04351105913519859,
0.012768835760653019,
-0.005783914588391781,
0.04032418131828308,
-0.0001251746725756675,
-0.00297291437163949,
-0.028118811547756195,
0.08720452338457108,
-0.026482146233320236,
0.004545432515442371,
-0.02054252102971077,
0.08720584958791733,
-0.001356904860585928,
-0.014508688822388649,
0.0365474559366703,
-0.05113610252737999,
-0.02904985286295414,
-0.005871285684406757,
0.05711347237229347,
-0.01615610159933567,
0.1046447679400444,
0.007435995154082775,
0.002378274453803897,
0.05551305413246155,
-0.017741655930876732,
0.023703983053565025,
-0.11765246838331223,
-0.10533887147903442,
0.033000484108924866,
-0.030806630849838257,
0.05074060708284378,
0.09019513428211212,
-0.040674977004528046,
-0.07116182148456573,
-0.07844420522451401,
0.029322657734155655,
-0.062593974173069,
0.08308420330286026,
0.04163484275341034,
-0.09106215834617615,
0.04791000485420227,
0.0236384067684412,
0.02489188313484192,
-0.031037544831633568,
0.02566622570157051,
-0.08643591403961182,
0.04541148617863655,
0.027435483410954475,
-0.029576847329735756,
0.009570819325745106,
0.044630762189626694,
-0.037028077989816666,
0.004905594978481531,
-0.0570346936583519,
0.004538627807050943,
0.025235788896679878,
0.006440701428800821,
0.06450410187244415,
-0.048578355461359024,
0.01743544451892376,
0.0053811888210475445,
-0.027388490736484528,
-0.008339747786521912,
1.2932535381777411e-33,
0.0463147833943367,
0.014385581016540527,
0.05765552073717117,
0.055223144590854645,
0.08020638674497604,
-0.026339147239923477,
-0.02051081880927086,
0.08032893389463425,
-0.005442393012344837,
0.005854581017047167,
-0.0062919133342802525,
-0.02797381766140461,
-0.051908720284700394,
0.06651116162538528,
0.017479320988059044,
0.01779697649180889,
-0.07185577601194382,
-0.0008997851400636137,
-0.04529806599020958,
-0.0006057160790078342,
-0.0740945115685463,
0.022378956899046898,
-0.039635371416807175,
-0.023889267817139626,
0.011375851929187775,
-0.013365358114242554,
0.015368668362498283,
-0.08200589567422867,
-0.07598947733640671,
-0.015190553851425648,
-0.07723238319158554,
0.004658734891563654,
0.05393998697400093,
-0.023585908114910126,
0.03312615677714348,
-0.07845710217952728,
-0.020481057465076447,
0.008162357844412327,
0.058066800236701965,
-0.00906013511121273,
-0.02596372924745083,
0.01169705018401146,
-0.06380309164524078,
-0.008898009546101093,
-0.02494136057794094,
-0.022714857012033463,
-0.06976301223039627,
0.06933771073818207,
-0.02366110496222973,
0.02678726799786091,
0.07896924018859863,
0.03646887466311455,
-0.06884551048278809,
-0.07358086109161377,
-0.01639513485133648,
-0.013906670734286308,
0.004994733724743128,
0.01585637778043747,
-0.024717263877391815,
-0.03765266761183739,
0.03842545673251152,
0.026365451514720917,
0.002773509593680501,
0.06721341609954834,
-0.004774897824972868,
0.026606030762195587,
0.029365595430135727,
0.02560201846063137,
0.023265095427632332,
0.003427983494475484,
-0.04272281378507614,
0.055655620992183685,
-0.0627327486872673,
-0.029368113726377487,
0.06911659240722656,
0.03732173144817352,
0.024787522852420807,
-0.11490532010793686,
0.033557794988155365,
0.09905469417572021,
0.0465024895966053,
-0.09526604413986206,
0.03118712268769741,
-0.05156858637928963,
-0.058803439140319824,
-0.05233404412865639,
0.04199082776904106,
-0.06538430601358414,
0.004482405725866556,
-0.0401979424059391,
-0.03445044159889221,
-0.018981847912073135,
-0.07911462336778641,
0.005399836227297783,
0.01840839348733425,
-1.5097446082751703e-33,
-0.0012157794553786516,
0.020601283758878708,
-0.04179590567946434,
0.03076084516942501,
-0.02142515778541565,
0.00800986960530281,
-0.015467911958694458,
0.06174638867378235,
0.055552463978528976,
0.016260981559753418,
-0.011401191353797913,
-0.0241395290941,
-0.0011470533208921552,
0.014274798333644867,
0.031035158783197403,
-0.012675726786255836,
0.010973935015499592,
0.03274943679571152,
-0.039777111262083054,
0.07143016904592514,
0.04857126250863075,
0.143501877784729,
-0.13906912505626678,
0.01756303571164608,
-0.008670332841575146,
0.07777276635169983,
-0.02227693237364292,
0.0720084086060524,
-0.04512885585427284,
-0.013747469522058964,
-0.009143149480223656,
0.020023344084620476,
0.011848784983158112,
-0.09073711186647415,
-0.0914885476231575,
0.01725572533905506,
0.06783891469240189,
-0.01668035052716732,
-0.024243313819169998,
0.07071013003587723,
0.14654867351055145,
-0.04445596784353256,
-0.1048252135515213,
-0.023669874295592308,
-0.11211664974689484,
-0.09022752940654755,
-0.03029821068048477,
-0.01848391816020012,
-0.0011882339604198933,
0.04019889980554581,
-0.06627658009529114,
0.0008110581547953188,
-0.0718294009566307,
0.03458058089017868,
-0.01931295171380043,
-0.08400499075651169,
-0.03685203939676285,
0.005304464604705572,
0.05982052907347679,
0.0009075882844626904,
-0.03605477511882782,
0.052321549504995346,
0.033152952790260315,
-0.0250565093010664,
-0.005533379968255758,
-0.008358808234333992,
-0.07760974764823914,
0.02740987204015255,
0.016458779573440552,
-0.004809336271136999,
0.15429455041885376,
-0.02436521090567112,
-0.06131716072559357,
0.010610184632241726,
0.02076140232384205,
0.02709195949137211,
0.07538553327322006,
-0.005398803856223822,
-0.008490328676998615,
-0.08081524819135666,
0.06185152381658554,
-0.03263244032859802,
0.07906585186719894,
0.07011649012565613,
0.13366122543811798,
0.05428065359592438,
0.04648794233798981,
0.020720407366752625,
0.022911924868822098,
0.024355167523026466,
-0.0446503683924675,
0.016899405047297478,
-0.014276159927248955,
0.1377929151058197,
0.018983161076903343,
-3.8223596732223086e-8,
-0.052770547568798065,
0.05180990323424339,
-0.013265286572277546,
0.06637421995401382,
-0.006131602451205254,
-0.023347368463873863,
-0.08835597336292267,
0.08476196974515915,
-0.060801222920417786,
-0.03215807303786278,
0.08794386684894562,
0.041206952184438705,
-0.03657113015651703,
-0.03973978012800217,
-0.016855385154485703,
0.09139657765626907,
-0.03088333085179329,
-0.005220149178057909,
0.008453994989395142,
0.03906182199716568,
0.02588004246354103,
-0.02452574484050274,
-0.003618640825152397,
0.018595809116959572,
0.006848577875643969,
-0.07284326106309891,
-0.014234893955290318,
0.08382473886013031,
0.05063330754637718,
-0.005158623214811087,
-0.01466432400047779,
0.04875124990940094,
-0.03824367746710777,
-0.027010370045900345,
0.030012695118784904,
0.054934918880462646,
0.0033593534026294947,
-0.058871131390333176,
-0.028941625729203224,
-0.046565841883420944,
-0.02175881341099739,
0.05157707631587982,
-0.009753856807947159,
-0.01861656829714775,
0.043719854205846786,
0.01461781281977892,
0.050012584775686264,
-0.15040569007396698,
0.03448748588562012,
0.0013971987646073103,
-0.09834378212690353,
-0.017424168065190315,
-0.04049605876207352,
0.11903142184019089,
-0.022436341270804405,
-0.000674092851113528,
-0.022349556908011436,
0.013186697848141193,
-0.01617310754954815,
0.03554719313979149,
-0.021548103541135788,
0.06556744128465652,
0.0290204044431448,
0.01172617543488741
] |
Add-it: Training-Free Object Insertion in Images With Pretrained Diffusion Models | https://openreview.net/forum?id=ZeaTvXw080 | [
"Yoad Tewel",
"Rinon Gal",
"Dvir Samuel",
"Yuval Atzmon",
"Lior Wolf",
"Gal Chechik"
] | Poster | Adding Object into images based on text instructions is a challenging task in semantic image editing, requiring a balance between preserving the original scene and seamlessly integrating the new object in a fitting location. Despite extensive efforts, existing models often struggle with this balance, particularly with finding a natural location for adding an object in complex scenes. We introduce Add-it, a training-free approach that extends diffusion models' attention mechanisms to incorporate information from three key sources: the scene image, the text prompt, and the generated image itself. Our weighted extended-attention mechanism maintains structural consistency and fine details while ensuring natural object placement. Without task-specific fine-tuning, Add-it achieves state-of-the-art results on both real and generated image insertion benchmarks, including our newly constructed "Additing Affordance Benchmark" for evaluating object placement plausibility, outperforming supervised methods. Human evaluations show that Add-it is preferred in over 80% of cases, and it also demonstrates improvements in various automated metrics. | Diffusion, Editing, Affordance | We presents a method to insert objects into existing images using pretrained diffusion models, guided by text prompts, without additional training. | 1,318 | null | [
0.06636758893728256,
-0.06514272838830948,
0.04316086694598198,
0.09748972952365875,
0.0953945517539978,
-0.020967306569218636,
-0.054129958152770996,
-0.0029634549282491207,
0.03919374942779541,
-0.01084186416119337,
0.0429646335542202,
-0.037805523723363876,
0.036859653890132904,
0.06222720444202423,
0.03031889908015728,
-0.005264460574835539,
0.0959668681025505,
0.008138861507177353,
-0.10420579463243484,
-0.021743789315223694,
-0.007273238152265549,
-0.045774493366479874,
0.15081587433815002,
-0.04299589619040489,
0.02533368207514286,
0.020486218854784966,
-0.051793720573186874,
-0.08511129021644592,
0.03575805202126503,
-0.05746772140264511,
0.03295949473977089,
-0.040044985711574554,
0.007801057770848274,
0.10517790168523788,
-0.040647607296705246,
0.05963104963302612,
-0.048501696437597275,
0.055912017822265625,
0.011808564886450768,
-0.07260485738515854,
0.02215622365474701,
-0.026300573721528053,
0.0030324191320687532,
0.0358097106218338,
0.09407278150320053,
-0.016654929146170616,
0.027957657352089882,
-0.05838777497410774,
0.015082045458257198,
-0.05081702023744583,
-0.04812770336866379,
-0.06636383384466171,
-0.0784175768494606,
0.06881465762853622,
-0.08305247128009796,
0.11435770243406296,
0.10471808165311813,
-0.020098848268389702,
0.04800686985254288,
-0.019653158262372017,
0.02111790142953396,
-0.02161002904176712,
0.04176412150263786,
-0.014220972545444965,
0.01980200596153736,
-0.0374123677611351,
0.023860108107328415,
0.0008223889744840562,
-0.054739076644182205,
-0.007036968600004911,
0.08370345085859299,
0.06827496737241745,
0.009205861017107964,
0.01856507547199726,
0.037801701575517654,
0.017702607437968254,
-0.013122739270329475,
-0.0000831462602945976,
-0.008350944146513939,
-0.05728732421994209,
0.02792067639529705,
-0.017585162073373795,
0.08259239047765732,
0.02064328081905842,
0.08273905515670776,
-0.016567032784223557,
0.02253367379307747,
0.03183070197701454,
-0.04797537997364998,
-0.012266279198229313,
-0.03045637346804142,
-0.05448726937174797,
-0.04083910584449768,
-0.025862976908683777,
0.014620012603700161,
0.013328030705451965,
-0.03410330042243004,
-0.09678637236356735,
-0.08305247128009796,
0.03286101296544075,
-0.01104605384171009,
-0.005122890695929527,
-0.0024088509380817413,
0.00018234543676953763,
0.05748484656214714,
-0.017671948298811913,
-0.02217797003686428,
-0.028852611780166626,
-0.014215943403542042,
-0.024955011904239655,
0.06484539806842804,
0.04136332869529724,
-0.014434091746807098,
-0.037579357624053955,
-0.016971459612250328,
-0.030302491039037704,
-0.022326473146677017,
0.012541312724351883,
0.01211303099989891,
-0.005004452541470528,
0.04566125571727753,
0.022096719592809677,
-0.036360159516334534,
0.029321009293198586,
-0.004312323871999979,
-0.060274600982666016,
-0.06759936362504959,
2.291884111001261e-33,
0.004995172377675772,
-0.0035341866314411163,
0.034085117280483246,
0.04622437059879303,
0.08982373774051666,
0.015111868269741535,
-0.047869592905044556,
0.0540725514292717,
-0.06049450859427452,
-0.06254790723323822,
-0.020184917375445366,
-0.07907558232545853,
-0.05610242113471031,
0.1450490802526474,
0.04010792449116707,
-0.022890880703926086,
-0.024590227752923965,
0.09144788980484009,
-0.04245932772755623,
0.056321918964385986,
-0.060508206486701965,
-0.04230158403515816,
-0.06879548728466034,
-0.0393003486096859,
0.014762834645807743,
0.03356768190860748,
0.07864703983068466,
-0.050794217735528946,
-0.0010904389200732112,
0.0038197289686650038,
-0.11890202015638351,
-0.023600859567523003,
0.04824593663215637,
0.05365828797221184,
0.0005512141506187618,
-0.07057082653045654,
-0.03394017741084099,
-0.03628954291343689,
0.08080969750881195,
-0.027308408170938492,
0.03248811885714531,
0.08447321504354477,
-0.0444769561290741,
0.02276507019996643,
-0.07480186969041824,
0.01908786967396736,
0.002196239773184061,
0.03322966769337654,
0.000008250802238762844,
0.027409173548221588,
0.025122299790382385,
-0.005310961045324802,
-0.02075178362429142,
-0.0022277934476733208,
-0.09131644666194916,
-0.0335933081805706,
0.05519292131066322,
-0.038568347692489624,
0.05860619246959686,
-0.033826909959316254,
0.09932547062635422,
0.07086554169654846,
-0.07890135794878006,
0.12704043090343475,
0.019097456708550453,
-0.03457871451973915,
0.03852979466319084,
0.06542297452688217,
0.051918286830186844,
0.032209612429142,
-0.05320292338728905,
0.04031079635024071,
-0.06079263612627983,
0.010242022573947906,
0.06646180152893066,
-0.05336378142237663,
-0.014325040392577648,
-0.045193541795015335,
0.020512644201517105,
0.021694593131542206,
-0.018062766641378403,
-0.03514135628938675,
-0.010065886192023754,
-0.019418949261307716,
-0.03668470308184624,
-0.04752205312252045,
-0.003120789071545005,
-0.09172307699918747,
-0.006219195667654276,
-0.030990375205874443,
0.03126249834895134,
-0.042268313467502594,
-0.07747461646795273,
0.05231271684169769,
-0.013191750273108482,
-4.187920407527621e-34,
0.06233949959278107,
-0.005551868584007025,
-0.005635361652821302,
0.0373237244784832,
0.0010601587127894163,
0.017879320308566093,
-0.018848299980163574,
0.02511349320411682,
0.0630093663930893,
-0.05662626400589943,
-0.0525958351790905,
0.03536012023687363,
-0.0500677153468132,
-0.06750186532735825,
-0.03378315642476082,
0.0038205499295145273,
-0.02335350215435028,
0.037507086992263794,
-0.05117803439497948,
-0.00025511736748740077,
0.038807522505521774,
0.03535981848835945,
0.005821278318762779,
-0.016340086236596107,
-0.08194832503795624,
0.04739905893802643,
-0.03213692829012871,
0.034269653260707855,
-0.022237060591578484,
-0.047953736037015915,
-0.10670806467533112,
0.0046938699670135975,
0.03129267692565918,
-0.03314581885933876,
-0.05803535878658295,
0.022816305980086327,
0.03194832429289818,
-0.010164892300963402,
-0.04954754561185837,
0.09383291751146317,
0.07667666673660278,
-0.04137440398335457,
-0.06446679681539536,
0.06848664581775665,
-0.08242528140544891,
-0.026415815576910973,
-0.05974271148443222,
-0.009526422247290611,
-0.009641218930482864,
0.14386776089668274,
-0.06518165022134781,
0.0017511422047391534,
-0.10869363695383072,
-0.019773246720433235,
-0.031639326363801956,
-0.015559562481939793,
-0.0782914087176323,
-0.03687133267521858,
0.09885338693857193,
-0.05993067845702171,
-0.059908922761678696,
0.005478454288095236,
0.021731412038207054,
-0.06442280113697052,
-0.011601920239627361,
-0.010876154527068138,
-0.0794583410024643,
-0.06135094538331032,
0.03705582767724991,
0.02225547470152378,
0.08098684996366501,
-0.03449469059705734,
-0.016362950205802917,
0.029314054176211357,
-0.013827409595251083,
-0.0010367193026468158,
0.06464600563049316,
-0.018388984724879265,
0.05258746072649956,
-0.1076047420501709,
-0.007335740607231855,
-0.02390996553003788,
-0.027147376909852028,
0.04227166622877121,
0.07717692106962204,
0.08057226240634918,
0.05964105948805809,
0.007306766230612993,
-0.01501344796270132,
0.04802035167813301,
-0.08640015125274658,
0.08177868276834488,
0.04260212928056717,
0.05793291702866554,
-0.06694148480892181,
-3.972321138689949e-8,
-0.07230141758918762,
0.004241753835231066,
-0.004668686538934708,
0.0747794657945633,
-0.0020241790916770697,
-0.07284516096115112,
0.03535829856991768,
0.14046964049339294,
-0.030131056904792786,
-0.0347529835999012,
0.040367063134908676,
0.04227444529533386,
-0.01285037025809288,
-0.05774626508355141,
-0.05747207999229431,
0.06729501485824585,
-0.019519207999110222,
0.03660723939538002,
-0.0328754261136055,
-0.013273614458739758,
0.04116306081414223,
0.08831590414047241,
0.08784522116184235,
-0.032135553658008575,
0.03874067962169647,
-0.03199721872806549,
-0.034298013895750046,
0.05517423152923584,
0.024593107402324677,
-0.053549692034721375,
0.07070573419332504,
0.002089365851134062,
-0.0048430622555315495,
0.07726237922906876,
0.020616423338651657,
0.06911086291074753,
-0.007991778664290905,
-0.003496549092233181,
-0.04154055193066597,
-0.07472144812345505,
0.07048912346363068,
0.03757162019610405,
0.027905073016881943,
0.024782055988907814,
0.08339357376098633,
0.01780885085463524,
0.03813906013965607,
-0.06240027770400047,
0.038149066269397736,
-0.045729395002126694,
-0.013665723614394665,
-0.024404603987932205,
-0.040138114243745804,
0.04965035617351532,
-0.03407934308052063,
0.023558087646961212,
0.08502187579870224,
0.012959197163581848,
0.06462536007165909,
0.1534375697374344,
0.013563459739089012,
0.028701700270175934,
-0.019293159246444702,
-0.025315457955002785
] |
RecDreamer: Consistent Text-to-3D Generation via Uniform Score Distillation | https://openreview.net/forum?id=aucMP9hGYv | [
"Chenxi Zheng",
"Yihong Lin",
"Bangzhen Liu",
"Xuemiao Xu",
"Yongwei Nie",
"Shengfeng He"
] | Poster | Current text-to-3D generation methods based on score distillation often suffer from geometric inconsistencies, leading to repeated patterns across different poses of 3D assets. This issue, known as the Multi-Face Janus problem, arises because existing methods struggle to maintain consistency across varying poses and are biased toward a canonical pose. While recent work has improved pose control and approximation, these efforts are still limited by this inherent bias, which skews the guidance during generation.
To address this, we propose a solution called RecDreamer, which reshapes the underlying data distribution to achieve more consistent pose representation. The core idea behind our method is to rectify the prior distribution, ensuring that pose variation is uniformly distributed rather than biased toward a canonical form. By modifying the prescribed distribution through an auxiliary function, we can reconstruct the density of the distribution to ensure compliance with specific marginal constraints. In particular, we ensure that the marginal distribution of poses follows a uniform distribution, thereby eliminating the biases introduced by the prior knowledge.
We incorporate this rectified data distribution into existing score distillation algorithms, a process we refer to as uniform score distillation. To efficiently compute the posterior distribution required for the auxiliary function, RecDreamer introduces a training-free classifier that estimates pose categories in a plug-and-play manner. Additionally, we utilize various approximation techniques for noisy states, significantly improving system performance.
Our experimental results demonstrate that RecDreamer effectively mitigates the Multi-Face Janus problem, leading to more consistent 3D asset generation across different poses. | 3D generation | A new text-to-3D generation method with superior view-consistency. | 1,315 | 2502.12640 | [
-0.04199874401092529,
-0.022510869428515434,
0.03508606553077698,
-0.06583816558122635,
0.025660699233412743,
0.00046548323007300496,
-0.028694327920675278,
-0.06578848510980606,
0.03524116054177284,
0.025805164128541946,
-0.03618822246789932,
-0.010280250571668148,
0.11338373273611069,
0.06247438117861748,
0.005583535879850388,
-0.02577519416809082,
0.03154076635837555,
0.14115861058235168,
-0.04885680601000786,
0.0880330428481102,
-0.002136173425242305,
0.007004912942647934,
0.015095905400812626,
-0.03334549441933632,
0.05842713639140129,
-0.025006085634231567,
0.09140408784151077,
-0.04882754758000374,
0.013467731885612011,
-0.07598859071731567,
0.0875631719827652,
0.04854004830121994,
-0.033629100769758224,
0.008497143164277077,
-0.06266068667173386,
0.08679676800966263,
-0.10664879530668259,
0.07050582021474838,
-0.0352732315659523,
0.05782410129904747,
0.03094588965177536,
0.11659080535173416,
0.011754297651350498,
0.03851717337965965,
0.14280621707439423,
0.11700969189405441,
-0.027116812765598297,
0.011597187258303165,
0.08558370918035507,
0.022201837971806526,
-0.03654029220342636,
-0.046963658183813095,
0.009099915623664856,
0.004348615650087595,
-0.04434845596551895,
-0.02547471411526203,
0.04616234079003334,
-0.08050284534692764,
0.04160057753324509,
-0.01156268548220396,
0.05207463726401329,
-0.032606396824121475,
-0.008651095442473888,
-0.027250003069639206,
-0.005442750174552202,
-0.032837629318237305,
0.011473911814391613,
0.00987813901156187,
-0.031214242801070213,
0.07575178891420364,
-0.03612822666764259,
0.05969516187906265,
0.030137918889522552,
0.024214228615164757,
0.04903748258948326,
0.06853814423084259,
-0.017291506752371788,
-0.060798995196819305,
-0.03081653080880642,
-0.09542544931173325,
0.029626168310642242,
0.026602795347571373,
0.0868324562907219,
0.03190971165895462,
0.0737559050321579,
0.008003970608115196,
-0.034836385399103165,
0.07188532501459122,
0.020971469581127167,
0.026103632524609566,
-0.12128131091594696,
-0.03736848011612892,
0.021373344585299492,
-0.04521311819553375,
-0.010085293091833591,
0.0014429143629968166,
0.025482846423983574,
-0.07236172258853912,
0.0350981242954731,
0.07237020879983902,
0.010022816248238087,
0.04449354484677315,
0.0661160945892334,
-0.00747488671913743,
0.008727156557142735,
-0.004663094412535429,
0.021268300712108612,
-0.010115000419318676,
-0.014463888481259346,
-0.09786254167556763,
0.05266701430082321,
-0.0397917777299881,
0.0035394360311329365,
0.026919333264231682,
-0.014258691109716892,
-0.0341637022793293,
-0.02500908635556698,
-0.02115785703063011,
-0.0014356216415762901,
-0.08223719894886017,
0.015736086294054985,
-0.037165723741054535,
0.05894303321838379,
-0.008070151321589947,
-0.04729575291275978,
-0.013363667763769627,
0.013387465849518776,
2.43711607070337e-33,
-0.00344167766161263,
0.0606539212167263,
0.048857808113098145,
0.06616874784231186,
0.07415951788425446,
-0.0016658171080052853,
-0.06473805755376816,
0.033455971628427505,
-0.02701282873749733,
0.04121154919266701,
-0.0020991757046431303,
-0.07305153459310532,
-0.033591195940971375,
0.08202052116394043,
0.05402597412467003,
-0.017639491707086563,
0.004734426271170378,
0.0707380548119545,
-0.10068745911121368,
0.06493677198886871,
-0.008887282572686672,
-0.004167200066149235,
-0.03204070031642914,
-0.08367042243480682,
-0.051368363201618195,
0.05236845463514328,
0.036671239882707596,
0.01118291076272726,
-0.054562922567129135,
-0.05366004258394241,
-0.003929546102881432,
0.0027971044182777405,
0.056289467960596085,
0.013054978102445602,
0.03618333488702774,
-0.08081082254648209,
0.005795060191303492,
-0.062254466116428375,
-0.006078836973756552,
0.030393525958061218,
0.08365730941295624,
0.08146212995052338,
-0.024254044517874718,
-0.05688461288809776,
-0.052110880613327026,
0.032957255840301514,
0.047289714217185974,
0.1136864498257637,
-0.008985188789665699,
-0.03437031805515289,
0.06703504920005798,
0.07864070683717728,
-0.012752777896821499,
-0.10431670397520065,
-0.07058552652597427,
-0.08401205390691757,
-0.06319458782672882,
-0.06158442422747612,
0.02350376546382904,
0.03786742314696312,
-0.033402372151613235,
0.011021245270967484,
-0.05270810052752495,
0.04855877533555031,
-0.0023351022973656654,
0.00447869161143899,
0.006829443853348494,
-0.014169538393616676,
0.12718021869659424,
0.08867410570383072,
-0.026175126433372498,
-0.025527961552143097,
-0.08074900507926941,
-0.054935272783041,
0.024652335792779922,
0.039627790451049805,
-0.016256345435976982,
0.02294541895389557,
-0.01615731045603752,
0.013623101636767387,
-0.04187215119600296,
-0.010547542944550514,
-0.0492553636431694,
-0.035693276673555374,
-0.11776860058307648,
-0.06704773753881454,
0.05747184902429581,
-0.03840554133057594,
-0.04529392719268799,
-0.03519605100154877,
-0.01745135709643364,
-0.07260385900735855,
-0.05022648349404335,
0.015960151329636574,
0.038923248648643494,
-1.555723977972088e-33,
-0.0158896092325449,
0.009517325088381767,
0.012909987941384315,
-0.004829649813473225,
0.037232909351587296,
-0.08002676069736481,
-0.017890363931655884,
-0.021928653120994568,
-0.012160019017755985,
-0.055210672318935394,
-0.1231430396437645,
-0.07977237552404404,
0.000878192251548171,
-0.04230434074997902,
0.014777295291423798,
0.06919535994529724,
-0.0008218798320740461,
0.017053280025720596,
-0.07542114704847336,
-0.01119100023061037,
0.043382711708545685,
0.09935164451599121,
-0.012846802361309528,
0.027223946526646614,
-0.01921946555376053,
0.10951171070337296,
0.03116368129849434,
0.06464382261037827,
-0.039477039128541946,
0.041867945343256,
-0.056297946721315384,
-0.004686641972512007,
0.004559598863124847,
-0.00039750186260789633,
-0.10144277662038803,
-0.010300615802407265,
-0.027972007170319557,
0.025098497048020363,
-0.031583625823259354,
0.06183329597115517,
0.033181410282850266,
0.021949920803308487,
-0.07953986525535583,
0.042691223323345184,
0.025505853816866875,
-0.019288619980216026,
-0.03373027965426445,
0.049988601356744766,
0.01758044958114624,
0.1040678322315216,
-0.10821172595024109,
0.06760508567094803,
-0.04461127147078514,
-0.03638787195086479,
-0.03249019756913185,
-0.09713371098041534,
-0.12459813058376312,
-0.02725096605718136,
0.1164894551038742,
0.00601211516186595,
-0.054673902690410614,
-0.04124912619590759,
0.02036552131175995,
0.04065358638763428,
0.00993809662759304,
0.022606050595641136,
0.000701495970133692,
-0.027817681431770325,
-0.02840409427881241,
0.08505292981863022,
0.0182832982391119,
-0.02900538593530655,
0.06261540949344635,
0.02098444662988186,
-0.0010639827232807875,
-0.05695468187332153,
0.035431474447250366,
0.026991313323378563,
0.0073360661044716835,
-0.010460336692631245,
0.04274068400263786,
-0.05942029878497124,
0.05393458157777786,
0.12057580053806305,
0.04462900012731552,
0.08461690694093704,
-0.021571828052401543,
0.0025412424001842737,
0.005256409291177988,
0.05742303654551506,
-0.025555577129125595,
0.04366942122578621,
-0.009993546642363071,
0.05465121939778328,
-0.008116349577903748,
-4.1279051288256596e-8,
-0.0861419215798378,
0.015924546867609024,
-0.056998785585165024,
0.015577680431306362,
-0.05635344609618187,
0.011989731341600418,
-0.0026038517244160175,
0.04264136403799057,
-0.04894765466451645,
-0.07166504859924316,
0.08051183819770813,
-0.06541399657726288,
-0.011925296857953072,
-0.07452216744422913,
-0.0010557412169873714,
0.08122086524963379,
-0.026121877133846283,
0.037299297749996185,
-0.046419981867074966,
0.015344323590397835,
0.015027280896902084,
0.05317111685872078,
-0.007329848594963551,
0.016018422320485115,
-0.002213288564234972,
-0.023388100787997246,
-0.07864145934581757,
0.02872488833963871,
0.06965859979391098,
0.021054986864328384,
0.026589620858430862,
-0.024072783067822456,
0.06807348877191544,
0.061237093061208725,
0.006525334902107716,
-0.027996251359581947,
-0.04705744609236717,
-0.018173370510339737,
0.030364004895091057,
-0.03429032862186432,
0.002687920117750764,
0.13686349987983704,
0.006817937362939119,
0.027177032083272934,
0.0023593285586684942,
0.06974416226148605,
0.02998538687825203,
-0.04040808975696564,
0.032300904393196106,
-0.07663697004318237,
-0.03450286388397217,
-0.09325708448886871,
0.007504328154027462,
0.03675884008407593,
0.016949165612459183,
0.01747915707528591,
-0.04207266494631767,
-0.005270820111036301,
0.07288657873868942,
0.014081240631639957,
-0.030049249529838562,
0.00010162511898670346,
-0.053677719086408615,
-0.03942522034049034
] |
Aria-MIDI: A Dataset of Piano MIDI Files for Symbolic Music Modeling | https://openreview.net/forum?id=X5hrhgndxW | [
"Louis Bradshaw",
"Simon Colton"
] | Poster | We introduce an extensive new dataset of MIDI files, created by transcribing audio recordings of piano performances into their constituent notes. The data pipeline we use is multi-stage, employing a language model to autonomously crawl and score audio recordings from the internet based on their metadata, followed by a stage of pruning and segmentation using an audio classifier. The resulting dataset contains over one million distinct MIDI files, comprising roughly 100,000 hours of transcribed audio. We provide an in-depth analysis of our techniques, offering statistical insights, and investigate the content by extracting metadata tags, which we also provide. Dataset available at https://github.com/loubbrad/aria-midi. | music, symbolic music, piano transcription, dataset, midi | We introduce a dataset of MIDI files comprising roughly 100,000 hours of transcribed piano recordings. | 1,313 | null | [
-0.018676863983273506,
-0.0779150128364563,
0.006993126589804888,
-0.05156134068965912,
0.039945535361766815,
0.03992881253361702,
-0.040307991206645966,
-0.0791105404496193,
0.0011392522137612104,
-0.06585099548101425,
-0.029951585456728935,
-0.035135697573423386,
-0.010379183106124401,
-0.01346929743885994,
-0.027401352301239967,
-0.013345330953598022,
-0.024729346856474876,
0.022821376100182533,
-0.027738993987441063,
-0.040775373578071594,
0.02275244891643524,
0.022792097181081772,
-0.003186053829267621,
0.002978922799229622,
0.08743524551391602,
0.11710058897733688,
0.011790795251727104,
0.050959452986717224,
0.011324248276650906,
-0.015300299040973186,
0.02382820099592209,
0.055227700620889664,
0.07480983436107635,
-0.029514439404010773,
-0.01610170118510723,
0.07209903001785278,
-0.0017240408342331648,
-0.05047643184661865,
-0.07562604546546936,
-0.045461174100637436,
0.014005374163389206,
0.03919505700469017,
0.000622933148406446,
0.033904045820236206,
-0.020382722839713097,
-0.008796343579888344,
-0.06306462734937668,
-0.10502835363149643,
-0.06089608743786812,
0.08925147354602814,
-0.09023844450712204,
-0.053529951721429825,
-0.011645794846117496,
0.09424170106649399,
-0.03292089328169823,
0.015444186516106129,
0.09979129582643509,
0.005455960985273123,
0.03258109092712402,
0.014905638992786407,
-0.02047530747950077,
-0.038161132484674454,
-0.008714603260159492,
-0.02937443181872368,
0.0058997212909162045,
0.03198827803134918,
0.015368316322565079,
0.028422579169273376,
0.016460200771689415,
-0.01964545249938965,
-0.07106195390224457,
0.02097146585583687,
0.029646964743733406,
0.09362885355949402,
0.025203276425600052,
-0.006271237973123789,
-0.0363914780318737,
-0.03934063762426376,
-0.027274122461676598,
-0.10707738250494003,
-0.025037668645381927,
-0.04498489201068878,
0.04557681083679199,
0.020673541352152824,
0.031398385763168335,
-0.029720285907387733,
0.029817119240760803,
0.0009154175058938563,
-0.026256320998072624,
0.020735060796141624,
-0.09123675525188446,
0.009198908694088459,
0.06918063014745712,
-0.044177427887916565,
-0.011241348460316658,
0.14103619754314423,
0.03781082108616829,
-0.01565512828528881,
0.09796251356601715,
0.09108784049749374,
-0.01897330768406391,
0.06182253733277321,
0.010687559843063354,
-0.028628505766391754,
-0.016000384464859962,
-0.027579952031373978,
0.04999808222055435,
0.014099112711846828,
0.02521342970430851,
-0.0435049831867218,
0.08207418024539948,
0.010821938514709473,
-0.050663385540246964,
-0.04756142199039459,
0.08960326761007309,
0.006925093941390514,
-0.0128638856112957,
-0.010821587406098843,
0.025715412572026253,
0.11404047161340714,
-0.015278567560017109,
0.019124669954180717,
-0.009696234948933125,
0.09252272546291351,
-0.003885875456035137,
-0.015587316825985909,
-0.10475827008485794,
-2.845070329048039e-34,
0.06601924449205399,
-0.005510745104402304,
0.04642295837402344,
-0.024416964501142502,
0.07396692037582397,
-0.14472462236881256,
-0.06409730762243271,
0.026618370786309242,
0.00312099140137434,
-0.023357979953289032,
0.006004233844578266,
0.02895311452448368,
-0.02864041179418564,
0.02759670652449131,
-0.014181244187057018,
-0.0013470350531861186,
-0.04581747204065323,
0.05007071793079376,
-0.05516144633293152,
0.019708186388015747,
0.03146469593048096,
-0.05773339420557022,
0.046890269964933395,
0.012217020615935326,
0.03339065983891487,
0.06443606317043304,
-0.030065054073929787,
-0.07000415772199631,
0.005813730880618095,
0.01483792532235384,
-0.03979124873876572,
-0.09708744287490845,
0.02981417439877987,
-0.009916605427861214,
0.01760922186076641,
0.0020083195995539427,
-0.057061247527599335,
0.03162730485200882,
-0.01877710223197937,
-0.055481716990470886,
0.0008014730992726982,
-0.04999716952443123,
-0.0035619682166725397,
-0.08003705739974976,
-0.11029179394245148,
-0.0317356139421463,
-0.046338751912117004,
0.013664350844919682,
0.09406813979148865,
-0.020501675084233284,
0.08073852211236954,
0.05160569027066231,
-0.004739630967378616,
0.05086619406938553,
0.003436279483139515,
-0.02718125842511654,
0.01981543004512787,
0.03118337318301201,
-0.027223026379942894,
0.027859361842274666,
0.07436294853687286,
0.017205944284796715,
0.07439804077148438,
0.0037584914825856686,
-0.0014963371213525534,
0.053563814610242844,
0.10345781594514847,
0.017930056899785995,
0.11093088984489441,
-0.0018146068323403597,
-0.10643885284662247,
-0.02167273871600628,
0.030572451651096344,
-0.01845969818532467,
-0.010243477299809456,
-0.015342948026955128,
-0.00921478122472763,
-0.08267306536436081,
-0.049557775259017944,
0.006992954760789871,
-0.031772464513778687,
0.02905968762934208,
0.02227126993238926,
-0.0328357107937336,
-0.04051821678876877,
0.021593008190393448,
0.02668740600347519,
-0.1561528742313385,
-0.074895940721035,
-0.020834913477301598,
-0.09144442528486252,
0.026120170950889587,
-0.12187836319208145,
-0.050572946667671204,
0.020453225821256638,
-7.072680890071442e-34,
0.029329977929592133,
0.07497315108776093,
0.0444345623254776,
0.013320252299308777,
-0.05592398717999458,
-0.008652218617498875,
-0.025032008066773415,
0.0818287655711174,
0.019319666549563408,
0.07311055064201355,
0.010301603935658932,
-0.05687364935874939,
0.013641765341162682,
-0.12491001933813095,
-0.03178227320313454,
-0.027870075777173042,
0.010395989753305912,
0.01977154240012169,
0.0630350410938263,
0.021424271166324615,
-0.05985627323389053,
0.028477298095822334,
-0.008929074741899967,
0.0022994314786046743,
-0.025553202256560326,
0.019597146660089493,
-0.024177003651857376,
0.014449539594352245,
-0.017712188884615898,
-0.04774754121899605,
-0.017970938235521317,
-0.07385503500699997,
-0.07731722295284271,
-0.1267564743757248,
-0.07474230229854584,
-0.01957845501601696,
0.023787910118699074,
0.02090439200401306,
0.022971412166953087,
0.06345462054014206,
0.0514339841902256,
0.0369500108063221,
-0.036722976714372635,
-0.020225806161761284,
-0.051271311938762665,
-0.06229793280363083,
-0.0699491947889328,
0.17554078996181488,
0.010125400498509407,
-0.04549609497189522,
0.09151589125394821,
-0.01233756821602583,
0.0034015735145658255,
-0.1013081967830658,
-0.02432132512331009,
-0.029174048453569412,
-0.049079280346632004,
-0.0254177488386631,
-0.06380610167980194,
0.06623147428035736,
-0.046548496931791306,
-0.008161481469869614,
-0.032486237585544586,
-0.015100179240107536,
0.04980723187327385,
-0.009469058364629745,
0.004099112004041672,
-0.05377517268061638,
-0.059710726141929626,
0.0491248294711113,
0.11808853596448898,
0.01913616992533207,
-0.03241061046719551,
0.1102842316031456,
-0.06194652244448662,
0.007764076814055443,
-0.010813956148922443,
-0.023096274584531784,
-0.027739645913243294,
-0.03870667889714241,
0.0212208554148674,
0.005307694431394339,
0.0640825554728508,
0.008566714823246002,
0.06285542994737625,
0.08473919332027435,
0.043524179607629776,
0.012634375132620335,
-0.019253799691796303,
0.03158712014555931,
0.020554477348923683,
0.08660680055618286,
0.004637277685105801,
0.04201027378439903,
-0.005671264138072729,
-3.0359078806441175e-8,
-0.05657074227929115,
0.01968121901154518,
-0.0683591365814209,
0.006651122123003006,
-0.015080930665135384,
0.058725081384181976,
0.01958601176738739,
0.013057343661785126,
-0.017633939161896706,
-0.02448355033993721,
0.12536706030368805,
-0.07534743845462799,
-0.03166918829083443,
0.016269205138087273,
-0.10178633779287338,
0.06484222412109375,
0.03694740682840347,
0.12122031301259995,
-0.07609917223453522,
0.01174221746623516,
0.051980867981910706,
0.08726634085178375,
0.004974969197064638,
-0.05467037484049797,
0.0572151243686676,
-0.025866109877824783,
0.0025864841882139444,
0.11567264795303345,
0.00919194146990776,
-0.014069310389459133,
-0.023186733946204185,
0.021970689296722412,
-0.0028921125922352076,
-0.05566960200667381,
0.054907944053411484,
0.05647112429141998,
-0.03861680626869202,
-0.038994818925857544,
-0.0960494726896286,
-0.01441345177590847,
0.054808709770441055,
0.05747781693935394,
-0.07283470034599304,
-0.006459662225097418,
0.002512795152142644,
-0.05030908063054085,
0.015195908024907112,
-0.020113900303840637,
0.0701054185628891,
0.04046769440174103,
-0.016181660816073418,
-0.03940342366695404,
-0.015589150600135326,
0.04072944074869156,
0.018893780186772346,
0.06486274302005768,
-0.04171266779303551,
0.08973268419504166,
0.03646010905504227,
0.000685201259329915,
-0.007581669837236404,
-0.016220059245824814,
0.03607935085892677,
-0.006591417361050844
] |
Efficient and Context-Aware Label Propagation for Zero-/Few-Shot Training-Free Adaptation of Vision-Language Model | https://openreview.net/forum?id=D10yarGQNk | [
"Yushu Li",
"Yongyi Su",
"Adam Goodge",
"Kui Jia",
"Xun Xu"
] | Poster | Vision-language models (VLMs) have revolutionized machine learning by leveraging large pre-trained models to tackle various downstream tasks. Although label, training, and data efficiency have improved, many state-of-the-art VLMs still require task-specific hyperparameter tuning and fail to fully exploit test samples. To overcome these challenges, we propose a graph-based approach for label-efficient adaptation and inference. Our method dynamically constructs a graph over text prompts, few-shot examples, and test samples, using label propagation for inference without task-specific tuning. Unlike existing zero-shot label propagation techniques, our approach requires no additional unlabeled support set and effectively leverages the test sample manifold through dynamic graph expansion. We further introduce a context-aware feature re-weighting mechanism to improve task adaptation accuracy. Additionally, our method supports efficient graph expansion, enabling real-time inductive inference. Extensive evaluations on downstream tasks, such as fine-grained categorization and out-of-distribution generalization, demonstrate the effectiveness of our approach. The source code is available at https://github.com/Yushu-Li/ECALP. | Vision-Language Model, Label Propagation, Training-Free | A graph-based method for VLM label-efficient adaptation, enabling dynamic graph construction with computationally efficient label propagation and context-aware feature re-weighting without task-specific tuning. | 1,297 | null | [
0.012797834351658821,
-0.0440613329410553,
-0.0010561638046056032,
0.06639526039361954,
0.12791742384433746,
0.03458067402243614,
-0.010007631964981556,
-0.04337159916758537,
-0.04800689220428467,
-0.07056701183319092,
-0.005835007410496473,
-0.11398860812187195,
0.03019758313894272,
0.10002531111240387,
0.042670540511608124,
-0.02164563164114952,
0.17323242127895355,
0.06038646027445793,
-0.06419568508863449,
-0.013373537920415401,
0.00766996294260025,
0.011340641416609287,
0.09185779839754105,
0.045143090188503265,
0.080462746322155,
-0.021608896553516388,
-0.038515664637088776,
-0.06327206641435623,
0.04097701236605644,
-0.01678052544593811,
0.023387806490063667,
-0.017094960436224937,
-0.11562635749578476,
0.04343706741929054,
0.009168094955384731,
-0.0014793650479987264,
-0.009837874211370945,
0.07817164808511734,
-0.002563272602856159,
0.037061404436826706,
0.010379940271377563,
-0.012071039527654648,
0.013596700504422188,
0.03773980587720871,
0.0348428450524807,
0.06411004811525345,
-0.016322031617164612,
-0.01734626479446888,
-0.017783304676413536,
0.057015560567379,
-0.10723777115345001,
-0.03601939603686333,
-0.0335281603038311,
0.012013361789286137,
-0.06337519735097885,
0.02250857464969158,
0.08391546458005905,
0.0016873219283297658,
0.054805509746074677,
0.03359004482626915,
-0.0066561149433255196,
-0.11461017280817032,
-0.002746481914073229,
0.03563053160905838,
-0.07094241678714752,
0.029867352917790413,
0.010454838164150715,
0.0011487092124298215,
-0.02619926445186138,
-0.0041998038068413734,
-0.008625500835478306,
0.10269353538751602,
0.022417327389121056,
0.0738738551735878,
0.036273472011089325,
0.039177779108285904,
0.016482872888445854,
-0.02681911364197731,
0.07476311177015305,
-0.052057910710573196,
0.038037482649087906,
0.02668098174035549,
-0.0004269591881893575,
-0.03280467540025711,
0.1128094494342804,
0.06236184388399124,
0.006472447421401739,
-0.006768220569938421,
-0.024397823959589005,
0.006546972785145044,
-0.09843284636735916,
-0.07876990735530853,
0.004217989277094603,
-0.008384998887777328,
0.028345055878162384,
0.0003304348501842469,
-0.024600611999630928,
-0.09224194288253784,
-0.04371771588921547,
0.05033893883228302,
-0.019079694524407387,
-0.007073154207319021,
0.038485653698444366,
-0.0668429359793663,
-0.0009282981627620757,
0.04053938761353493,
0.029346352443099022,
0.045304521918296814,
0.0020983582362532616,
-0.07663577049970627,
0.06584989279508591,
0.06467383354902267,
0.009811047464609146,
-0.02393786609172821,
0.001333874766714871,
-0.03834386169910431,
-0.004994161892682314,
-0.07583863288164139,
-0.0055722640827298164,
0.0430525466799736,
-0.03712764382362366,
0.03286183997988701,
-0.04656105488538742,
-0.046036090701818466,
-0.009340006858110428,
0.014752025716006756,
0.008726492524147034,
2.521969635512706e-34,
0.08531161397695541,
0.06275999546051025,
0.10564222186803818,
0.025375720113515854,
0.05666939914226532,
-0.0639604777097702,
-0.005548418499529362,
0.04589497298002243,
-0.03756284713745117,
-0.03405484929680824,
0.005951406434178352,
-0.039406802505254745,
-0.041117776185274124,
0.07300321757793427,
0.016638638451695442,
-0.02067265659570694,
-0.03046286292374134,
0.03110875003039837,
0.01800827868282795,
0.028739996254444122,
-0.06403286755084991,
-0.03403784707188606,
-0.008808105252683163,
-0.04453481361269951,
0.02800082042813301,
-0.0023026899434626102,
0.034110795706510544,
-0.023504704236984253,
-0.048836007714271545,
0.007766853552311659,
-0.11600648611783981,
0.005303021986037493,
0.04153234884142876,
0.025368323549628258,
0.025936655700206757,
-0.06651216745376587,
-0.022375188767910004,
-0.037690669298172,
0.03614605963230133,
-0.008337845094501972,
0.0036668777465820312,
0.04742201417684555,
0.043959468603134155,
-0.03612353280186653,
-0.07838141173124313,
-0.06938409060239792,
-0.05175202339887619,
0.0489426925778389,
-0.06495114415884018,
-0.025440938770771027,
0.05044938996434212,
0.013918345794081688,
-0.011546061374247074,
-0.005022334400564432,
-0.04272458329796791,
-0.004855899605900049,
0.005208443384617567,
0.029544338583946228,
0.003986237570643425,
-0.0008967149769887328,
0.04993922635912895,
0.0019349679350852966,
-0.05061763897538185,
0.12350593507289886,
0.05659076198935509,
0.03272535279393196,
-0.018992578610777855,
0.015834394842386246,
0.031803153455257416,
-0.01278422586619854,
-0.01551285944879055,
0.008359585888683796,
-0.051895081996917725,
-0.034758713096380234,
0.0685090646147728,
0.03353087231516838,
0.019057875499129295,
-0.14088264107704163,
0.011866741813719273,
0.09912147372961044,
-0.004762770142406225,
-0.018858077004551888,
0.0570976696908474,
0.001587124657817185,
-0.00854051485657692,
-0.055899281054735184,
0.04890584573149681,
-0.11711128801107407,
0.029586387798190117,
-0.10496819764375687,
-0.04026876762509346,
-0.05730262026190758,
-0.05745663866400719,
-0.008255415596067905,
0.0722825676202774,
1.5535190077093345e-33,
0.019228655844926834,
0.06674445420503616,
0.022570021450519562,
0.05283159017562866,
0.010041714645922184,
0.010689307004213333,
0.09455260634422302,
0.05078065022826195,
0.008743906393647194,
-0.013758820481598377,
-0.06369578093290329,
-0.07387527078390121,
-0.03120819479227066,
-0.019060000777244568,
-0.0019228694727644324,
0.007901339791715145,
-0.05719766765832901,
0.035923201590776443,
-0.03739127144217491,
0.03169121593236923,
0.01729612424969673,
0.11639617383480072,
-0.05148950591683388,
-0.0008939451654441655,
0.00744338845834136,
0.03773130849003792,
-0.024920936673879623,
0.0512545183300972,
0.007971788756549358,
-0.021244298666715622,
0.0013161561219021678,
0.028363803401589394,
0.04100235551595688,
-0.06194191426038742,
-0.015453079715371132,
0.08913824707269669,
0.04866829141974449,
-0.03194234520196915,
0.002562464913353324,
0.1277376413345337,
0.04762628301978111,
-0.012451497837901115,
-0.0748547688126564,
0.0000940793615882285,
-0.04145772010087967,
0.03184576332569122,
-0.04016381502151489,
-0.041618842631578445,
-0.05602356418967247,
0.049111805856227875,
-0.11850226670503616,
0.06601713597774506,
-0.08954474329948425,
0.059623174369335175,
-0.08138857781887054,
-0.14190910756587982,
-0.007815580815076828,
0.01553119532763958,
0.050598930567502975,
-0.06235203891992569,
-0.0772080048918724,
0.04031296446919441,
0.01798870414495468,
-0.018676355481147766,
0.044836919754743576,
-0.05391650274395943,
-0.012402944266796112,
-0.030560901388525963,
0.030712604522705078,
-0.005297681782394648,
0.06704604625701904,
-0.032219164073467255,
-0.044432688504457474,
-0.04230055585503578,
-0.029711520299315453,
-0.029261576011776924,
-0.017696725204586983,
-0.02838071994483471,
-0.027048734948039055,
-0.04581918567419052,
0.06692694872617722,
-0.06652788817882538,
-0.017617251724004745,
0.0774424746632576,
0.13558483123779297,
0.11953160166740417,
0.08595003932714462,
0.06652718037366867,
0.04975612461566925,
0.06658728420734406,
-0.045655060559511185,
-0.010382529348134995,
-0.010658584535121918,
0.07363733649253845,
-0.016741447150707245,
-3.892506938996121e-8,
-0.12409889698028564,
0.08115445077419281,
-0.040132809430360794,
0.07146816700696945,
-0.02959533780813217,
0.03387288749217987,
-0.008888030424714088,
0.12344839423894882,
0.005544361658394337,
-0.002234645653516054,
0.021997369825839996,
0.01871023327112198,
0.03594048321247101,
-0.04139936342835426,
-0.05484055355191231,
0.060075536370277405,
-0.0023041232489049435,
0.06748010218143463,
0.029709694907069206,
0.03779299557209015,
0.002086015185341239,
0.013725167140364647,
0.03870707377791405,
0.0017850914737209678,
-0.012038638815283775,
-0.11507663875818253,
0.02430405467748642,
0.07081804424524307,
0.10683777928352356,
-0.09287740290164948,
-0.03115007095038891,
0.041115980595350266,
-0.007933163084089756,
-0.04098661243915558,
-0.01430311519652605,
0.06236035376787186,
-0.0005259386380203068,
0.005942957475781441,
0.0012192836729809642,
-0.0020167508628219366,
-0.04136419668793678,
0.09250260889530182,
-0.026596838608384132,
0.0047824871726334095,
-0.023916540667414665,
0.015677211806178093,
-0.0015413761138916016,
-0.07236552238464355,
-0.009452802129089832,
-0.07411711663007736,
-0.003346596146002412,
-0.02401079051196575,
-0.053342048078775406,
0.008546773344278336,
0.0026754627469927073,
0.012054402381181717,
-0.007952815853059292,
-0.019593743607401848,
0.023445649072527885,
0.08966542035341263,
-0.002000385895371437,
0.06788540631532669,
-0.052331309765577316,
-0.03585796803236008
] |
Methods with Local Steps and Random Reshuffling for Generally Smooth Non-Convex Federated Optimization | https://openreview.net/forum?id=TrJ36UfD9P | [
"Yury Demidovich",
"Petr Ostroukhov",
"Grigory Malinovsky",
"Samuel Horváth",
"Martin Takáč",
"Peter Richtárik",
"Eduard Gorbunov"
] | Poster | Non-convex Machine Learning problems typically do not adhere to the standard smoothness assumption. Based on empirical findings, Zhang et al. (2020b) proposed a more realistic generalized $(L_0,L_1)$-smoothness assumption, though it remains largely unexplored. Many existing algorithms designed for standard smooth problems need to be revised. However, in the context of Federated Learning, only a few works address this problem but rely on additional limiting assumptions. In this paper, we address this gap in the literature: we propose and analyze new methods with local steps, partial participation of clients, and Random Reshuffling without extra restrictive assumptions beyond generalized smoothness. The proposed methods are based on the proper interplay between clients' and server's stepsizes and gradient clipping. Furthermore, we perform the first analysis of these methods under the Polyak-Łojasiewicz condition. Our theory is consistent with the known results for standard smooth problems, and our experimental results support the theoretical insights. | Optimization, Federated Learning, Distributed Optimization, Local Training, Random Reshuffling, Generalized Smoothness | null | 1,294 | 2412.02781 | [
-0.11227274686098099,
-0.05748353525996208,
0.06310167908668518,
0.029377305880188942,
0.03913946822285652,
-0.058769647032022476,
-0.0005589656066149473,
-0.10709429532289505,
-0.02277316339313984,
0.009553448297083378,
-0.03585211932659149,
0.10313025116920471,
0.032196350395679474,
-0.03319554403424263,
-0.004371969029307365,
0.015672501176595688,
0.07032573223114014,
0.03098146617412567,
-0.017125051468610764,
0.0008173180394805968,
-0.07392824441194534,
-0.02033781073987484,
-0.04576912149786949,
0.0050539602525532246,
0.05347198620438576,
-0.08127385377883911,
0.04020925238728523,
-0.0240719486027956,
0.013337763026356697,
0.021046951413154602,
0.030467916280031204,
-0.007524525746703148,
-0.07224543392658234,
-0.06378551572561264,
0.03421799838542938,
0.08168338984251022,
-0.0451795868575573,
0.09576823562383652,
0.004132227040827274,
0.042922236025333405,
0.017241142690181732,
-0.013600323349237442,
0.004988860804587603,
0.0018954372499138117,
0.09660661965608597,
-0.03204749524593353,
-0.009239214472472668,
-0.06649861484766006,
-0.062806136906147,
0.034102704375982285,
-0.03406880050897598,
-0.06835352629423141,
-0.06342232972383499,
0.07181167602539062,
-0.04394666850566864,
-0.015492398291826248,
-0.011094004847109318,
0.030226411297917366,
-0.006707561202347279,
0.08276695758104324,
-0.04512939602136612,
-0.12476281821727753,
-0.03682718053460121,
0.03609388321638107,
-0.002988896332681179,
-0.012115049175918102,
0.033193040639162064,
0.021386459469795227,
-0.025308430194854736,
0.07288485020399094,
-0.02277214638888836,
0.0346536748111248,
-0.07606518268585205,
-0.01118544489145279,
-0.012897680513560772,
0.004809449892491102,
0.06092654913663864,
0.06312306225299835,
0.006197568494826555,
-0.05883273482322693,
-0.054536063224077225,
0.003489695256575942,
0.029926100745797157,
-0.01079332921653986,
0.012033019214868546,
-0.055087752640247345,
-0.033082012087106705,
0.0028141604270786047,
0.029426638036966324,
-0.03003603033721447,
-0.038477782160043716,
0.10303692519664764,
-0.04851146042346954,
-0.016583152115345,
-0.00694658700376749,
-0.020365118980407715,
0.062193576246500015,
-0.027135126292705536,
0.05940092355012894,
0.14253084361553192,
-0.05845999717712402,
0.08934783935546875,
0.04600990191102028,
-0.05604377016425133,
0.057225555181503296,
-0.018494175747036934,
0.0069632939994335175,
0.03171286731958389,
0.04691918566823006,
-0.06708376854658127,
-0.013354649767279625,
0.04266637563705444,
0.07207996398210526,
0.011319906450808048,
-0.010034365579485893,
-0.02806748077273369,
0.04659400135278702,
-0.07123148441314697,
-0.04094194620847702,
0.05508799105882645,
-0.02840292640030384,
-0.010872808285057545,
0.06963218748569489,
0.02239159494638443,
-0.01446172408759594,
0.03986543044447899,
-0.007365439552813768,
3.6651524476135215e-33,
0.0069804927334189415,
-0.03924969583749771,
0.005432162433862686,
-0.09155341982841492,
0.08622914552688599,
-0.00584313552826643,
0.051797375082969666,
-0.02292528748512268,
-0.025056764483451843,
-0.018425514921545982,
0.03464868664741516,
-0.012327178381383419,
-0.006028379313647747,
0.04237908869981766,
0.01605329103767872,
-0.06074124574661255,
0.03136875480413437,
0.08607591688632965,
-0.004713116679340601,
-0.10530586540699005,
0.047384392470121384,
-0.06921162456274033,
-0.01587115228176117,
-0.006809623446315527,
-0.02664656564593315,
-0.053198717534542084,
0.029445059597492218,
0.05988201126456261,
0.018738429993391037,
0.03977908939123154,
-0.03059014305472374,
-0.005921002943068743,
-0.04374518245458603,
0.062390465289354324,
-0.016740772873163223,
0.034422826021909714,
0.010957048274576664,
-0.0478539764881134,
-0.027122922241687775,
0.04865771904587746,
0.010564452037215233,
0.04278819262981415,
0.04661745950579643,
-0.07528265565633774,
-0.16925616562366486,
-0.07150088995695114,
-0.06243468075990677,
0.01940891332924366,
-0.06186515465378761,
-0.041510406881570816,
0.012928562238812447,
0.03466339036822319,
0.021237606182694435,
-0.04363764822483063,
-0.06884991377592087,
0.009860238060355186,
-0.0033113923855125904,
0.03986227884888649,
0.021247228607535362,
0.08113723993301392,
0.004865764640271664,
-0.004633249714970589,
-0.03234711289405823,
0.03309362754225731,
-0.07200593501329422,
0.07273553311824799,
0.028433658182621002,
0.033129241317510605,
0.06822492182254791,
-0.1055474802851677,
0.06165652349591255,
0.0618537999689579,
0.006698458921164274,
-0.05170692130923271,
0.08531715720891953,
-0.05210302397608757,
0.01478082500398159,
-0.019283398985862732,
-0.0006153375725261867,
-0.040201686322689056,
-0.018620245158672333,
0.07000463455915451,
-0.06427527964115143,
-0.05847766622900963,
-0.03659842908382416,
-0.0719062089920044,
0.031464170664548874,
-0.027769746258854866,
-0.06727125495672226,
0.003932532388716936,
-0.16564111411571503,
0.03610891103744507,
0.0007967691635712981,
0.04560133442282677,
0.05488361045718193,
-3.0401299789958183e-33,
-0.030205348506569862,
-0.0045514474622905254,
-0.053581055253744125,
0.04882272705435753,
0.0359487421810627,
0.012632180005311966,
-0.05582384392619133,
-0.01731078140437603,
0.010185606777668,
-0.039956316351890564,
-0.03554455563426018,
-0.010271917097270489,
0.03703412786126137,
-0.010618908330798149,
-0.05989564210176468,
-0.0027317609637975693,
0.021570978686213493,
0.03653559461236,
0.006401965394616127,
0.02780221775174141,
0.024686938151717186,
-0.009418126195669174,
-0.0843367874622345,
0.02767058275640011,
0.06058801710605621,
0.05483390390872955,
-0.13460494577884674,
0.16469977796077728,
-0.04948925971984863,
-0.05281636491417885,
-0.025683879852294922,
-0.0462123304605484,
-0.1106148362159729,
-0.09897810220718384,
-0.01965910755097866,
0.0874190628528595,
-0.028914282098412514,
0.087929368019104,
-0.014135854318737984,
0.09692186117172241,
0.08311010152101517,
-0.05844249948859215,
-0.06969776004552841,
0.06267594546079636,
0.07738832384347916,
-0.007457837462425232,
-0.012371784076094627,
-0.024713866412639618,
-0.011255903169512749,
-0.012975105084478855,
-0.05980880558490753,
0.05160931125283241,
-0.042848896235227585,
0.10474557429552078,
-0.07228517532348633,
0.03096037544310093,
0.014884102158248425,
-0.0016639461973682046,
-0.0061684721149504185,
-0.015976212918758392,
-0.011680779978632927,
-0.029283354058861732,
-0.07163111120462418,
0.0702374130487442,
0.07816272974014282,
-0.021901650354266167,
0.05479760840535164,
0.019472895190119743,
-0.007720643188804388,
0.03216679021716118,
-0.028279397636651993,
0.054396502673625946,
-0.016964351758360863,
0.01551812794059515,
0.029432062059640884,
0.037770356982946396,
0.10055191814899445,
-0.08627977222204208,
-0.027015702798962593,
0.015249144285917282,
0.08528167754411697,
0.0032408658880740404,
0.065803661942482,
0.004906395450234413,
0.0428411029279232,
-0.002402552403509617,
0.03640712797641754,
0.05135999992489815,
0.039229754358530045,
-0.06833438575267792,
0.01170576922595501,
-0.03384006395936012,
-0.023288415744900703,
0.03632243722677231,
-0.005099182948470116,
-4.702902245412588e-8,
-0.05008922517299652,
0.008992289192974567,
-0.033573295921087265,
0.007010117173194885,
0.09833003580570221,
0.008943810127675533,
-0.03587254509329796,
0.046596914529800415,
-0.040983621031045914,
-0.06465647369623184,
0.05162996053695679,
-0.045306943356990814,
0.05385098233819008,
-0.06808898597955704,
-0.04864251986145973,
0.0634537935256958,
-0.04913697391748428,
0.03975145146250725,
-0.03391391411423683,
0.05127551034092903,
-0.050452135503292084,
0.042604729533195496,
0.026367060840129852,
-0.041670188307762146,
0.026232047006487846,
-0.04327133670449257,
0.06368178129196167,
0.03568759933114052,
0.0693255066871643,
0.009461532346904278,
-0.15218167006969452,
0.09191515296697617,
0.021555908024311066,
0.030772432684898376,
0.026562843471765518,
0.06976324319839478,
0.003930686041712761,
-0.008058331906795502,
-0.039153091609478,
-0.002112187212333083,
0.024866951629519463,
0.049054622650146484,
-0.028669612482190132,
-0.023836683481931686,
0.04895678907632828,
-0.04730573296546936,
-0.0271356962621212,
-0.03353147581219673,
0.03350852429866791,
-0.028131617233157158,
0.03124958649277687,
-0.023541530594229698,
0.04378025233745575,
0.06350066512823105,
0.11653303354978561,
-0.031887639313936234,
0.014728995971381664,
-0.03331148251891136,
0.050595298409461975,
0.0894186869263649,
0.020179444923996925,
0.036910805851221085,
-0.056943152099847794,
-0.0014752361457794905
] |
From Risk to Uncertainty: Generating Predictive Uncertainty Measures via Bayesian Estimation | https://openreview.net/forum?id=cWfpt2t37q | [
"Nikita Kotelevskii",
"Vladimir Kondratyev",
"Martin Takáč",
"Eric Moulines",
"Maxim Panov"
] | Poster | There are various measures of predictive uncertainty in the literature, but their relationships to each other remain unclear. This paper uses a decomposition of statistical pointwise risk into components associated with different sources of predictive uncertainty: namely, aleatoric uncertainty (inherent data variability) and epistemic uncertainty (model-related uncertainty). Together with Bayesian methods applied as approximations, we build a framework that allows one to generate different predictive uncertainty measures.
We validate measures, derived from our framework on image datasets by evaluating its performance in detecting out-of-distribution and misclassified instances using the AUROC metric. The experimental results confirm that the measures derived from our framework are useful for the considered downstream tasks. | Uncertainty quantification, Bayesian methods, Statistics | We show, how one can get predictive uncertainty measures using Bayesian approximations of statistical risk. | 1,292 | 2402.10727 | [
-0.0008531646453775465,
-0.06464817374944687,
0.0046169026754796505,
0.03601732477545738,
0.0766090527176857,
-0.01664024218916893,
0.052032649517059326,
-0.0012856071116402745,
0.05229898542165756,
-0.00741258729249239,
-0.0009260665974579751,
-0.1377338320016861,
0.09783997386693954,
0.033863455057144165,
-0.11053745448589325,
-0.05977281928062439,
0.03376764431595802,
0.02021303027868271,
-0.05312935262918472,
0.021108334884047508,
-0.011111048050224781,
0.037293437868356705,
-0.008621079847216606,
0.012400205247104168,
0.08745106309652328,
0.00856960192322731,
0.033402133733034134,
0.02276909537613392,
-0.03439090773463249,
-0.04800837114453316,
0.047065578401088715,
-0.027182796970009804,
-0.018616443499922752,
-0.09124600887298584,
0.03235657513141632,
0.026883326470851898,
-0.04157250002026558,
0.05333709344267845,
0.06878835707902908,
-0.01808713935315609,
0.004027745686471462,
0.00043344483128748834,
0.036071352660655975,
0.07160531729459763,
-0.012131438590586185,
0.0485563725233078,
0.007686320226639509,
-0.10690350085496902,
-0.004756442271173,
0.022943241521716118,
-0.05712779611349106,
-0.023094674572348595,
-0.026512976735830307,
-0.019956735894083977,
-0.07451985031366348,
0.06217259168624878,
0.03493497148156166,
-0.07237010449171066,
0.048559896647930145,
-0.004358016885817051,
0.013742289505898952,
-0.09057076275348663,
-0.07322424650192261,
0.008242486976087093,
0.026480939239263535,
0.06283312290906906,
0.02404438890516758,
-0.05542522296309471,
0.05232839286327362,
-0.01841636747121811,
-0.02857353538274765,
0.07270554453134537,
-0.042332179844379425,
0.04342129826545715,
-0.04030660167336464,
0.017085175961256027,
0.06499118357896805,
0.002762961434200406,
-0.012045730836689472,
-0.055032793432474136,
0.015050038695335388,
0.054103754460811615,
0.013986684381961823,
-0.009696437045931816,
0.15463761985301971,
0.039756521582603455,
0.057252101600170135,
0.05651087313890457,
-0.013827919028699398,
0.015331398695707321,
-0.09166774153709412,
-0.04341386631131172,
0.05852869153022766,
0.003645160235464573,
-0.025938306003808975,
0.030931225046515465,
-0.011480924673378468,
-0.11753162741661072,
0.05462426692247391,
0.029601121321320534,
-0.016236193478107452,
0.04114158824086189,
0.037745121866464615,
0.013311917893588543,
0.03485230356454849,
-0.0336269736289978,
-0.010473043657839298,
-0.03851839527487755,
0.06631746143102646,
-0.034893572330474854,
-0.009143469855189323,
-0.0448756217956543,
0.011901576071977615,
-0.04141422361135483,
0.012975740246474743,
-0.02377469465136528,
-0.015249096788465977,
-0.003650849685072899,
-0.0431065708398819,
0.038976892828941345,
-0.02318054810166359,
-0.02827328070998192,
0.08796728402376175,
-0.022257857024669647,
0.04201565310359001,
-0.05111926794052124,
-0.012400481849908829,
1.738094395991481e-33,
0.05558345094323158,
-0.08985607326030731,
0.06771578639745712,
0.043501339852809906,
0.021147338673472404,
0.056801147758960724,
-0.1048121303319931,
-0.06837944686412811,
0.015119154937565327,
-0.019415875896811485,
-0.0024841229896992445,
0.006293745245784521,
-0.09802060574293137,
0.02471579983830452,
0.11972185969352722,
0.07183968275785446,
-0.05757564306259155,
0.05301787331700325,
-0.03194158151745796,
0.0069422051310539246,
0.014024821110069752,
-0.08661540597677231,
-0.06318087875843048,
-0.015741582959890366,
-0.008884929120540619,
0.014926950447261333,
0.06597781181335449,
0.013072168454527855,
-0.018892204388976097,
0.005676296539604664,
-0.08682965487241745,
0.03952023759484291,
0.0748407393693924,
-0.030705951154232025,
0.0008988305344246328,
-0.05061086267232895,
-0.0715547725558281,
0.0036883968859910965,
-0.00738763390108943,
-0.015616101212799549,
0.060383956879377365,
-0.016255894675850868,
0.06630344688892365,
0.042389221489429474,
-0.03462794050574303,
-0.05671916902065277,
0.048753637820482254,
-0.06967761367559433,
-0.03450162708759308,
-0.04005487635731697,
0.02708626165986061,
-0.0368758887052536,
0.019154662266373634,
0.019446879625320435,
-0.05129587650299072,
0.0034238151274621487,
-0.004083407577127218,
0.013870085589587688,
-0.02755936048924923,
-0.06321799755096436,
-0.03784780949354172,
0.03297446295619011,
-0.002048519905656576,
-0.0017828779527917504,
0.021437877789139748,
0.0645081177353859,
0.04066295176744461,
0.027267416939139366,
-0.07004300504922867,
0.05227245017886162,
-0.03655822575092316,
-0.009222235530614853,
-0.0012591611593961716,
-0.061436157673597336,
-0.013754315674304962,
0.03952621668577194,
-0.0018265977269038558,
0.03263018652796745,
0.01579468883574009,
0.023184120655059814,
-0.09207450598478317,
-0.05421299487352371,
0.007117318920791149,
0.016946369782090187,
0.005254976451396942,
0.008018258959054947,
0.055476270616054535,
-0.011220603249967098,
-0.14134082198143005,
0.023438265547156334,
-0.03464451804757118,
-0.007045378442853689,
-0.013009633868932724,
0.05884496122598648,
-0.06142812222242355,
-3.733741440958048e-33,
-0.08752832561731339,
0.0479414202272892,
0.08007204532623291,
0.034121472388505936,
0.01139247976243496,
-0.05523160472512245,
-0.0028636381030082703,
0.07106562703847885,
0.01838897168636322,
-0.07467411458492279,
-0.08973383158445358,
-0.0029242767486721277,
-0.04072410240769386,
-0.06169493496417999,
-0.04890371859073639,
0.01600983925163746,
0.018370574340224266,
-0.008007481694221497,
-0.06908867508172989,
-0.03376789763569832,
0.056768689304590225,
-0.02925516478717327,
-0.010686175897717476,
0.018447713926434517,
-0.01959569938480854,
0.07868658751249313,
0.01397079136222601,
0.02012990415096283,
-0.08408248424530029,
-0.0619235560297966,
-0.056635260581970215,
0.03523048385977745,
0.04701750725507736,
-0.0021423178259283304,
-0.041172776371240616,
-0.009355458430945873,
0.01216100063174963,
0.009248352609574795,
-0.08331099897623062,
0.0448073148727417,
-0.00722668319940567,
0.04461894929409027,
-0.14718246459960938,
-0.08618004620075226,
-0.01831415295600891,
-0.07809145748615265,
0.08408044278621674,
-0.051230307668447495,
0.04692375287413597,
-0.02561620995402336,
-0.021776702255010605,
0.090088851749897,
-0.08583984524011612,
0.11392410099506378,
-0.053825195878744125,
-0.01025377307087183,
-0.03529459610581398,
-0.06910460442304611,
0.07307352125644684,
0.07840330898761749,
-0.03804166615009308,
-0.0011188684729859233,
0.012145495973527431,
0.03730493038892746,
-0.012375925667583942,
0.02116527408361435,
-0.1580982506275177,
-0.00924244150519371,
-0.013581700623035431,
0.0958276242017746,
0.011155172251164913,
-0.03700427711009979,
0.004905879497528076,
-0.025469902902841568,
-0.008634106256067753,
-0.05650314688682556,
0.03945524990558624,
0.007109758444130421,
0.0981510654091835,
0.03457928076386452,
0.058900926262140274,
0.02560466155409813,
0.05729296803474426,
0.07134884595870972,
-0.012207362800836563,
0.11348491162061691,
0.029852576553821564,
-0.05433475598692894,
0.08353148400783539,
0.0006506147910840809,
-0.04956887289881706,
0.028298113495111465,
0.006334549747407436,
0.005900457501411438,
-0.027213837951421738,
-4.3702129914890975e-8,
-0.016083162277936935,
0.08388850092887878,
0.07060667872428894,
-0.04924623668193817,
0.016879750415682793,
0.022172391414642334,
-0.06356576085090637,
0.049967750906944275,
-0.014997050166130066,
-0.08308359980583191,
0.08148008584976196,
-0.04733788222074509,
-0.05605293810367584,
0.007177222985774279,
0.021184150129556656,
0.07521773129701614,
0.06236778199672699,
0.13631197810173035,
0.005311899818480015,
-0.011299731209874153,
0.01951361820101738,
-0.009650848805904388,
0.00046927991206757724,
-0.021877488121390343,
-0.012573308311402798,
-0.05531098321080208,
-0.012830394320189953,
0.12902085483074188,
0.01392645388841629,
0.07418769598007202,
0.036859963089227676,
-0.06047334149479866,
0.03435715287923813,
0.03357269987463951,
-0.016507958993315697,
0.07307366281747818,
0.0077932002022862434,
-0.11867688596248627,
-0.07892807573080063,
0.004476395435631275,
0.008882864378392696,
0.03221319615840912,
-0.11116346716880798,
0.07222740352153778,
0.009612680412828922,
0.04638676345348358,
0.07843918353319168,
0.012531022541224957,
0.036301981657743454,
-0.015382951125502586,
0.030711255967617035,
-0.00010245887096971273,
-0.04465277120471001,
0.0750705674290657,
0.047307103872299194,
-0.02308298461139202,
0.058228351175785065,
-0.10046571493148804,
-0.036400970071554184,
0.015162641182541847,
0.019727610051631927,
-0.024516258388757706,
-0.06896939128637314,
0.015431554056704044
] |
Timer-XL: Long-Context Transformers for Unified Time Series Forecasting | https://openreview.net/forum?id=KMCJXjlDDr | [
"Yong Liu",
"Guo Qin",
"Xiangdong Huang",
"Jianmin Wang",
"Mingsheng Long"
] | Poster | We present Timer-XL, a causal Transformer for unified time series forecasting. To uniformly predict multidimensional time series, we generalize next token prediction, predominantly adopted for 1D token sequences, to multivariate next token prediction. The paradigm formulates various forecasting tasks as a long-context prediction problem. We opt for decoder-only Transformers that capture causal dependencies from varying-length contexts for unified forecasting, making predictions on non-stationary univariate time series, multivariate series with complicated dynamics and correlations, as well as covariate-informed contexts that include exogenous variables. Technically, we propose a universal TimeAttention to capture fine-grained intra- and inter-series dependencies of flattened time series tokens (patches), which is further enhanced by deft position embedding for temporal causality and variable equivalence. Timer-XL achieves state-of-the-art performance across task-specific forecasting benchmarks through a unified approach. Based on large-scale pre-training, Timer-XL achieves state-of-the-art zero-shot performance, making it a promising architecture for pre-trained time series models. Code is available at this repository: https://github.com/thuml/Timer-XL. | Time Series Forecasting, Transformer, Foundation Model | We introduce a generative Transformer for multi-dimensional time series, which achieves SOTA performance in supervised training and zero-shot forecasting. | 1,290 | null | [
-0.12727242708206177,
-0.07587691396474838,
0.05904154106974602,
0.05712286755442619,
0.04980984330177307,
0.009936417452991009,
-0.09495427459478378,
-0.002761213341727853,
0.03808369114995003,
-0.03136969730257988,
-0.1089622974395752,
-0.054046448320150375,
-0.04892769455909729,
0.07671845704317093,
-0.02377510629594326,
-0.03318426012992859,
0.013767401687800884,
0.003672440303489566,
-0.008763429708778858,
-0.0524713508784771,
0.05890689417719841,
0.050430383533239365,
-0.04506372660398483,
0.038247644901275635,
0.05885517597198486,
0.0011296806624159217,
0.011225290596485138,
-0.023168886080384254,
-0.007187935523688793,
0.0560004748404026,
0.013151728548109531,
-0.0001804597704904154,
-0.08912951499223709,
-0.03764182701706886,
-0.09252426773309708,
0.0713183656334877,
0.008360110223293304,
-0.0014348619151860476,
0.03717460483312607,
-0.0448266863822937,
0.08668571710586548,
-0.021700764074921608,
0.01578526571393013,
0.06081373989582062,
-0.010563777759671211,
0.05153767019510269,
0.09630292654037476,
-0.05370195955038071,
-0.036772720515728,
0.014338846318423748,
-0.02023239992558956,
0.034009892493486404,
-0.006124876905232668,
0.008990037254989147,
0.048900555819272995,
0.04580451548099518,
0.009716757573187351,
0.06506539136171341,
0.012803619727492332,
-0.032000359147787094,
-0.09901224821805954,
-0.07099755853414536,
0.002030879957601428,
-0.0019154889741912484,
-0.03671428933739662,
0.0063575743697583675,
-0.004535461775958538,
0.13206523656845093,
-0.032581888139247894,
0.036747097969055176,
-0.06036640703678131,
0.06258942186832428,
-0.06075510382652283,
0.04283398762345314,
-0.008679737336933613,
0.07443202286958694,
0.046928443014621735,
-0.027871103957295418,
0.01971372589468956,
-0.06816416233778,
0.025655517354607582,
0.05299580097198486,
0.03167802467942238,
-0.05020392686128616,
0.08316733688116074,
-0.0012126774527132511,
0.05849824100732803,
0.07217665016651154,
0.02610844001173973,
0.047732967883348465,
0.004243528936058283,
-0.07388508319854736,
0.1226164698600769,
0.00484576728194952,
-0.007492390926927328,
0.023521536961197853,
-0.0072961486876010895,
-0.0032154894433915615,
0.0032893940806388855,
0.07798473536968231,
0.021828584372997284,
0.02910696715116501,
0.05285734310746193,
0.029118144884705544,
0.014096781611442566,
-0.06180847808718681,
-0.015504100359976292,
0.0210233423858881,
-0.012914367020130157,
-0.06773358583450317,
0.014325203374028206,
0.008127299137413502,
0.05563681945204735,
-0.05462132394313812,
-0.0038786802906543016,
-0.011430026963353157,
-0.02546694129705429,
-0.04961114376783371,
0.03369049355387688,
0.07803104817867279,
0.03564494103193283,
-0.02180824801325798,
-0.05875324085354805,
0.05143601819872856,
0.020153604447841644,
0.009117704816162586,
0.005988867022097111,
3.7112967079437908e-34,
-0.028486603870987892,
0.03991737216711044,
0.0896117240190506,
0.02143135480582714,
-0.021153731271624565,
0.023363040760159492,
-0.0024025659076869488,
0.02177223190665245,
0.01577608659863472,
0.004675679374486208,
0.016775071620941162,
0.08334998786449432,
-0.11605013161897659,
0.100164994597435,
-0.020494306460022926,
0.007609759923070669,
0.004622532054781914,
0.07111310213804245,
0.03847562149167061,
-0.006310345605015755,
0.017614182084798813,
-0.03815227001905441,
-0.10460321605205536,
-0.029961759224534035,
0.0073614283464848995,
-0.025851115584373474,
0.119327113032341,
-0.10101659595966339,
-0.0059954640455543995,
0.002041703322902322,
-0.06940008699893951,
0.0014574188971892,
-0.01217588409781456,
0.01637835055589676,
0.016323259100317955,
-0.07805827260017395,
-0.05680364370346069,
-0.037228092551231384,
-0.03257197514176369,
-0.0034243378322571516,
0.00470636785030365,
0.02081913873553276,
-0.027523161843419075,
-0.036991991102695465,
-0.05347534269094467,
-0.14939989149570465,
-0.00528976134955883,
0.019354835152626038,
0.0036602315958589315,
0.007046538405120373,
0.08871905505657196,
0.00890428014099598,
-0.031013328582048416,
-0.11829320341348648,
-0.01866726763546467,
-0.00021222780924290419,
0.02005651220679283,
-0.03724881261587143,
-0.02268107235431671,
0.014361580833792686,
-0.05954713001847267,
0.017688335850834846,
-0.014784176833927631,
-0.020899813622236252,
-0.07650284469127655,
0.06603245437145233,
0.03818536549806595,
0.09040921926498413,
0.005031900480389595,
-0.0007828858215361834,
-0.017724057659506798,
0.005596826318651438,
0.017075568437576294,
-0.06713887304067612,
0.07771997153759003,
0.054552361369132996,
0.08616270124912262,
-0.07027147710323334,
-0.03244795277714729,
0.09677565842866898,
0.005274499766528606,
-0.043685995042324066,
0.04680698737502098,
0.009201539680361748,
0.01704087294638157,
-0.021532045677304268,
0.04858190938830376,
0.006139525678008795,
-0.06257008761167526,
-0.10949654132127762,
-0.05272310599684715,
-0.026790186762809753,
-0.019781379029154778,
0.049676984548568726,
-0.03816153109073639,
-1.4217200097012472e-33,
0.06897816807031631,
-0.03915902227163315,
0.011960912495851517,
-0.03175186365842819,
0.03484676405787468,
-0.05481854826211929,
-0.06510074436664581,
0.026794714853167534,
0.007846793159842491,
-0.0081523722037673,
-0.00786991324275732,
-0.08825841546058655,
0.03377438336610794,
-0.0020158446859568357,
0.007678280118852854,
0.006163663696497679,
-0.0017504766583442688,
0.017032727599143982,
-0.05476995185017586,
-0.05624008551239967,
0.03117774799466133,
0.00941420253366232,
-0.05948121100664139,
0.02644846960902214,
0.012133006937801838,
-0.01666533201932907,
0.014668791554868221,
0.10950671881437302,
-0.05128098279237747,
0.010283872485160828,
-0.061095815151929855,
-0.07895145565271378,
0.08020665496587753,
-0.0413990318775177,
-0.03310714289546013,
0.0681358203291893,
0.07451213896274567,
-0.032374147325754166,
-0.045473936945199966,
0.0012514705304056406,
0.13038842380046844,
-0.005491037853062153,
-0.017696211114525795,
0.03293902426958084,
-0.04597548395395279,
0.019783109426498413,
-0.06689733266830444,
0.06405804306268692,
0.07065768539905548,
0.05117621272802353,
-0.06617522984743118,
0.08841338008642197,
-0.047056909650564194,
0.011275218799710274,
-0.03295369818806648,
-0.08394075930118561,
-0.0019118626369163394,
-0.07581085711717606,
0.0259045772254467,
-0.033673565834760666,
-0.01513278391212225,
-0.019122779369354248,
0.05418591946363449,
-0.055866263806819916,
0.01738450489938259,
-0.06480476260185242,
0.02582906372845173,
-0.07329737395048141,
0.03089647740125656,
-0.01017963606864214,
0.06188450753688812,
0.016476662829518318,
-0.0711025595664978,
-0.07673133909702301,
-0.029555708169937134,
0.011949227191507816,
0.0601581335067749,
-0.09402460604906082,
-0.049402736127376556,
-0.019287804141640663,
-0.0349956676363945,
-0.0021992181427776814,
0.08180437982082367,
-0.006993808317929506,
0.004610789008438587,
0.0919669046998024,
0.11997807770967484,
0.07249229401350021,
0.06132005155086517,
0.011113041080534458,
-0.12363582104444504,
0.011710061691701412,
-0.03321109712123871,
0.11306614428758621,
-0.007595529314130545,
-3.940817450143186e-8,
-0.02642974443733692,
0.0689542144536972,
0.018919337540864944,
0.009936429560184479,
-0.008083352819085121,
-0.01736990548670292,
-0.013576631434261799,
0.011064780876040459,
0.024856967851519585,
0.031614113599061966,
0.09697414934635162,
-0.09480933099985123,
-0.03532608970999718,
-0.03913786634802818,
0.00007913917215773836,
0.029527634382247925,
-0.008891570381820202,
-0.0344766341149807,
-0.043277185410261154,
-0.035866156220436096,
0.02839859388768673,
0.0644804835319519,
-0.01095779612660408,
-0.04064125195145607,
0.0822959765791893,
-0.08367261290550232,
-0.0461571179330349,
0.10589839518070221,
0.09932804852724075,
-0.024939965456724167,
-0.0217290036380291,
0.0502798818051815,
-0.05256841704249382,
-0.09857138246297836,
-0.10671623796224594,
-0.0009300410747528076,
0.05255696550011635,
0.021206816658377647,
-0.0016823422629386187,
-0.01924353651702404,
0.013988313265144825,
0.029130255803465843,
-0.09785562008619308,
0.060567669570446014,
0.014404010027647018,
-0.0728851854801178,
0.009367515332996845,
-0.05201209336519241,
0.015116174705326557,
-0.03053230233490467,
0.04494732990860939,
-0.02580096572637558,
0.03948933631181717,
0.04823598638176918,
0.12862427532672882,
0.048578426241874695,
0.003256586380302906,
-0.028654467314481735,
0.010004467330873013,
0.04248793423175812,
0.015326366759836674,
-0.00241298065520823,
-0.05571432411670685,
0.03417644649744034
] |
Reflective Gaussian Splatting | https://openreview.net/forum?id=xPxHQHDH2u | [
"Yuxuan Yao",
"Zixuan Zeng",
"Chun Gu",
"Xiatian Zhu",
"Li Zhang"
] | Poster | Novel view synthesis has experienced significant advancements owing to increasingly capable NeRF- and 3DGS-based methods. However, reflective object reconstruction remains challenging, lacking a proper solution to achieve real-time, high-quality rendering while accommodating inter-reflection. To fill this gap, we introduce a Reflective Gaussian splatting (Ref-Gaussian) framework characterized with two components: (I) Physically based deferred rendering that empowers the rendering equation with pixel-level material properties via formulating split-sum approximation; (II) Gaussian-grounded inter-reflection that realizes the desired inter-reflection function within a Gaussian splatting paradigm for the first time. To enhance geometry modeling, we further introduce material-aware normal propagation and an initial per-Gaussian shading stage, along with 2D Gaussian primitives. Extensive experiments on standard datasets demonstrate that Ref-Gaussian surpasses existing approaches in terms of quantitative metrics, visual quality, and compute efficiency. Further, we show that our method serves as a unified solution for both reflective and non-reflective scenes, going beyond the previous alternatives focusing on only reflective scenes. Also, we illustrate that Ref-Gaussian supports more applications such as relighting and editing. | Gaussain-Splatting, Physically based Rendering, Deferred-Rendering, Inter-Reflection | null | 1,286 | 2412.19282 | [
-0.03444276750087738,
-0.0850256085395813,
0.09913953393697739,
0.00828103069216013,
0.05654725804924965,
-0.07474742084741592,
0.0031320720445364714,
-0.03983287513256073,
-0.024117613211274147,
-0.030961796641349792,
-0.0692993625998497,
-0.06374769657850266,
-0.01277894340455532,
0.034540992230176926,
0.002322260057553649,
-0.02264646254479885,
0.15058143436908722,
0.054991498589515686,
-0.037067681550979614,
0.041200876235961914,
0.07088641822338104,
-0.07853084057569504,
-0.08148842304944992,
-0.07198864966630936,
0.11703204363584518,
-0.014060135930776596,
0.05517091229557991,
-0.03387971594929695,
0.04737116023898125,
0.016265632584691048,
0.028655216097831726,
0.002453686436638236,
-0.06093185395002365,
-0.06117965281009674,
-0.09041481465101242,
0.03784836828708649,
0.04469463974237442,
0.02823730930685997,
0.020107842981815338,
-0.03722265362739563,
0.026926737278699875,
0.042248375713825226,
-0.03362732008099556,
0.009514517150819302,
0.04880017042160034,
0.0007305299513973296,
0.07350730150938034,
-0.11164165288209915,
0.029911482706665993,
-0.049009792506694794,
-0.008218491449952126,
0.005168744828552008,
-0.06460796296596527,
0.003977810498327017,
0.05919396132230759,
0.0997626930475235,
0.0015507403295487165,
-0.06054050102829933,
0.047890182584524155,
0.008461030200123787,
-0.0013282649451866746,
0.00275602238252759,
0.06947522610425949,
-0.029126359149813652,
0.02110155299305916,
0.0027407126035541296,
0.027353666722774506,
-0.060364674776792526,
-0.0010320504661649466,
0.09291676431894302,
-0.10409405082464218,
0.0400858148932457,
0.03538220003247261,
-0.070458322763443,
-0.03507935255765915,
0.07468948513269424,
-0.006939050741493702,
-0.04101453348994255,
-0.03590191900730133,
-0.03852134570479393,
0.09478092193603516,
0.03528242185711861,
0.013398358598351479,
0.040932513773441315,
-0.03795376792550087,
0.0301323551684618,
-0.015133225359022617,
0.1246819868683815,
0.013646356761455536,
0.02988678403198719,
0.012827190570533276,
0.03835289180278778,
-0.03667128458619118,
-0.04137638583779335,
-0.0728350281715393,
-0.07018860429525375,
0.006311031058430672,
-0.08069149404764175,
0.08653761446475983,
0.042728032916784286,
0.04966418445110321,
0.021863501518964767,
0.05261492729187012,
-0.035948678851127625,
-0.0008847948047332466,
-0.036263301968574524,
0.04464195668697357,
0.09807470440864563,
-0.07774721086025238,
0.050946176052093506,
0.049386367201805115,
0.03639296442270279,
-0.03940543532371521,
-0.00163698208052665,
-0.02927802875638008,
0.04279197007417679,
-0.0022187125869095325,
-0.026737365871667862,
0.011075278744101524,
0.06713354587554932,
0.04850441589951515,
0.024084823206067085,
-0.00207479833625257,
0.06276585161685944,
0.052391331642866135,
-0.058435969054698944,
0.00649453466758132,
1.8712931506320082e-33,
0.029993677511811256,
0.06753301620483398,
0.02919895015656948,
0.021663576364517212,
0.06356488913297653,
0.02908697910606861,
0.016945255920290947,
0.030942143872380257,
0.008419116027653217,
0.002725563943386078,
-0.015067243948578835,
-0.01211325079202652,
-0.013000081293284893,
0.04674293100833893,
0.021950164809823036,
0.08937248587608337,
-0.03737000748515129,
-0.011401106603443623,
-0.06283123791217804,
0.03478042781352997,
-0.020787164568901062,
-0.012091363780200481,
0.0331323966383934,
-0.011273550800979137,
-0.0852232277393341,
0.016427744179964066,
0.0027582552284002304,
-0.005180991720408201,
-0.057808272540569305,
-0.023194417357444763,
-0.018662545830011368,
0.01846933551132679,
-0.002270376542583108,
0.03538709506392479,
-0.04268241673707962,
-0.027967484667897224,
0.037153925746679306,
-0.0885554701089859,
-0.02331439033150673,
-0.004345995374023914,
-0.09837213158607483,
0.11293916404247284,
0.09736911207437515,
0.009072920307517052,
-0.08248935639858246,
-0.005164065398275852,
0.054658934473991394,
0.08636006712913513,
0.015180586837232113,
0.04339691624045372,
0.06104223430156708,
-0.05498151108622551,
-0.04880569875240326,
-0.04408058896660805,
-0.088686503469944,
-0.04855167120695114,
0.05079742521047592,
0.02294979989528656,
0.0743006244301796,
-0.023224271833896637,
-0.00618391390889883,
0.006724681705236435,
-0.03537068888545036,
0.05148882418870926,
-0.10746366530656815,
0.04702650010585785,
-0.029171479865908623,
0.011134553700685501,
-0.05123741179704666,
0.07809363305568695,
-0.05086160823702812,
-0.0006020673899911344,
0.02003522776067257,
0.024335075169801712,
0.06130683794617653,
-0.04824259504675865,
-0.02771146409213543,
-0.09084164351224899,
0.03665823116898537,
0.10671549290418625,
-0.10215786844491959,
0.04003720358014107,
-0.026934823021292686,
-0.10792521387338638,
-0.0524817518889904,
-0.04694173485040665,
-0.007018971722573042,
-0.011861493811011314,
-0.05907206982374191,
-0.12882913649082184,
-0.0806201845407486,
-0.046510446816682816,
-0.03543521463871002,
-0.05691525712609291,
-0.049316830933094025,
-1.7375000366603463e-33,
0.04647368937730789,
-0.0034776534885168076,
-0.008916953578591347,
0.0758935883641243,
-0.012401538901031017,
-0.058942761272192,
0.0030062454752624035,
-0.0017377602634951472,
0.039793286472558975,
-0.02337205410003662,
-0.02755376510322094,
0.02946408838033676,
0.059927474707365036,
0.08144228160381317,
-0.0728391632437706,
-0.051289353519678116,
-0.020623646676540375,
-0.06527484953403473,
-0.08566728234291077,
-0.10645196586847305,
0.031034298241138458,
-0.010185584425926208,
0.050936635583639145,
0.010407363064587116,
-0.046849098056554794,
0.1251116544008255,
-0.018957320600748062,
-0.0009018551791086793,
-0.02591206319630146,
0.015236668288707733,
-0.01113804243505001,
-0.03128000348806381,
0.0696021094918251,
-0.04837248474359512,
-0.061001818627119064,
0.04758762940764427,
0.08596891164779663,
-0.0010382821783423424,
-0.030181031674146652,
-0.04866114258766174,
-0.0068083275109529495,
-0.0010815896093845367,
0.03974452242255211,
0.07203961908817291,
-0.04711908474564552,
0.03856918588280678,
0.007924371398985386,
-0.02396027185022831,
0.0016292333602905273,
0.02864772640168667,
-0.030848225578665733,
0.0555252842605114,
-0.021772315725684166,
0.022771060466766357,
0.02367382124066353,
-0.08854450285434723,
-0.14158184826374054,
0.005019054748117924,
0.1114388108253479,
0.02678573876619339,
0.00821619015187025,
0.008104635402560234,
-0.07436486333608627,
-0.041881971061229706,
0.011334150098264217,
0.027890082448720932,
-0.035174861550331116,
-0.06947578489780426,
0.05557829141616821,
0.05850904434919357,
-0.03719385340809822,
0.026852747425436974,
-0.06754133105278015,
-0.02142694778740406,
-0.0001952009042724967,
0.006812714040279388,
0.03353523090481758,
0.04342728853225708,
-0.0012576142325997353,
0.07370784878730774,
0.07300073653459549,
-0.010093639604747295,
0.04450850188732147,
0.09100332111120224,
0.059604596346616745,
0.01479378342628479,
-0.07593017816543579,
0.032113347202539444,
0.026620004326105118,
-0.002553773345425725,
-0.045237235724925995,
0.06568700820207596,
0.006032843608409166,
0.025841161608695984,
-0.007853935472667217,
-4.546686582784787e-8,
0.0046468209475278854,
0.05467148870229721,
0.007630058564245701,
-0.010062779299914837,
-0.03137461096048355,
-0.054612863808870316,
-0.014263372868299484,
-0.0459052249789238,
-0.03837623819708824,
-0.1154475137591362,
0.01168388593941927,
-0.14327090978622437,
-0.02326999418437481,
-0.02001482993364334,
0.060740187764167786,
0.10318538546562195,
0.01498489361256361,
-0.04127173498272896,
-0.05668097734451294,
0.02310151793062687,
-0.0882345661520958,
-0.018428806215524673,
-0.038513485342264175,
0.015972573310136795,
-0.008829428814351559,
0.06996747106313705,
-0.03385097533464432,
0.057579126209020615,
0.11205495148897171,
0.027495985850691795,
0.0026543347630649805,
-0.031277261674404144,
0.014704291708767414,
0.03081381879746914,
-0.05513361096382141,
0.008690458722412586,
-0.01337102148681879,
-0.0015385652659460902,
0.030761880800127983,
-0.03716413676738739,
-0.037046823650598526,
0.03773682191967964,
0.019375858828425407,
0.03198467195034027,
0.08038759231567383,
0.02274322509765625,
-0.018012240529060364,
-0.04393787682056427,
-0.030554290860891342,
0.017141234129667282,
0.04870004579424858,
-0.039418790489435196,
-0.04974699765443802,
0.08542690426111221,
0.03129572421312332,
-0.03416389971971512,
0.03205491602420807,
0.01628820225596428,
0.07112453132867813,
-0.006154898554086685,
0.02950703352689743,
-0.004817838314920664,
-0.05772995948791504,
0.033703844994306564
] |
Data Center Cooling System Optimization Using Offline Reinforcement Learning | https://openreview.net/forum?id=W8xukd70cU | [
"Xianyuan Zhan",
"Xiangyu Zhu",
"Peng Cheng",
"Xiao Hu",
"Ziteng He",
"Hanfei Geng",
"Jichao Leng",
"Huiwen Zheng",
"Chenhui Liu",
"Tianshun Hong",
"Yan Liang",
"Yunxin Liu",
"Feng Zhao"
] | Poster | The recent advances in information technology and artificial intelligence have fueled a rapid expansion of the data center (DC) industry worldwide, accompanied by an immense appetite for electricity to power the DCs. In a typical DC, around 30-40% of the energy is spent on the cooling system rather than on computer servers, posing a pressing need for developing new energy-saving optimization technologies for DC cooling systems. However, optimizing such real-world industrial systems faces numerous challenges, including but not limited to a lack of reliable simulation environments, limited historical data, and stringent safety and control robustness requirements. In this work, we present a novel physics-informed offline reinforcement learning (RL) framework for energy efficiency optimization of DC cooling systems. The proposed framework models the complex dynamical patterns and physical dependencies inside a server room using a purposely designed graph neural network architecture that is compliant with the fundamental time-reversal symmetry. Because of its well-behaved and generalizable state-action representations, the model enables sample-efficient and robust latent space offline policy learning using limited real-world operational data. Our framework has been successfully deployed and verified in a large-scale production DC for closed-loop control of its air-cooling units (ACUs). We conducted a total of 2000 hours of short and long-term experiments in the production DC environment. The results show that our method achieves 14-21% energy savings in the DC cooling system, without any violation of the safety or operational constraints. We have also conducted a comprehensive evaluation of our approach in a real-world DC testbed environment. Our results have demonstrated the significant potential of offline RL in solving a broad range of data-limited, safety-critical real-world industrial control problems. | Offline Reinforcement learning, data center optimization, cooling system, energy saving | We developed and deployed a sample-efficient offline RL framework for energy efficiency optimization of data center's cooling system | 1,285 | 2501.15085 | [
-0.09882552921772003,
-0.03517408296465874,
0.030363311991095543,
0.1149914339184761,
0.027058670297265053,
-0.02294391766190529,
-0.07194004952907562,
0.015069527551531792,
-0.06268198788166046,
0.08192802965641022,
-0.018671439960598946,
-0.005473538767546415,
0.03316613659262657,
-0.012506301514804363,
0.009773296304047108,
0.07459001988172531,
0.09735744446516037,
-0.024145808070898056,
-0.09419851750135422,
-0.09321632236242294,
0.023866022005677223,
-0.0020695796702057123,
-0.029844479635357857,
-0.014941781759262085,
0.04818802699446678,
0.11377832293510437,
0.04227464273571968,
0.0464010052382946,
-0.041353508830070496,
0.05440645292401314,
0.008689187467098236,
-0.10732617974281311,
-0.02760704606771469,
-0.08896303176879883,
-0.023924220353364944,
0.07600909471511841,
-0.09286779165267944,
0.019320741295814514,
0.03229574114084244,
0.0048063406720757484,
-0.0021550129167735577,
-0.019038543105125427,
0.0536651536822319,
0.04140191525220871,
0.0749305710196495,
0.03988838195800781,
-0.08072339743375778,
-0.02171357348561287,
0.03306769207119942,
0.03037211112678051,
0.02105852961540222,
0.0016472749412059784,
0.015511705540120602,
0.04300396516919136,
0.03917906433343887,
0.07875218242406845,
0.057691916823387146,
-0.007601931691169739,
0.024378573521971703,
-0.012810437940061092,
0.04313064366579056,
-0.12880799174308777,
0.011856509372591972,
-0.053327541798353195,
-0.07712843269109726,
0.03100382536649704,
-0.031153269112110138,
0.08342371135950089,
0.038018498569726944,
-0.048070237040519714,
-0.03435627371072769,
0.07099002599716187,
0.008379773236811161,
-0.03727882355451584,
0.014838458970189095,
-0.023020850494503975,
0.030250608921051025,
-0.04892826825380325,
0.03629610314965248,
-0.017851492390036583,
0.03030724823474884,
0.009854686446487904,
-0.04674460366368294,
0.014681319706141949,
-0.0005411436432041228,
0.0029405306559056044,
0.03621706739068031,
0.050055328756570816,
0.11489564180374146,
0.010649806819856167,
0.03646072372794151,
0.11552649736404419,
0.03515695780515671,
-0.06133607402443886,
0.002290497301146388,
0.0015389514155685902,
0.015254680067300797,
-0.02125418558716774,
-0.07267190515995026,
0.013985916040837765,
0.024357620626688004,
0.059529684484004974,
0.06751590967178345,
0.03637157008051872,
0.03395959362387657,
-0.07847044616937637,
-0.04538556560873985,
0.053882986307144165,
0.059001218527555466,
-0.14385446906089783,
-0.01417459174990654,
0.04748515039682388,
-0.023530030623078346,
0.015888376161456108,
0.0020463867112994194,
-0.009074500761926174,
0.04407479241490364,
-0.10696006566286087,
-0.010712225921452045,
0.05923346430063248,
-0.024648649618029594,
-0.005136874504387379,
-0.02775910310447216,
0.0010039667831733823,
0.041775915771722794,
-0.06966883689165115,
-0.03720499202609062,
-6.388255044023675e-35,
-0.03748783841729164,
-0.01866050250828266,
0.03339044004678726,
-0.01653435081243515,
0.09755661338567734,
-0.030493078753352165,
0.0115553829818964,
0.014257468283176422,
0.0516931377351284,
0.01222163625061512,
0.01397046446800232,
-0.002723166486248374,
-0.043064214289188385,
0.08644766360521317,
0.05992637574672699,
-0.028066273778676987,
0.01986386626958847,
-0.03622358664870262,
0.04949218034744263,
-0.10116007179021835,
0.047562092542648315,
-0.014802654273808002,
0.017870843410491943,
0.03438054770231247,
0.06427129358053207,
0.04980867728590965,
0.0861172080039978,
0.02350720390677452,
0.017663614824414253,
-0.02193334326148033,
0.0637395828962326,
0.005752161610871553,
-0.049681443721055984,
0.049973588436841965,
0.008945914916694164,
-0.0014273335691541433,
-0.004235590808093548,
0.0015972591936588287,
0.011536845937371254,
0.006947284564375877,
0.009094753302633762,
0.006061427760869265,
-0.029838064685463905,
0.03349311649799347,
-0.08371587097644806,
-0.0811922550201416,
0.0751049816608429,
-0.07058613002300262,
-0.018142199143767357,
-0.06906040012836456,
-0.025795310735702515,
0.01005298551172018,
0.027462493628263474,
-0.011520136147737503,
0.009450672194361687,
-0.032550882548093796,
0.03480584919452667,
0.04876363277435303,
0.03340606018900871,
0.024035951122641563,
-0.042558129876852036,
-0.0029774941504001617,
0.026997318491339684,
0.0449649803340435,
-0.02998826652765274,
0.03211579844355583,
-0.0610872320830822,
0.01113932579755783,
-0.018477540463209152,
-0.04641690105199814,
0.06478296220302582,
0.039482422173023224,
0.0401250422000885,
-0.0918297991156578,
0.11921974271535873,
0.039512891322374344,
0.001546933432109654,
-0.04645245894789696,
-0.10901758074760437,
-0.03472499921917915,
-0.02758987806737423,
-0.01887458749115467,
0.0068285297602415085,
0.007045222911983728,
0.014864487573504448,
-0.15312643349170685,
-0.02094266004860401,
0.010674610733985901,
-0.02025221474468708,
-0.06496383249759674,
-0.11734867841005325,
-0.05508587136864662,
0.0016006140504032373,
0.07022932916879654,
0.016835056245326996,
1.952724327923772e-34,
-0.0012893818784505129,
-0.020719939842820168,
0.04172057658433914,
0.028498830273747444,
0.018420400097966194,
0.017785297706723213,
-0.05204914137721062,
-0.05877683311700821,
0.014758408069610596,
0.039262257516384125,
-0.0308422539383173,
-0.022119447588920593,
-0.016140928491950035,
-0.056394387036561966,
0.07937929034233093,
0.0349828265607357,
-0.08728024363517761,
-0.013318436220288277,
0.02347675897181034,
0.05361546203494072,
-0.01774095743894577,
0.0035837427712976933,
-0.07143145054578781,
0.005834691692143679,
0.008826498873531818,
0.035475410521030426,
-0.033328596502542496,
0.011426332406699657,
0.04856918007135391,
0.09319083392620087,
-0.08262056112289429,
-0.02667953073978424,
0.013533205725252628,
0.02138420194387436,
-0.06200146675109863,
0.05327660217881203,
0.064402736723423,
0.011141382157802582,
-0.016065968200564384,
0.05241968855261803,
0.06985865533351898,
-0.08437126129865646,
-0.10758282989263535,
0.009899789467453957,
-0.014817317016422749,
0.0025572997983545065,
-0.04736575856804848,
0.010418927296996117,
-0.07207880169153214,
0.04760146141052246,
-0.017676547169685364,
-0.002304832451045513,
-0.039180196821689606,
-0.009622050449252129,
-0.04128964617848396,
-0.02767174132168293,
0.025768354535102844,
0.08716356754302979,
0.02574516087770462,
-0.011127248406410217,
0.0069070360623300076,
-0.08765190094709396,
-0.006456607021391392,
0.0007510380819439888,
-0.07791528105735779,
0.04339609295129776,
0.029864177107810974,
0.03996843472123146,
0.043042123317718506,
-0.05177571251988411,
-0.0049619111232459545,
0.007796772290021181,
-0.01097642257809639,
-0.06019718572497368,
0.03314153477549553,
-0.04124166816473007,
0.015732135623693466,
-0.02856227196753025,
-0.01630960777401924,
0.06950019299983978,
-0.02262384258210659,
0.03066290356218815,
0.04983598738908768,
0.04518711194396019,
0.07615301012992859,
0.009413488209247589,
-0.011734150350093842,
0.062405262142419815,
-0.0062899817712605,
-0.015208208933472633,
-0.024774808436632156,
-0.03662260249257088,
-0.01592300459742546,
0.08483105152845383,
-0.06640355288982391,
-3.840118623088529e-8,
-0.14141598343849182,
-0.0013698464026674628,
0.16358743607997894,
0.008966704830527306,
0.1115325465798378,
-0.01789288967847824,
0.08086416125297546,
0.008979306556284428,
0.04343133419752121,
0.023240951821208,
0.12148574739694595,
-0.030648237094283104,
0.02379303239285946,
-0.10083655267953873,
-0.054663851857185364,
0.026094632223248482,
-0.0007632299093529582,
0.0030234409496188164,
0.07476934045553207,
0.021006835624575615,
0.035517409443855286,
-0.03004406951367855,
-0.016495870426297188,
-0.01611189916729927,
0.07146772742271423,
-0.0560813844203949,
-0.023882891982793808,
0.056561704725027084,
0.039728909730911255,
0.0008859597146511078,
-0.017389433458447456,
-0.0428290031850338,
0.12001343071460724,
-0.04900537058711052,
0.040388621389865875,
-0.018339093774557114,
0.1113463044166565,
-0.09563225507736206,
0.019564155489206314,
-0.032929323613643646,
-0.013686553575098515,
0.0724550113081932,
-0.06876809895038605,
-0.020533407106995583,
0.020012447610497475,
-0.04227140173316002,
-0.04124471917748451,
-0.06560578942298889,
0.09721191972494125,
0.021324647590517998,
-0.005469599738717079,
-0.023310475051403046,
-0.016976702958345413,
0.027230488136410713,
0.05246645212173462,
-0.025719353929162025,
0.0015285665867850184,
-0.04275191202759743,
0.011289498768746853,
0.0744214728474617,
-0.01585487462580204,
-0.008778364397585392,
-0.1674460768699646,
-0.006347006186842918
] |
Bias Mitigation in Graph Diffusion Models | https://openreview.net/forum?id=CSj72Rr2PB | [
"Meng Yu",
"Kun Zhan"
] | Poster | Most existing graph diffusion models have significant bias problems. We observe that the forward diffusion’s maximum perturbation distribution in most models deviates from the standard Gaussian distribution, while reverse sampling consistently starts from a standard Gaussian distribution, which results in a reverse-starting bias. Together with the inherent exposure bias of diffusion models, this results in degraded generation quality. This paper proposes a comprehensive approach to mitigate both biases. To mitigate reverse-starting bias, we employ a newly designed Langevin sampling algorithm to align with the forward maximum perturbation distribution, establishing a new reverse-starting point. To address the exposure bias, we introduce a score correction mechanism based on a newly defined score difference. Our approach, which requires no network modifications, is validated across multiple models, datasets, and tasks, achieving state-of-the-art results. | Diffusion models, Graph learning, Bias analysis | null | 1,279 | null | [
-0.05254397913813591,
-0.028465313836932182,
0.07158201187849045,
0.07011433690786362,
0.06242366507649422,
-0.06979542225599289,
-0.09842382371425629,
-0.026094546541571617,
0.008840370923280716,
0.0532224178314209,
-0.036582332104444504,
-0.00036492772051133215,
0.06546285003423691,
-0.007817862555384636,
-0.02251136489212513,
-0.04999275505542755,
0.049850378185510635,
0.05009952932596207,
-0.017377890646457672,
-0.04402538016438484,
-0.07218805700540543,
0.054573286324739456,
-0.03439496085047722,
-0.025717105716466904,
0.0988643690943718,
-0.07030796259641647,
0.039036113768815994,
-0.045688025653362274,
-0.008196269161999226,
-0.033924710005521774,
0.06542172282934189,
-0.0234253890812397,
-0.02081179805099964,
0.015317539684474468,
-0.024824701249599457,
-0.009937930852174759,
-0.06148605793714523,
0.02496699057519436,
0.0917435958981514,
0.04763296619057655,
0.053041934967041016,
0.01117037609219551,
0.022927114740014076,
0.0565728023648262,
0.06405884027481079,
0.0009216665639542043,
-0.01596875861287117,
0.01705520786345005,
0.029266903176903725,
0.07580582797527313,
0.04728309065103531,
-0.0015417365357279778,
-0.06587698310613632,
-0.005826313514262438,
-0.016225138679146767,
-0.006889558397233486,
0.04407182335853577,
0.033048465847969055,
0.028969280421733856,
0.0516764335334301,
0.006091096438467503,
-0.11018659919500351,
-0.05040926858782768,
-0.00238565681502223,
0.0025324977468699217,
0.021171903237700462,
0.03220955654978752,
0.08370625972747803,
-0.04066392034292221,
0.1364266574382782,
0.031889658421278,
0.04945508763194084,
-0.11222980171442032,
0.006041237618774176,
0.00850270688533783,
0.110218346118927,
-0.04932180047035217,
0.015009718015789986,
0.014651906676590443,
-0.04180437698960304,
0.0330965518951416,
-0.017670173197984695,
0.0860772654414177,
-0.015861153602600098,
0.002158920280635357,
-0.024205584079027176,
0.08635938912630081,
-0.010093032382428646,
0.07883153855800629,
0.004192409571260214,
-0.037553027272224426,
0.09711205214262009,
-0.009607362560927868,
0.04900345578789711,
-0.038786523044109344,
0.028528554365038872,
0.0070247771218419075,
-0.06574505567550659,
0.04933487996459007,
0.06198228895664215,
-0.015105451457202435,
0.031017418950796127,
0.01404829416424036,
-0.024418717250227928,
0.051159437745809555,
-0.027055159211158752,
0.060525164008140564,
0.05135658010840416,
-0.0046598478220403194,
-0.03253351151943207,
-0.001068144105374813,
0.09786931425333023,
0.05575580894947052,
-0.007930864579975605,
0.0436740443110466,
-0.08104108273983002,
0.03696121647953987,
0.01142109278589487,
0.020199839025735855,
0.060386791825294495,
-0.022590722888708115,
0.033855024725198746,
-0.02556055597960949,
-0.04938211292028427,
-0.02522171288728714,
0.013629108667373657,
-0.00834392849355936,
1.8260559035795434e-33,
0.0363956019282341,
-0.029441488906741142,
0.05370989069342613,
-0.027870193123817444,
0.08429425209760666,
0.020870275795459747,
-0.04338179901242256,
-0.08685970306396484,
-0.057824648916721344,
-0.044917091727256775,
-0.032884981483221054,
0.005614601541310549,
-0.09618856012821198,
0.07968580722808838,
0.06370618939399719,
0.05997638404369354,
-0.022553738206624985,
0.07404497265815735,
0.0450088195502758,
-0.008591067045927048,
0.03682593256235123,
-0.11610715091228485,
-0.011657171882689,
-0.013335670344531536,
0.03358220309019089,
0.09249640256166458,
0.006329621188342571,
-0.0256933756172657,
-0.018206164240837097,
0.0014447292778640985,
-0.07202025502920151,
0.04393347352743149,
0.016390463337302208,
-0.02779700979590416,
-0.021200008690357208,
0.02834242582321167,
0.013455325737595558,
-0.002501147799193859,
-0.019694391638040543,
-0.07083248347043991,
-0.04514329135417938,
0.09036370366811752,
-0.012264049611985683,
-0.04759844020009041,
-0.12024080753326416,
-0.01564611680805683,
-0.01594551093876362,
-0.08029236644506454,
0.021623698994517326,
-0.03833797574043274,
-0.0017762824427336454,
-0.04719701036810875,
-0.0025285091251134872,
0.0015075295232236385,
-0.12005118280649185,
0.006550568621605635,
0.07010724395513535,
-0.051446136087179184,
-0.02968006394803524,
0.027454614639282227,
0.06857189536094666,
0.03026004321873188,
-0.06813596934080124,
0.03538098186254501,
0.06475654989480972,
0.04386643320322037,
-0.007743850350379944,
-0.007248584181070328,
0.0017301320331171155,
-0.0031272799242287874,
-0.025233402848243713,
0.020739974454045296,
-0.08174236118793488,
-0.0939757451415062,
-0.031571827828884125,
-0.038332149386405945,
-0.02071239985525608,
0.05477883294224739,
-0.013078715652227402,
-0.03933679312467575,
-0.07758338004350662,
-0.10073645412921906,
-0.07231981307268143,
-0.11657899618148804,
0.0027044974267482758,
-0.019691552966833115,
-0.0072172959335148335,
-0.0031798118725419044,
-0.03003549762070179,
-0.008921635337173939,
0.0334695428609848,
-0.028489556163549423,
-0.023010129109025,
0.07794784009456635,
-0.051346804946660995,
-1.6374112844626364e-33,
-0.10875850170850754,
0.08762959390878677,
0.034513261169195175,
0.10207952558994293,
-0.005549951456487179,
-0.04839125648140907,
0.033256400376558304,
0.027840152382850647,
0.05167476460337639,
0.030120909214019775,
-0.08234882354736328,
-0.07384692132472992,
-0.00103973678778857,
0.01433888915926218,
0.006281699053943157,
-0.06292255222797394,
0.03032308630645275,
-0.0036920327693223953,
-0.04797585308551788,
-0.033667899668216705,
0.006724674720317125,
0.07828976213932037,
-0.0660143718123436,
0.023989923298358917,
0.07585227489471436,
0.03828079253435135,
-0.064860999584198,
0.05721845477819443,
-0.011523412540555,
-0.059286851435899734,
-0.09952639788389206,
0.05589137598872185,
0.04243035614490509,
-0.044812534004449844,
-0.013845209963619709,
0.08905881643295288,
0.04214449226856232,
0.04705963656306267,
0.021532244980335236,
0.07097036391496658,
0.06517203152179718,
-0.009664366953074932,
-0.1285230815410614,
0.007115266751497984,
0.04962804540991783,
0.041007690131664276,
-0.00996863842010498,
0.06400743126869202,
-0.06480958312749863,
-0.02852478250861168,
-0.03517288714647293,
0.0653923973441124,
-0.0019455786095932126,
0.08647152036428452,
-0.06263633072376251,
-0.09564150869846344,
0.0018405887531116605,
0.029899412766098976,
0.06038721650838852,
0.03247124329209328,
-0.08282153308391571,
-0.038217298686504364,
-0.013268518261611462,
-0.03455116227269173,
-0.09631313383579254,
0.012993118725717068,
-0.03882888704538345,
0.007329090498387814,
0.07026698440313339,
0.03603702783584595,
-0.00956693384796381,
0.02433166652917862,
0.03929243981838226,
-0.04999113827943802,
0.012556301429867744,
-0.0819997563958168,
0.013563125394284725,
0.04166062921285629,
-0.028730304911732674,
-0.004392857197672129,
-0.000691610446665436,
0.08040742576122284,
0.04330475255846977,
-0.0073407478630542755,
0.041940074414014816,
-0.0036405317950993776,
0.056784503161907196,
-0.007996524684131145,
-0.01214934978634119,
0.040687717497348785,
-0.07687204331159592,
-0.04608853906393051,
-0.012611192651093006,
0.0624224916100502,
-0.0895472839474678,
-4.3409439598462995e-8,
-0.127369686961174,
0.07861417531967163,
0.004844377748668194,
0.11006343364715576,
0.041403915733098984,
0.01971224695444107,
0.0381297841668129,
0.11065450310707092,
0.005138037726283073,
-0.014765361323952675,
0.04390983656048775,
-0.008183262310922146,
-0.020671462640166283,
-0.028653431683778763,
-0.004989976063370705,
0.05978721007704735,
-0.025910690426826477,
0.06537318974733353,
0.023281583562493324,
-0.023184850811958313,
-0.06334219872951508,
0.03961380198597908,
0.03923460468649864,
-0.025973722338676453,
0.08333845436573029,
-0.014980840496718884,
0.009162033908069134,
-0.05475429445505142,
0.06035421043634415,
-0.020262062549591064,
-0.043481241911649704,
-0.0023133475333452225,
0.03570404648780823,
0.0332588367164135,
-0.001865330501459539,
0.07850431650876999,
-0.020091837272047997,
-0.02377253770828247,
0.03420415148139,
0.05036533996462822,
-0.04267989099025726,
0.10098814964294434,
-0.05102245882153511,
-0.008197704330086708,
-0.06625460088253021,
0.016637330874800682,
0.0338309109210968,
0.05386140197515488,
0.039536722004413605,
-0.12732338905334473,
-0.014555809088051319,
-0.11723922938108444,
-0.05004182085394859,
-0.06564413756132126,
0.00005778406557510607,
-0.033204078674316406,
-0.028922835364937782,
0.01949191652238369,
0.007090276572853327,
0.013977127149701118,
0.07172181457281113,
-0.040151312947273254,
-0.0606994703412056,
-0.00590458232909441
] |
The Utility and Complexity of In- and Out-of-Distribution Machine Unlearning | https://openreview.net/forum?id=HVFMooKrHX | [
"Youssef Allouah",
"Joshua Kazdan",
"Rachid Guerraoui",
"Sanmi Koyejo"
] | Poster | Machine unlearning, the process of selectively removing data from trained models, is increasingly crucial for addressing privacy concerns and knowledge gaps post-deployment. Despite this importance, existing approaches are often heuristic and lack formal guarantees. In this paper, we analyze the fundamental utility, time, and space complexity trade-offs of approximate unlearning, providing rigorous certification analogous to differential privacy. For in-distribution forget data—data similar to the retain set—we show that a surprisingly simple and general procedure, empirical risk minimization with output perturbation, achieves tight unlearning-utility-complexity trade-offs, addressing a previous theoretical gap on the separation from unlearning ``for free" via differential privacy, which inherently facilitates the removal of such data. However, such techniques fail with out-of-distribution forget data—data significantly different from the retain set—where unlearning time complexity can exceed that of retraining, even for a single sample. To address this, we propose a new robust and noisy gradient descent variant that provably amortizes unlearning time complexity without compromising utility. | machine unlearning, differential privacy, optimization, theory | null | 1,277 | null | [
-0.1371592879295349,
-0.04618510603904724,
-0.02982725389301777,
0.09615480154752731,
0.07111199200153351,
-0.02772665210068226,
-0.004184250254184008,
-0.08709676563739777,
0.03182625770568848,
0.007790598087012768,
0.04570590704679489,
-0.007347614038735628,
0.019614633172750473,
-0.007723216898739338,
-0.023906251415610313,
-0.04303514584898949,
0.025556199252605438,
0.06874031573534012,
-0.10786081850528717,
-0.04523957893252373,
-0.027077950537204742,
0.021511081606149673,
-0.035374127328395844,
0.07459498196840286,
0.05275995656847954,
-0.03279758244752884,
0.05280794948339462,
-0.03382530063390732,
-0.045399606227874756,
-0.04730590432882309,
0.08912181854248047,
0.00782843865454197,
-0.009044461883604527,
-0.027246547862887383,
0.03867022320628166,
0.02018101140856743,
-0.01341388002038002,
0.02575766295194626,
0.026555702090263367,
-0.03191148117184639,
0.03664620593190193,
-0.046719059348106384,
-0.053878992795944214,
0.05371023342013359,
-0.024552861228585243,
0.0014993983786553144,
0.028376709669828415,
-0.0823092833161354,
-0.021839598193764687,
0.003183051012456417,
0.01640397682785988,
0.010916396975517273,
-0.07681535184383392,
-0.0025615603663027287,
-0.015497434884309769,
-0.017501771450042725,
0.022323133423924446,
0.020695922896265984,
-0.09292495995759964,
0.05514459311962128,
-0.0664898157119751,
-0.10701750963926315,
-0.046239275485277176,
0.033632636070251465,
-0.03982972726225853,
0.03157060593366623,
0.05359572172164917,
0.03513018414378166,
0.04121366888284683,
0.009792394004762173,
-0.029067717492580414,
0.09912506490945816,
-0.10195954889059067,
0.04281407594680786,
0.022694775834679604,
0.015709245577454567,
0.04789983853697777,
-0.030215751379728317,
-0.011304766871035099,
-0.031451817601919174,
0.008232190273702145,
0.04090000316500664,
0.012700035236775875,
0.03931266814470291,
0.028810100629925728,
-0.039161067456007004,
0.024982279166579247,
0.05279024317860603,
0.09214670956134796,
-0.03475892171263695,
-0.03731822595000267,
-0.003031309926882386,
-0.043540097773075104,
0.016490774229168892,
0.048312924802303314,
-0.01183015201240778,
0.05092168226838112,
-0.023363681510090828,
-0.03244103118777275,
0.09091004729270935,
-0.03614005818963051,
0.08398643136024475,
-0.03451940789818764,
-0.018894199281930923,
0.006315253674983978,
0.0010594639461487532,
0.02818523347377777,
0.02713998779654503,
0.015366115607321262,
-0.008378051221370697,
0.06827934831380844,
0.014305644668638706,
-0.030669888481497765,
0.020036492496728897,
0.008332387544214725,
0.030955510213971138,
-0.05206757038831711,
0.011911160312592983,
0.0022768350318074226,
0.11845546960830688,
-0.017016299068927765,
-0.005179522559046745,
0.0769495740532875,
-0.007297534961253405,
-0.06403379142284393,
0.018576152622699738,
-0.03739989548921585,
9.724878539677208e-34,
-0.06949379295110703,
0.012108160182833672,
-0.02241680398583412,
-0.0702420026063919,
0.047313038259744644,
0.028090178966522217,
0.004091559909284115,
0.10949810594320297,
-0.036229271441698074,
0.06755142658948898,
0.017515458166599274,
0.04375239834189415,
-0.044513750821352005,
-0.020093074068427086,
0.06630802154541016,
0.02024834044277668,
0.013357833959162235,
0.10245873779058456,
0.03421538695693016,
-0.036688707768917084,
0.1631181687116623,
-0.04978444054722786,
-0.01697741448879242,
-0.03990745171904564,
0.010344083420932293,
-0.06284306943416595,
0.09281138330698013,
-0.014737071469426155,
0.053196072578430176,
0.04638127237558365,
-0.06438186764717102,
0.07061344385147095,
0.05755455046892166,
0.03217872232198715,
0.08877794444561005,
0.0020090241450816393,
-0.010414394550025463,
-0.07424679398536682,
0.03960263729095459,
0.002539165550842881,
-0.0015892494702711701,
0.016384879127144814,
-0.016647102311253548,
-0.008690250106155872,
-0.106768399477005,
-0.07114424556493759,
0.03506733477115631,
-0.04700468108057976,
0.010796314105391502,
-0.0810391902923584,
0.009138899855315685,
0.0001905794779304415,
-0.06035603955388069,
-0.07903897762298584,
-0.10440166294574738,
-0.04216822609305382,
0.0633828341960907,
-0.017120907083153725,
-0.005829894449561834,
0.045720767229795456,
-0.017300892621278763,
0.07298187911510468,
-0.002809370867908001,
0.0072393156588077545,
0.025507455691695213,
-0.011769533157348633,
-0.011064903810620308,
-0.009698707610368729,
-0.0008547685574740171,
-0.07461950182914734,
0.002304556081071496,
-0.0042930529452860355,
-0.07172930985689163,
-0.10559804737567902,
0.07007194310426712,
-0.0021337082143872976,
0.11341482400894165,
0.006257341708987951,
0.00934345182031393,
-0.06925442069768906,
0.05843225121498108,
-0.006079185288399458,
-0.03223488852381706,
0.017280014231801033,
-0.0647047758102417,
-0.07457519322633743,
0.018991034477949142,
0.051746416836977005,
-0.08216559886932373,
0.030764088034629822,
-0.09244970232248306,
0.0011954508954659104,
-0.05728920176625252,
-0.05138292536139488,
-0.019952887669205666,
-2.346812941634271e-33,
-0.09366098046302795,
0.001485063461586833,
-0.03987443819642067,
0.03268193081021309,
0.06714071333408356,
-0.06455473601818085,
-0.07258675992488861,
0.023968324065208435,
0.029083063825964928,
-0.04418448358774185,
-0.08815909922122955,
-0.02651897445321083,
0.049164123833179474,
-0.06627961248159409,
0.03344118967652321,
0.04425633326172829,
-0.05073338374495506,
0.012823536060750484,
-0.07073863595724106,
0.06849770992994308,
0.009020624682307243,
0.03208354115486145,
-0.07441166043281555,
0.035739898681640625,
-0.004579898435622454,
0.025656791403889656,
-0.12054412066936493,
0.11464115232229233,
0.03420566767454147,
0.03438866510987282,
-0.10358402878046036,
-0.03438727930188179,
-0.06829165667295456,
-0.09904938191175461,
-0.07391904294490814,
0.009884350933134556,
0.013980600982904434,
0.09455167502164841,
-0.03380761295557022,
0.08335813134908676,
-0.007746212650090456,
0.03542102128267288,
-0.16832852363586426,
-0.027461519464850426,
0.05281234160065651,
-0.0396796353161335,
-0.045871954411268234,
-0.023814942687749863,
0.10386546701192856,
0.05239966884255409,
-0.0021093804389238358,
0.05005374550819397,
-0.0025041368789970875,
0.05475716292858124,
-0.03179570659995079,
-0.020106734707951546,
0.02988485060632229,
0.01090280618518591,
0.025196192786097527,
0.007926499471068382,
-0.005254691932350397,
0.008449345827102661,
-0.03328065201640129,
0.03216279670596123,
0.0001108092037611641,
-0.04017022252082825,
0.0142119275406003,
0.013051696121692657,
-0.037535835057497025,
0.06414669007062912,
-0.013045335188508034,
-0.018887681886553764,
-0.047909751534461975,
-0.04884878545999527,
0.010274953208863735,
0.010298362001776695,
0.0073582082986831665,
-0.04191884770989418,
-0.052429042756557465,
0.061204444617033005,
0.10238239914178848,
0.007881549187004566,
0.01887807995080948,
-0.030408594757318497,
0.05582157522439957,
0.011688732542097569,
0.12219388782978058,
-0.056762974709272385,
0.0173944141715765,
0.001028233440592885,
-0.027793509885668755,
0.006599592976272106,
-0.04822876304388046,
0.011645913124084473,
-0.059449244290590286,
-4.399208464178628e-8,
-0.0028475604485720396,
-0.053432416170835495,
0.05577491596341133,
0.028946394100785255,
0.12676316499710083,
-0.06749292463064194,
-0.003766686422750354,
0.10848897695541382,
-0.06363505870103836,
-0.03585951775312424,
0.10858762264251709,
-0.009432844817638397,
-0.036112383008003235,
-0.010903988033533096,
-0.0427616685628891,
0.08859619498252869,
-0.01779024861752987,
0.0009704872500151396,
-0.0203816220164299,
0.030511269345879555,
-0.03679004684090614,
0.0021384158171713352,
-0.007149983663111925,
-0.023430494591593742,
0.018424201756715775,
-0.04131828621029854,
0.051979921758174896,
0.057298965752124786,
0.12130989134311676,
0.07198454439640045,
-0.03870461881160736,
-0.08131694793701172,
0.05532005801796913,
-0.011117364279925823,
-0.03585716336965561,
0.07943151891231537,
0.001261817873455584,
-0.03331828862428665,
-0.041562438011169434,
-0.0008085404988378286,
0.022432567551732063,
0.09483038634061813,
-0.0003366362943779677,
0.0806884840130806,
-0.004484101198613644,
-0.07796666771173477,
-0.007676206529140472,
-0.052989911288022995,
0.01860613003373146,
0.005056871101260185,
0.028603445738554,
0.0020576827228069305,
-0.04218141734600067,
0.09964993596076965,
0.0206738468259573,
0.0051511540077626705,
0.0025970677379518747,
-0.02051251009106636,
-0.018859632313251495,
0.08176711201667786,
0.030680930241942406,
0.004107956774532795,
-0.02204069308936596,
-0.023372294381260872
] |
SynCamMaster: Synchronizing Multi-Camera Video Generation from Diverse Viewpoints | https://openreview.net/forum?id=m8Rk3HLGFx | [
"Jianhong Bai",
"Menghan Xia",
"Xintao Wang",
"Ziyang Yuan",
"Zuozhu Liu",
"Haoji Hu",
"Pengfei Wan",
"Di ZHANG"
] | Poster | Recent advancements in video diffusion models demonstrate remarkable capabilities in simulating real-world dynamics and 3D consistency. This progress motivates us to explore the potential of these models to maintain dynamic consistency across diverse viewpoints, a feature highly sought after in applications like virtual filming. Unlike existing methods focused on multi-view generation of single objects for 4D reconstruction, our interest lies in generating open-world videos from arbitrary viewpoints, incorporating six degrees of freedom (6 DoF) camera poses.
To achieve this, we propose a plug-and-play module that enhances a pre-trained text-to-video model for multi-camera video generation, ensuring consistent content across different viewpoints. Specifically, we introduce a multi-view synchronization module designed to maintain appearance and geometry consistency across these viewpoints. Given the scarcity of high-quality training data, we also propose a progressive training scheme that leverages multi-camera images and monocular videos as a supplement to Unreal Engine-rendered multi-camera videos. This comprehensive approach significantly benefits our model.
Experimental results demonstrate the superiority of our proposed method over existing competitors and several baselines. Furthermore, our method enables intriguing extensions, such as re-rendering a video from multiple novel viewpoints. Project webpage: https://jianhongbai.github.io/SynCamMaster/ | Video Generation, Diffusion Model | A efficient method is proposed to lift pre-trained text-to-video models for open-domain multi-camera video generation from diverse viewpoints | 1,274 | 2412.07760 | [
-0.032182153314352036,
-0.10985913127660751,
-0.02614275924861431,
-0.04988311231136322,
0.08007010817527771,
0.03397085517644882,
-0.09879200905561447,
-0.07894733548164368,
0.052328743040561676,
-0.07593481987714767,
-0.02076217532157898,
-0.04138098284602165,
-0.03786400333046913,
0.0512281209230423,
-0.007091580890119076,
-0.01870799995958805,
0.055660851299762726,
0.12661801278591156,
-0.10829479247331619,
0.03631483390927315,
-0.057113584131002426,
-0.05578160658478737,
0.09462790191173553,
-0.016305265948176384,
0.048403091728687286,
0.04821860045194626,
0.06120714545249939,
-0.06574099510908127,
0.02914060465991497,
-0.005051317159086466,
0.05761152505874634,
-0.029520131647586823,
-0.05633317679166794,
0.027921171858906746,
-0.04577142000198364,
0.02896587736904621,
-0.03410525247454643,
0.013829194009304047,
-0.057378917932510376,
-0.00789080560207367,
0.0412401519715786,
0.0929201990365982,
0.019910095259547234,
0.007388940081000328,
0.04873637855052948,
0.06795287877321243,
-0.02368401549756527,
-0.006537044886499643,
0.07826335728168488,
-0.012425017543137074,
-0.04391148313879967,
-0.07611647993326187,
-0.041542790830135345,
0.002225924050435424,
-0.018047312274575233,
0.07533017545938492,
0.02702472172677517,
-0.05068151280283928,
0.0321086086332798,
0.005203971639275551,
0.07126116007566452,
0.015080190263688564,
0.039668526500463486,
0.017162593081593513,
0.010887949727475643,
-0.048708342015743256,
0.07254321873188019,
0.016856521368026733,
-0.03847448527812958,
0.021911950781941414,
-0.11278652399778366,
0.028096329420804977,
0.02179555967450142,
-0.05881945416331291,
-0.03131869062781334,
0.021197011694312096,
-0.008336315862834454,
-0.08615828305482864,
-0.029383229091763496,
-0.05611788108944893,
0.06587520986795425,
0.0033095290418714285,
0.025309570133686066,
0.03191331401467323,
-0.018079975619912148,
-0.008942068554461002,
0.0008984863525256515,
0.08191215991973877,
-0.020316243171691895,
0.009461114183068275,
-0.08802337199449539,
0.004257074557244778,
-0.012914873659610748,
-0.06398437172174454,
-0.028001999482512474,
-0.006965483073145151,
0.029456226155161858,
-0.06336040049791336,
0.0435592383146286,
0.02491847425699234,
0.020809974521398544,
-0.006321982014924288,
0.03863786533474922,
-0.013123972341418266,
0.06241193413734436,
-0.022548237815499306,
0.05975344777107239,
0.024560941383242607,
-0.029892902821302414,
0.012283052317798138,
0.07180611044168472,
-0.006412127986550331,
-0.005527555476874113,
0.03816770017147064,
0.019153783097863197,
0.05760791897773743,
-0.029821595177054405,
-0.0018389943288639188,
0.0038476844783872366,
0.04494288191199303,
0.03483650088310242,
-0.04061850905418396,
0.018550176173448563,
0.009891078807413578,
-0.0014179422287270427,
-0.021962516009807587,
0.04748130217194557,
-3.705554463957772e-34,
0.044164616614580154,
0.024185707792639732,
0.09463389962911606,
0.10179141908884048,
0.08791752904653549,
0.023818379268050194,
0.000996395479887724,
-0.006136480253189802,
-0.04546002298593521,
-0.033384744077920914,
0.03768830746412277,
-0.09614263474941254,
-0.024080945178866386,
0.10281826555728912,
0.025330090895295143,
-0.030522439628839493,
-0.04323139041662216,
0.018954792991280556,
0.02716330997645855,
0.16081084311008453,
0.038493331521749496,
-0.011223818175494671,
-0.0035216298419982195,
0.003736053593456745,
-0.07109030336141586,
0.04563279077410698,
0.05748796835541725,
0.05370454117655754,
-0.00006421427679015324,
-0.0361112542450428,
-0.0806780755519867,
-0.028958994895219803,
0.02471023239195347,
-0.036846455186605453,
0.07663624733686447,
-0.02481660060584545,
-0.06311200559139252,
-0.06860990822315216,
-0.06452201306819916,
-0.024237483739852905,
-0.02533758245408535,
0.06221005693078041,
-0.06455140560865402,
-0.08383972197771072,
-0.07487978041172028,
0.04225921258330345,
0.0541287325322628,
0.08752551674842834,
-0.09549965709447861,
0.0234793983399868,
0.060495391488075256,
0.0280908290296793,
-0.005432664882391691,
-0.13477671146392822,
-0.09479115903377533,
-0.07915566116571426,
0.009769591502845287,
-0.03762006759643555,
-0.002289536641910672,
0.005231836810708046,
0.01465674676001072,
0.0034075567964464426,
-0.10196614265441895,
0.019619537517428398,
-0.023055963218212128,
0.007102579344063997,
0.05115320906043053,
0.03187667578458786,
0.07077626883983612,
0.055025164037942886,
-0.02053803578019142,
-0.0020959400571882725,
-0.061789873987436295,
-0.0847654864192009,
0.05526711419224739,
0.04644932597875595,
-0.01283822488039732,
-0.04457459598779678,
0.03455847501754761,
0.04835261404514313,
-0.06007056683301926,
-0.011221544817090034,
0.056209176778793335,
-0.029781626537442207,
-0.041811101138591766,
-0.07771213352680206,
-0.03058980219066143,
-0.024619147181510925,
-0.006773028988391161,
-0.02875535748898983,
-0.05186406895518303,
-0.07662966102361679,
0.0053900666534900665,
-0.06256067007780075,
0.05384945869445801,
7.524346686453155e-34,
0.08348031342029572,
0.011362956836819649,
-0.055667318403720856,
0.032734084874391556,
0.05099603533744812,
-0.005860080011188984,
0.013517079874873161,
0.07284092903137207,
0.026058930903673172,
-0.10868987441062927,
-0.024157479405403137,
-0.03793540224432945,
-0.017985068261623383,
-0.0009514156845398247,
-0.0074692461639642715,
0.03974832594394684,
0.013777256943285465,
-0.027858005836606026,
-0.07430372387170792,
-0.008130948059260845,
0.09171693027019501,
-0.013467962853610516,
0.03620680421590805,
-0.0012734014308080077,
-0.02364886924624443,
0.0970393717288971,
-0.008458566851913929,
0.031237848103046417,
0.022760620340704918,
-0.0022293825168162584,
0.0639284998178482,
-0.029352983459830284,
0.040095485746860504,
-0.01535998284816742,
-0.1381537914276123,
0.06946314871311188,
0.051477011293172836,
0.017918778583407402,
-0.03441975265741348,
0.07847797870635986,
0.024705838412046432,
-0.07179097831249237,
-0.015959730371832848,
0.05007942020893097,
-0.028301047161221504,
0.02265312895178795,
-0.07160039246082306,
0.029298216104507446,
-0.04325200989842415,
0.08969826251268387,
-0.10993999987840652,
0.08137353509664536,
-0.06263214349746704,
-0.03724530711770058,
0.004065455868840218,
-0.10096254199743271,
-0.09919612109661102,
-0.03837788105010986,
0.04844396188855171,
-0.006155833136290312,
-0.01472910400480032,
-0.06594955921173096,
-0.030841641128063202,
0.014246693812310696,
0.004963917192071676,
0.019312864169478416,
0.00005949742262600921,
-0.03511034697294235,
-0.02856365591287613,
0.05939454957842827,
0.021550912410020828,
0.010241571813821793,
-0.017782142385840416,
-0.009545463137328625,
0.024381224066019058,
-0.026725882664322853,
0.0006122050108388066,
0.003241626312956214,
0.11847924441099167,
-0.032446444034576416,
0.004126101732254028,
-0.07652217149734497,
0.05000137537717819,
0.13680239021778107,
0.07523129135370255,
0.048060107976198196,
0.06312558054924011,
0.05144449695944786,
0.028040040284395218,
0.07960491627454758,
-0.011484072543680668,
0.03940124809741974,
-0.019081898033618927,
0.03154570609331131,
-0.007727688644081354,
-3.676465354374159e-8,
-0.11754057556390762,
0.001026158919557929,
-0.039662424474954605,
0.04868816211819649,
-0.11869850754737854,
-0.08862040191888809,
0.03483428806066513,
0.034425295889377594,
0.030695129185914993,
-0.050532761961221695,
0.014527029357850552,
-0.028121402487158775,
-0.01344473659992218,
0.0066880276426672935,
-0.0027250172570347786,
0.023945502936840057,
-0.04847586899995804,
-0.003727269358932972,
-0.03355816379189491,
0.04939071089029312,
-0.020469212904572487,
0.04111851751804352,
0.028348714113235474,
-0.035524025559425354,
0.03292379155755043,
0.024808675050735474,
-0.02459796704351902,
0.07865547388792038,
0.12554782629013062,
-0.0579526349902153,
0.02457062527537346,
-0.037276122719049454,
-0.017649684101343155,
-0.07226194441318512,
-0.05068587884306908,
-0.005042199976742268,
-0.009150231257081032,
-0.019513223320245743,
0.04984612390398979,
-0.04254313185811043,
-0.025607561692595482,
0.05481782183051109,
0.050649117678403854,
0.04105283319950104,
0.030332723632454872,
0.05357864499092102,
0.029484085738658905,
-0.0688415914773941,
0.017933402210474014,
-0.038873378187417984,
-0.10841777920722961,
-0.014398743398487568,
-0.0003091269754804671,
0.08028799295425415,
0.07416243106126785,
0.005229340400546789,
0.07424270361661911,
-0.011082342825829983,
0.050524912774562836,
0.07359926402568817,
0.017983093857765198,
-0.0023838416673243046,
-0.016003046184778214,
0.03440997377038002
] |
CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models | https://openreview.net/forum?id=E77uvbOTtp | [
"Hyungjin Chung",
"Jeongsol Kim",
"Geon Yeong Park",
"Hyelin Nam",
"Jong Chul Ye"
] | Poster | Classifier-free guidance (CFG) is a fundamental tool in modern diffusion models for text-guided generation. Although effective, CFG has notable drawbacks. For instance, DDIM with CFG lacks invertibility, complicating image editing; furthermore, high guidance scales, essential for high-quality outputs, frequently result in issues like mode collapse. Contrary to the widespread belief that these are inherent limitations of diffusion models, this paper reveals that the problems actually stem from the off-manifold phenomenon associated with CFG, rather than the diffusion models themselves. More specifically, inspired by the recent advancements of diffusion model-based inverse problem solvers (DIS), we reformulate text-guidance as an inverse problem with a text-conditioned score matching loss and develop CFG++, a novel approach that tackles the off-manifold challenges inherent in traditional CFG. CFG++ features a surprisingly simple fix to CFG, yet it offers significant improvements, including better sample quality for text-to-image generation, invertibility, smaller guidance scales, reduced etc. Furthermore, CFG++ enables seamless interpolation between unconditional and conditional sampling at lower guidance scales, consistently outperforming traditional CFG at all scales. Moreover, CFG++ can be easily integrated into the high-order diffusion solvers and naturally extends to distilled diffusion models. Experimental results confirm that our method significantly enhances performance in text-to-image generation, DDIM inversion, editing, and solving inverse problems, suggesting a wide-ranging impact and potential applications in various fields that utilize text guidance. Project Page: https://cfgpp-diffusion.github.io/anon | Diffusion models, Manifold, Classifier-free guidance | null | 1,270 | 2406.08070 | [
-0.05473197251558304,
-0.008242622017860413,
0.07217548042535782,
0.11355157941579819,
0.09679006040096283,
-0.025258822366595268,
-0.07956080883741379,
0.004924894776195288,
0.02703726850450039,
-0.02875569276511669,
-0.030795443803071976,
0.007797393016517162,
0.032117925584316254,
0.0495515838265419,
-0.032996732741594315,
-0.010732162743806839,
0.0283681470900774,
0.00663340138271451,
-0.128090500831604,
-0.05173290893435478,
-0.04863700270652771,
0.04808824509382248,
-0.022265378385782242,
-0.028808576986193657,
0.024522356688976288,
-0.012352852150797844,
0.03175848349928856,
-0.030507052317261696,
-0.012969971634447575,
-0.015312143601477146,
0.03479044511914253,
0.003849263535812497,
-0.018475502729415894,
-0.0026100606191903353,
-0.06087317317724228,
0.11239936202764511,
0.01124426070600748,
0.035933539271354675,
-0.03297644481062889,
-0.032033808529376984,
0.004474097862839699,
0.024748682975769043,
0.051490992307662964,
0.07835466414690018,
0.13947971165180206,
-0.033085957169532776,
-0.08778367936611176,
-0.017032986506819725,
0.0011652472894638777,
0.02539358288049698,
-0.0369189977645874,
-0.01235014945268631,
-0.11377008259296417,
0.010745507664978504,
-0.03915826976299286,
0.07765254378318787,
0.05828039348125458,
0.0249920841306448,
0.04085467755794525,
-0.020694853737950325,
-0.04571089893579483,
-0.09243645519018173,
-0.10028921812772751,
-0.009291091002523899,
-0.016548462212085724,
0.008602851070463657,
0.014488591812551022,
0.05654531717300415,
-0.04922281950712204,
0.1091655045747757,
-0.0629783645272255,
0.0012101831380277872,
-0.020294686779379845,
0.019244913011789322,
-0.0007867299718782306,
0.020822906866669655,
0.003649383084848523,
-0.01923029124736786,
0.03906374052166939,
-0.0002287567185703665,
0.08700994402170181,
-0.020262762904167175,
0.10704941302537918,
0.00908742193132639,
0.07140140235424042,
-0.02900581620633602,
-0.014846673235297203,
0.0005091380444355309,
0.021674850955605507,
0.004367918707430363,
0.005003561265766621,
-0.07923199981451035,
-0.06283547729253769,
-0.03525300696492195,
-0.060160081833601,
-0.010759107768535614,
-0.029483743011951447,
-0.055673446506261826,
0.02630169689655304,
0.1053476482629776,
-0.03447432816028595,
-0.010006137192249298,
-0.025323251262307167,
-0.023836536332964897,
-0.0024046837352216244,
-0.04649635776877403,
0.09249759465456009,
0.07902515679597855,
0.008206712082028389,
-0.07118325680494308,
0.07602500915527344,
-0.024894677102565765,
0.02030584029853344,
-0.009352334775030613,
0.08064191043376923,
-0.10837855935096741,
0.02893519401550293,
-0.042143311351537704,
0.05702286586165428,
0.04789971187710762,
-0.07544341683387756,
0.003413187572732568,
-0.06288275867700577,
-0.013796822167932987,
-0.005583639722317457,
-0.0058660428039729595,
0.010557230561971664,
5.218928962054083e-33,
-0.02455420233309269,
0.007646393962204456,
0.05984089523553848,
0.06290514767169952,
0.0904378741979599,
0.020283015444874763,
-0.047772690653800964,
-0.04325241968035698,
-0.04853260889649391,
0.03214181959629059,
-0.005427888128906488,
-0.037210315465927124,
-0.10500849783420563,
0.11400796473026276,
0.041120171546936035,
0.0024247190449386835,
0.020378539338707924,
0.06161437928676605,
-0.07407363504171371,
-0.039567239582538605,
0.017611010000109673,
-0.02893650345504284,
-0.0014041656395420432,
-0.09789963066577911,
0.02639882080256939,
0.018235648050904274,
0.045273736119270325,
-0.08426476269960403,
0.02000531181693077,
-0.020898528397083282,
-0.11640240997076035,
-0.03601772338151932,
0.007318931166082621,
0.06939715892076492,
-0.04936496913433075,
0.04450492933392525,
-0.043758854269981384,
-0.053733427077531815,
-0.05149838700890541,
-0.027463214471936226,
-0.031070344150066376,
0.08995561301708221,
0.007552371360361576,
-0.0022959362249821424,
-0.1241229921579361,
0.0468815341591835,
0.03660547360777855,
-0.008915425278246403,
-0.062092188745737076,
-0.04808267205953598,
0.00010736126569099724,
-0.0286649651825428,
-0.048958294093608856,
-0.092878058552742,
-0.02885490283370018,
-0.03237014263868332,
0.03480325639247894,
-0.004044626373797655,
0.016613049432635307,
0.024116432294249535,
0.09990877658128738,
0.04034598171710968,
-0.0507429838180542,
0.016767002642154694,
0.03416092321276665,
0.009934617206454277,
-0.006596007850021124,
0.00908659491688013,
0.006643503904342651,
0.015800027176737785,
-0.036842018365859985,
0.003044001292437315,
-0.03129302337765694,
-0.020423494279384613,
0.08204882591962814,
0.00012497841089498252,
0.04929133132100105,
-0.06441275775432587,
0.033789485692977905,
-0.01530715636909008,
-0.08301732689142227,
-0.047255873680114746,
0.020612992346286774,
-0.08533736318349838,
-0.05507509037852287,
-0.05818697437644005,
0.027158422395586967,
-0.04736578091979027,
-0.05222989618778229,
-0.05854317173361778,
-0.047394558787345886,
0.030492698773741722,
-0.04600575193762779,
0.1044330820441246,
0.05113222077488899,
-3.2074920892200954e-33,
-0.0024959249421954155,
0.06748969107866287,
0.01955804042518139,
0.009859582409262657,
-0.02265055850148201,
0.030569924041628838,
0.06004161387681961,
0.06161714717745781,
0.11517869681119919,
-0.11026202142238617,
-0.021952735260128975,
0.011495484970510006,
0.040285881608724594,
0.01730426773428917,
-0.06717414408922195,
-0.023226110264658928,
-0.03876229375600815,
-0.0061098737642169,
-0.03552461415529251,
0.05649933964014053,
-0.047327350825071335,
0.07196133583784103,
-0.020817270502448082,
0.048073917627334595,
-0.06977692991495132,
0.05711016431450844,
-0.01652257703244686,
0.09109457582235336,
0.004651986062526703,
-0.004755209665745497,
-0.03294895961880684,
0.02300001122057438,
0.03073222003877163,
-0.03941383212804794,
-0.098390132188797,
0.12051943689584732,
-0.035768359899520874,
-0.006372001487761736,
-0.046831224113702774,
0.13491030037403107,
0.039293937385082245,
0.034193940460681915,
-0.07921194285154343,
0.026799317449331284,
0.0333796851336956,
-0.01989223249256611,
0.0422334186732769,
-0.02115032821893692,
0.06252188235521317,
0.04564864560961723,
-0.0501326285302639,
0.038246337324380875,
-0.055021509528160095,
0.04189794883131981,
-0.06664786487817764,
-0.06360936164855957,
-0.10452780872583389,
-0.05850207805633545,
0.07464808225631714,
0.029307547956705093,
-0.01011496689170599,
0.017752185463905334,
0.06114547327160835,
-0.028036678209900856,
0.028060687705874443,
0.024891536682844162,
-0.005616610404103994,
0.04307843744754791,
-0.020590290427207947,
0.1106979101896286,
-0.01322779431939125,
0.01228271797299385,
0.019521960988640785,
-0.016169331967830658,
0.018557725474238396,
-0.05664808303117752,
0.020356889814138412,
0.01844523288309574,
-0.025820059701800346,
-0.049367599189281464,
0.05231770500540733,
0.04761502519249916,
0.0006909220246598125,
0.07913611829280853,
0.02077765390276909,
-0.0034464409109205008,
0.04519558697938919,
0.009083791635930538,
0.036475446075201035,
0.061934761703014374,
0.00009742168185766786,
-0.0016742886509746313,
0.06666731834411621,
0.07194766402244568,
-0.04724590852856636,
-4.706162215484255e-8,
-0.09815274924039841,
0.0392148457467556,
-0.003778791520744562,
0.005000952631235123,
-0.00665361899882555,
0.004541254136711359,
0.09858853369951248,
0.021131150424480438,
0.0017655924893915653,
-0.0035246326588094234,
-0.008657262660562992,
-0.01378603558987379,
-0.05944664776325226,
-0.055802661925554276,
-0.09584197402000427,
0.07776220887899399,
0.01920054666697979,
0.05692342296242714,
0.009425419382750988,
-0.002098550321534276,
0.02635449357330799,
0.020475879311561584,
0.0018795537762343884,
-0.019062520936131477,
0.10140447318553925,
-0.028047733008861542,
-0.059892091900110245,
0.01866074651479721,
0.09609650820493698,
-0.010665570385754108,
-0.021896759048104286,
0.005408414639532566,
0.01709386706352234,
0.07681676000356674,
0.02493223175406456,
0.09629938751459122,
-0.010893110185861588,
-0.029573626816272736,
-0.0826776772737503,
-0.025459587574005127,
-0.03232591971755028,
0.08001402020454407,
-0.048986371606588364,
0.004168324172496796,
0.027938226237893105,
0.055388178676366806,
0.14276331663131714,
-0.06162457540631294,
0.004242823459208012,
-0.0005853289039805532,
0.056661028414964676,
-0.06936752051115036,
0.013596619479358196,
0.004277585539966822,
0.09390201419591904,
0.021484242752194405,
0.04905621334910393,
0.054389119148254395,
0.03751073405146599,
0.045650627464056015,
-0.06420287489891052,
-0.02216072380542755,
0.007803612854331732,
-0.021127745509147644
] |
TDDBench: A Benchmark for Training data detection | https://openreview.net/forum?id=hpeyWG1PP6 | [
"Zhihao Zhu",
"Yi Yang",
"Defu Lian"
] | Poster | Training Data Detection (TDD) is a task aimed at determining whether a specific data instance is used to train a machine learning model. In the computer security literature, TDD is also referred to as Membership Inference Attack (MIA). Given its potential to assess the risks of training data breaches, ensure copyright authentication, and verify model unlearning, TDD has garnered significant attention in recent years, leading to the development of numerous methods. Despite these advancements, there is no comprehensive benchmark to thoroughly evaluate the effectiveness of TDD methods.
In this work, we introduce TDDBench, which consists of 13 datasets spanning three data modalities: image, tabular, and text. We benchmark 21 different TDD methods across four detection paradigms and evaluate their performance from five perspectives: average detection performance, best detection performance, memory consumption, and computational efficiency in both time and memory. With TDDBench, researchers can identify bottlenecks and areas for improvement in TDD algorithms, while practitioners can make informed trade-offs between effectiveness and efficiency when selecting TDD algorithms for specific use cases. Our extensive experiments also reveal the generally unsatisfactory performance of TDD algorithms across different datasets. To enhance accessibility and reproducibility, we open-source TDDBench for the research community at https://github.com/zzh9568/TDDBench. | Training data detection; Benchmark; Copyright certification | null | 1,267 | 2411.03363 | [
-0.059101276099681854,
-0.06296099722385406,
-0.01755933277308941,
0.022562293335795403,
0.14880019426345825,
0.02615351229906082,
0.043431274592876434,
-0.047556519508361816,
0.013422270305454731,
-0.005512944422662258,
-0.0075419251807034016,
-0.030815571546554565,
0.1187819391489029,
0.010696619749069214,
-0.05730734020471573,
-0.033680349588394165,
0.016209011897444725,
0.08010987937450409,
-0.09805269539356232,
-0.07089773565530777,
-0.07771109789609909,
-0.06106409430503845,
0.035532236099243164,
0.013801474124193192,
0.011074417270720005,
-0.0006421764264814556,
0.02636948600411415,
-0.061032164841890335,
-0.06598535180091858,
-0.03903421387076378,
0.05298912897706032,
-0.027794459834694862,
-0.04429342970252037,
0.062430769205093384,
-0.01898791454732418,
-0.03683589771389961,
0.02008301205933094,
-0.01920662634074688,
0.07347682118415833,
-0.020576056092977524,
-0.06054699420928955,
-0.06849872320890427,
0.015404203906655312,
0.042274706065654755,
0.022765416651964188,
0.031099369749426842,
-0.003353782696649432,
-0.006095359101891518,
-0.02481117844581604,
-0.0050381869077682495,
0.021221084520220757,
0.026472361758351326,
-0.022249823436141014,
0.05476200953125954,
-0.06446310132741928,
-0.009154248051345348,
0.046379879117012024,
0.029234420508146286,
0.014102494344115257,
0.10094455629587173,
-0.05959150940179825,
-0.05745512247085571,
-0.013689792715013027,
-0.007290789857506752,
-0.05745639279484749,
-0.013150072656571865,
-0.013717608526349068,
0.011749922297894955,
0.04659952595829964,
-0.035824503749608994,
0.011195909231901169,
0.1278998702764511,
-0.05506070703268051,
0.04799606651067734,
-0.01354247983545065,
0.0018732513999566436,
0.03506599739193916,
-0.028730399906635284,
0.09011951833963394,
-0.16195373237133026,
-0.011322407983243465,
0.023215480148792267,
0.049688294529914856,
0.028193868696689606,
0.12157072126865387,
0.0227137990295887,
0.009474354796111584,
0.03980886936187744,
-0.01564493402838707,
0.04830377548933029,
0.026535114273428917,
0.022775130346417427,
0.023082835599780083,
0.022335225716233253,
-0.03717561438679695,
0.01908375881612301,
-0.01988663338124752,
-0.013435312546789646,
-0.017190156504511833,
0.06693685054779053,
-0.09508561342954636,
0.087002232670784,
0.015757890418171883,
-0.04592621698975563,
0.058412227779626846,
-0.0021946392953395844,
0.13304679095745087,
-0.045274242758750916,
0.10067752748727798,
-0.054918985813856125,
0.0369262620806694,
0.06303434818983078,
-0.10073201358318329,
-0.014871411956846714,
0.02755969576537609,
0.010797242633998394,
-0.10161858052015305,
-0.030193716287612915,
0.0022494953591376543,
0.05452949181199074,
0.04222124069929123,
-0.013975732959806919,
0.011150776408612728,
0.07832469791173935,
-0.031964030116796494,
-0.04725175350904465,
-0.11195513606071472,
5.0768546740518704e-33,
-0.027376223355531693,
0.038421858102083206,
0.0018223850056529045,
-0.026497961953282356,
0.04994535446166992,
-0.06489238142967224,
-0.03370360657572746,
0.06691734492778778,
-0.0587439090013504,
0.045514676719903946,
-0.12443943321704865,
-0.038749366998672485,
-0.03203340247273445,
0.04001791402697563,
0.11667397618293762,
0.07394233345985413,
-0.1104334220290184,
0.09270084649324417,
-0.08819940686225891,
-0.014837777242064476,
0.0610688142478466,
-0.08111602067947388,
0.03447172790765762,
0.037444792687892914,
0.03949074074625969,
-0.09264631569385529,
-0.021340245380997658,
0.04480321332812309,
0.0653926357626915,
0.04586084187030792,
-0.04637858271598816,
-0.006704046856611967,
0.013706476427614689,
0.05066346377134323,
0.06558433920145035,
-0.0626809298992157,
0.031319037079811096,
0.02455689013004303,
-0.008234995417296886,
0.04485785588622093,
-0.027611492201685905,
-0.0060784476809203625,
0.048426203429698944,
-0.06531085073947906,
-0.0329933688044548,
-0.05670221149921417,
-0.02695990540087223,
-0.016482656821608543,
-0.04181979224085808,
-0.019939778372645378,
0.03550715371966362,
-0.022688671946525574,
0.010661245323717594,
-0.03247949853539467,
-0.04740353673696518,
-0.027298083528876305,
0.00867817085236311,
0.0072550103068351746,
0.13569509983062744,
0.07559528201818466,
-0.014574749395251274,
0.00928614754229784,
-0.020737074315547943,
0.08182570338249207,
-0.08779921382665634,
-0.030052054673433304,
0.019027085974812508,
-0.042585305869579315,
0.02574433572590351,
-0.017904246225953102,
0.011586776934564114,
0.008360602892935276,
0.010966149158775806,
-0.03029526397585869,
0.03791423514485359,
-0.04366995766758919,
0.09715588390827179,
0.04706461355090141,
-0.03694146126508713,
-0.0003199674829374999,
0.027797061949968338,
-0.046751245856285095,
0.0455041229724884,
-0.051986727863550186,
-0.08976928889751434,
-0.039542440325021744,
-0.010777860879898071,
-0.11730151623487473,
-0.047140397131443024,
0.07004179060459137,
-0.04186524823307991,
-0.015672555193305016,
-0.04119191691279411,
-0.04644482210278511,
0.018243299797177315,
-4.195047668548946e-33,
0.007193608675152063,
0.06993243843317032,
0.005609283689409494,
-0.01845775917172432,
-0.010523204691708088,
-0.04594958573579788,
-0.03341444954276085,
0.04051390662789345,
0.006394149735569954,
-0.08413016051054001,
0.009501121006906033,
-0.004965823609381914,
-0.023189116269350052,
-0.040256317704916,
-0.017449207603931427,
0.04065614193677902,
-0.018354251980781555,
0.014969835057854652,
-0.07880695909261703,
0.023218175396323204,
-0.026388654485344887,
0.09852079302072525,
-0.07386478781700134,
0.0016892376588657498,
0.04456927627325058,
0.032466839998960495,
-0.06700070947408676,
0.049393948167562485,
0.04671870172023773,
0.04871445521712303,
-0.07503525167703629,
0.033555302768945694,
-0.030213268473744392,
-0.009963799268007278,
-0.04392391815781593,
0.01645492948591709,
0.11527068167924881,
0.00939578004181385,
0.02383832447230816,
0.07947609573602676,
0.06364361196756363,
0.062250103801488876,
-0.05742804333567619,
0.02099386416375637,
-0.033457402139902115,
-0.029698446393013,
0.005734474863857031,
0.001635579508729279,
0.03979327902197838,
0.02060590870678425,
-0.00564415892586112,
-0.06553755700588226,
0.04330183193087578,
-0.0024092206731438637,
0.017270753160119057,
-0.024192029610276222,
-0.010147333145141602,
0.053126443177461624,
0.028980690985918045,
-0.00021939277939964086,
0.029794320464134216,
0.03523758798837662,
-0.02272169478237629,
0.016248485073447227,
-0.014130455441772938,
-0.06186080724000931,
-0.06466741114854813,
0.10875484347343445,
-0.07165355235338211,
0.053895097225904465,
-0.0461001843214035,
-0.06720147281885147,
0.004430451896041632,
0.05345925688743591,
-0.049306999891996384,
-0.020619919523596764,
-0.0013382612960413098,
-0.010586138814687729,
-0.0724407210946083,
0.049367886036634445,
0.058453820645809174,
-0.055819351226091385,
-0.03778677433729172,
-0.00019860942848026752,
0.09523990750312805,
0.11261267215013504,
0.0382215790450573,
-0.041065018624067307,
0.024764349684119225,
0.0036740426439791918,
-0.07974016666412354,
-0.01666419580578804,
-0.04366876184940338,
0.06451843678951263,
0.04545700177550316,
-4.527985808522317e-8,
-0.01036786288022995,
-0.01797986403107643,
-0.005981734488159418,
-0.06567802280187607,
0.05823475867509842,
0.016333958134055138,
-0.07445083558559418,
0.14194460213184357,
-0.013840833678841591,
-0.06893821805715561,
0.08207891136407852,
-0.013010546565055847,
-0.04770718514919281,
-0.10708604753017426,
-0.013947616331279278,
0.03193924203515053,
-0.05025513470172882,
-0.06556834280490875,
-0.03461470454931259,
0.011443790048360825,
-0.019862085580825806,
0.03886856883764267,
0.047632597386837006,
-0.06415324658155441,
-0.02423737198114395,
-0.01939331740140915,
-0.045986779034137726,
0.06775469332933426,
0.044761352241039276,
-0.030655819922685623,
-0.04082416743040085,
0.011716531589627266,
-0.0009915763512253761,
-0.01042225956916809,
0.08231944590806961,
0.12808911502361298,
-0.03790043294429779,
-0.01204763725399971,
-0.04150671511888504,
0.024305077269673347,
0.021260451525449753,
0.0042810700833797455,
0.014292729087173939,
-0.002949762623757124,
-0.00837672222405672,
0.01475355215370655,
0.06891842186450958,
-0.012693566270172596,
0.039010364562273026,
-0.018797416239976883,
0.023092422634363174,
-0.03524213656783104,
-0.022905567660927773,
0.12077841907739639,
-0.05118453502655029,
0.03197333589196205,
0.013606609776616096,
-0.03395365551114082,
0.04615778848528862,
0.0524415522813797,
0.03910082206130028,
-0.040825940668582916,
0.0028415739070624113,
0.003496468299999833
] |
Warm Diffusion: Recipe for Blur-Noise Mixture Diffusion Models | https://openreview.net/forum?id=rdSVgnLHQB | [
"Hao-Chien Hsueh",
"Wen-Hsiao Peng",
"Ching-Chun Huang"
] | Poster | Diffusion probabilistic models have achieved remarkable success in generative tasks across diverse data types. While recent studies have explored alternative degradation processes beyond Gaussian noise, this paper bridges two key diffusion paradigms: hot diffusion, which relies entirely on noise, and cold diffusion, which uses only blurring without noise. We argue that hot diffusion fails to exploit the strong correlation between high-frequency image detail and low-frequency structures, leading to random behaviors in the early steps of generation. Conversely, while cold diffusion leverages image correlations for prediction, it neglects the role of noise (randomness) in shaping the data manifold, resulting in out-of-manifold issues and partially explaining its performance drop. To integrate both strengths, we propose Warm Diffusion, a unified Blur-Noise Mixture Diffusion Model (BNMD), to control blurring and noise jointly. Our divide-and-conquer strategy exploits the spectral dependency in images, simplifying score model estimation by disentangling the denoising and deblurring processes. We further analyze the Blur-to-Noise Ratio (BNR) using spectral analysis to investigate the trade-off between model learning dynamics and changes in the data manifold. Extensive experiments across benchmarks validate the effectiveness of our approach for image generation. | Diffusion probabilistic models, Image generation | null | 1,263 | null | [
-0.027538714930415154,
-0.12645095586776733,
0.07113883644342422,
0.10094442963600159,
0.07322786748409271,
-0.03234788775444031,
-0.004889028146862984,
-0.14154964685440063,
0.04261019825935364,
-0.02670137956738472,
-0.03086603805422783,
0.023732135072350502,
0.06207654997706413,
0.03727709874510765,
-0.000848502095323056,
-0.028404148295521736,
0.02271598018705845,
0.04269525781273842,
-0.1200929507613182,
-0.06444243341684341,
-0.015950027853250504,
-0.017870988696813583,
-0.0779135450720787,
0.02406616136431694,
0.10686219483613968,
-0.00847858190536499,
0.06217791140079498,
-0.05511496588587761,
0.006248035002499819,
-0.01677843928337097,
0.06578439474105835,
-0.008649363182485104,
0.025074593722820282,
-0.019930072128772736,
-0.06761511415243149,
0.013842384330928326,
-0.007883468642830849,
0.03223571553826332,
0.03845343738794327,
0.0008572710794396698,
-0.02726236917078495,
0.05293358489871025,
-0.036822039633989334,
0.03874969482421875,
-0.0077917082235217094,
-0.0026381586212664843,
0.028558854013681412,
-0.07926984876394272,
0.019972626119852066,
0.0783650204539299,
-0.03632587194442749,
0.02513783983886242,
-0.06589888036251068,
0.05471266433596611,
-0.016974961385130882,
0.07331547141075134,
0.03829246759414673,
0.010919040068984032,
0.05167442187666893,
0.026419227942824364,
-0.052215203642845154,
-0.051340848207473755,
-0.02478947676718235,
0.007389793638139963,
0.053623661398887634,
0.007811013609170914,
0.07181772589683533,
0.019441595301032066,
-0.03369426354765892,
0.048465240746736526,
-0.06372567266225815,
0.06055692583322525,
-0.005483540706336498,
-0.010070491582155228,
-0.009922267869114876,
0.0051925270818173885,
0.06673268973827362,
-0.10236701369285583,
0.028831325471401215,
-0.011150467209517956,
0.08628621697425842,
0.0014204607577994466,
0.05978313088417053,
-0.07698490470647812,
0.052378833293914795,
-0.021639516577124596,
0.101478710770607,
-0.02107495255768299,
-0.0674615427851677,
-0.048479411751031876,
-0.06467185169458389,
0.05101579800248146,
-0.07830635458230972,
-0.008962704800069332,
0.006125165149569511,
0.07229014486074448,
0.009862511418759823,
-0.04309188947081566,
0.05945167317986488,
0.04816463217139244,
-0.009857595898211002,
-0.06666357815265656,
-0.040235184133052826,
0.02229113131761551,
0.07004249840974808,
-0.03323853760957718,
0.0902528464794159,
0.024723254144191742,
0.006614174228161573,
0.022735513746738434,
0.06098285689949989,
0.05936754122376442,
-0.010641946457326412,
0.04599139839410782,
0.01916714943945408,
-0.04048692435026169,
-0.010351591743528843,
-0.01133911032229662,
-0.01551932655274868,
0.04164151847362518,
0.002701801247894764,
0.012736666016280651,
-0.09453976899385452,
-0.01271168701350689,
0.0022460666950792074,
-0.0012924104230478406,
-0.04244357347488403,
2.195868079387166e-33,
0.03182864934206009,
-0.0694516971707344,
0.0099567212164402,
0.07057039439678192,
0.07007992267608643,
-0.05301744118332863,
-0.05440767481923103,
-0.020753709599375725,
-0.00826277770102024,
0.014370420016348362,
0.010911541990935802,
-0.008127572946250439,
-0.0785006508231163,
0.07567641884088516,
0.06739421933889389,
-0.03000921756029129,
-0.05593502148985863,
0.0557473860681057,
-0.029070714488625526,
0.01665358804166317,
-0.024964699521660805,
0.029782237485051155,
0.0006141539197415113,
0.016996966674923897,
0.028220392763614655,
0.0445302315056324,
0.0878823772072792,
-0.053248245269060135,
0.039937231689691544,
0.017114965245127678,
-0.12715472280979156,
-0.022058192640542984,
-0.05844637751579285,
0.029774446040391922,
0.01150309294462204,
-0.03930876404047012,
-0.018439237028360367,
0.06247711181640625,
-0.021034887060523033,
-0.017125627025961876,
0.0046560741029679775,
0.08005169779062271,
-0.05081307142972946,
-0.035998836159706116,
-0.10610286146402359,
-0.005163318943232298,
0.06525618582963943,
0.025396177545189857,
-0.043624792248010635,
-0.0018003091681748629,
-0.010755506344139576,
-0.06770193576812744,
0.031567979604005814,
-0.029496969655156136,
-0.06758248060941696,
-0.006394818890839815,
0.10726086050271988,
-0.10305419564247131,
0.06933373212814331,
-0.02425343543291092,
0.042599424719810486,
0.0018906880868598819,
-0.013523240573704243,
-0.027087420225143433,
0.05065200477838516,
0.006361384876072407,
-0.01129832025617361,
0.06978046894073486,
-0.026821505278348923,
-0.07189694046974182,
0.028310487046837807,
0.07442940026521683,
-0.06090676784515381,
-0.10139448195695877,
0.0679750069975853,
-0.002451889682561159,
0.022909866645932198,
-0.0032406155951321125,
0.0012503719190135598,
0.03354160115122795,
-0.07124528288841248,
-0.0610605850815773,
-0.051413796842098236,
-0.09856867045164108,
-0.05216435715556145,
0.03571745380759239,
-0.02224024012684822,
-0.07379041612148285,
-0.10748086869716644,
-0.10456390678882599,
-0.06794216483831406,
-0.040006015449762344,
0.01454162411391735,
0.0027288601268082857,
-0.02690526470541954,
-1.5855708815654197e-33,
-0.028269125148653984,
0.08188154548406601,
0.06538127362728119,
0.10146396607160568,
-0.02457754872739315,
-0.029069364070892334,
0.05106990784406662,
0.036663539707660675,
0.06859853863716125,
-0.011418121866881847,
0.00002705291444726754,
-0.05482170358300209,
-0.04526793211698532,
-0.047965455800294876,
-0.02071441151201725,
-0.054107457399368286,
0.015651600435376167,
-0.01774105802178383,
-0.09637913107872009,
0.052998680621385574,
-0.028475284576416016,
0.03488179296255112,
-0.004793982952833176,
0.01378559973090887,
-0.028200553730130196,
0.06481309235095978,
-0.06543460488319397,
0.08530191332101822,
0.04859942942857742,
-0.013655979186296463,
-0.053415898233652115,
0.00730058504268527,
-0.003055551555007696,
-0.08368076384067535,
-0.115637868642807,
0.12995082139968872,
0.01660073548555374,
-0.006099335383623838,
-0.021873464807868004,
0.054116807878017426,
0.058834508061409,
-0.0053203413262963295,
-0.0590517483651638,
0.05682138726115227,
0.046190135180950165,
-0.05681300535798073,
-0.011222751811146736,
-0.02606993354856968,
0.00047722848830744624,
0.05181156471371651,
0.031525321304798126,
0.1078479140996933,
-0.09338783472776413,
0.09018529206514359,
-0.07293113321065903,
-0.06831341236829758,
-0.12516449391841888,
0.0018490793881937861,
0.07102374732494354,
0.0590328611433506,
0.007310886401683092,
-0.02339886501431465,
-0.062697634100914,
-0.057260990142822266,
-0.08637198805809021,
-0.0009463614551350474,
-0.02232293412089348,
-0.026926938444375992,
0.0371360182762146,
0.10699830949306488,
0.005687955766916275,
0.012169984169304371,
0.027626506984233856,
-0.012813874520361423,
0.03873314708471298,
-0.06890869140625,
0.033065833151340485,
-0.027026595547795296,
0.025026420131325722,
0.02950337342917919,
0.06813414394855499,
-0.029468821361660957,
-0.008700101636350155,
-0.003630138700827956,
0.03871757164597511,
0.018215512856841087,
-0.0020067833829671144,
-0.06850726902484894,
0.03539145737886429,
-0.00837723258882761,
-0.023297468200325966,
-0.06033284589648247,
-0.012854621745646,
0.06592702120542526,
-0.05327383428812027,
-4.49857076034732e-8,
-0.013193375431001186,
0.009836604818701744,
0.03570224344730377,
-0.03462452068924904,
0.054119233042001724,
-0.0706891342997551,
0.05725661292672157,
0.08948435634374619,
0.06550971418619156,
-0.05150190740823746,
0.11167559772729874,
-0.027588417753577232,
-0.11300927400588989,
-0.04141866788268089,
-0.004593114834278822,
0.04996186122298241,
0.020711734890937805,
0.04342393949627876,
0.03823663666844368,
-0.03523506596684456,
-0.04596267268061638,
-0.020100543275475502,
0.027469903230667114,
0.01692320592701435,
0.049474943429231644,
0.014741679653525352,
0.019301053136587143,
0.024667935445904732,
0.1350439488887787,
-0.04250768944621086,
-0.01890072599053383,
0.020442521199584007,
0.027811503037810326,
0.009147757664322853,
-0.032119471579790115,
0.009213613346219063,
0.015579595230519772,
0.05380828306078911,
-0.03945315629243851,
0.006399102509021759,
-0.033199187368154526,
0.0766121968626976,
-0.029678482562303543,
0.010986772365868092,
0.0004249650228302926,
0.05791155621409416,
0.0925474688410759,
-0.029341528192162514,
0.06334865838289261,
0.020400499925017357,
0.011253311298787594,
-0.0304668340831995,
0.011316214688122272,
0.049472082406282425,
0.07002158463001251,
-0.04338732361793518,
0.03774288669228554,
0.05796096473932266,
0.06696370244026184,
0.03736215829849243,
0.02342710644006729,
-0.021683042868971825,
-0.04297535866498947,
-0.012867256067693233
] |
ZeroDiff: Solidified Visual-semantic Correlation in Zero-Shot Learning | https://openreview.net/forum?id=wy9FRV8O5s | [
"Zihan Ye",
"Shreyank N Gowda",
"Shiming Chen",
"Xiaowei Huang",
"Haotian Xu",
"Fahad Shahbaz Khan",
"Yaochu Jin",
"Kaizhu Huang",
"Xiaobo Jin"
] | Poster | Zero-shot Learning (ZSL) aims to enable classifiers to identify unseen classes. This is typically achieved by generating visual features for unseen classes based on learned visual-semantic correlations from seen classes. However, most current generative approaches heavily rely on having a sufficient number of samples from seen classes. Our study reveals that a scarcity of seen class samples results in a marked decrease in performance across many generative ZSL techniques. We argue, quantify, and empirically demonstrate that this decline is largely attributable to spurious visual-semantic correlations. To address this issue, we introduce ZeroDiff, an innovative generative framework for ZSL that incorporates diffusion mechanisms and contrastive representations to enhance visual-semantic correlations. ZeroDiff comprises three key components: (1) Diffusion augmentation, which naturally transforms limited data into an expanded set of noised data to mitigate generative model overfitting; (2) Supervised-contrastive (SC)-based representations that dynamically characterize each limited sample to support visual feature generation; and (3) Multiple feature discriminators employing a Wasserstein-distance-based mutual learning approach, evaluating generated features from various perspectives, including pre-defined semantics, SC-based representations, and the diffusion process. Extensive experiments on three popular ZSL benchmarks demonstrate that ZeroDiff not only achieves significant improvements over existing ZSL methods but also maintains robust performance even with scarce training data. Our codes are available at https://github.com/FouriYe/ZeroDiff_ICLR25. | Zero-shot Learning, Generative Model, Diffusion Mechanism, Effective Learning | We find, quantify and empirically prove a spurious visual-semantic correlation problem amplified by fewer training samples, and we propose a novel data-effective framework ZeroDiff to keep a robust performance under even 10% training set. | 1,261 | 2406.02929 | [
0.052303336560726166,
-0.06373036652803421,
0.045661561191082,
0.08834628015756607,
0.10167261213064194,
-0.014988820068538189,
-0.07473471015691757,
-0.06303152441978455,
0.02078208327293396,
-0.03454204276204109,
0.02199467457830906,
-0.09449215233325958,
0.057528719305992126,
0.04508255049586296,
-0.06408783048391342,
0.012880785390734673,
0.10329418629407883,
0.05541902780532837,
-0.08075810223817825,
-0.02492407336831093,
0.010153906419873238,
0.006603731773793697,
0.0880214050412178,
0.006590391043573618,
0.026169700548052788,
0.016343815252184868,
0.0009201884386129677,
-0.018305933102965355,
0.04019150510430336,
-0.03247544914484024,
0.04970517382025719,
-0.0020001977682113647,
-0.13314849138259888,
0.018232328817248344,
0.014944634400308132,
0.03346752002835274,
0.10563745349645615,
0.0999370813369751,
-0.03464561700820923,
0.029284875839948654,
0.013921806588768959,
0.062206752598285675,
0.00010240107076242566,
-0.007666910067200661,
0.06048488989472389,
0.07110657542943954,
-0.04694972559809685,
-0.057313352823257446,
0.0005285668885335326,
-0.01404654048383236,
-0.07191400974988937,
-0.01340563129633665,
-0.07118254899978638,
0.029329758137464523,
-0.009859706275165081,
0.056687988340854645,
0.030758272856473923,
0.012151040136814117,
-0.0061372737400233746,
0.006222025491297245,
-0.05795130133628845,
-0.10451050102710724,
0.035434454679489136,
-0.007775393780320883,
-0.027861112728714943,
-0.04039919376373291,
0.003246569074690342,
0.0035296939313411713,
-0.004129049368202686,
0.007501513697206974,
0.015206492505967617,
0.08673067390918732,
-0.038607966154813766,
0.07247795909643173,
-0.007670937106013298,
0.05770937725901604,
-0.030089421197772026,
-0.07117528468370438,
0.052305348217487335,
-0.07344399392604828,
0.050348106771707535,
0.0047979941591620445,
0.018181877210736275,
-0.022791389375925064,
0.08744094520807266,
0.04981007054448128,
0.03964851424098015,
0.04585694149136543,
-0.06157549098134041,
0.007301880978047848,
-0.07995762676000595,
0.0319148451089859,
-0.026524432003498077,
-0.011912400834262371,
-0.007661162875592709,
0.04434667155146599,
-0.01940850168466568,
-0.0266357883810997,
-0.02719888463616371,
0.03563773259520531,
0.02973923087120056,
0.010913500562310219,
-0.003093552775681019,
-0.02439950779080391,
0.015804097056388855,
0.012225920334458351,
0.051920291036367416,
0.01786682941019535,
-0.02971252240240574,
-0.04454059898853302,
0.05596774443984032,
-0.0030407754238694906,
-0.0326831191778183,
-0.07110020518302917,
-0.042973220348358154,
0.00832288432866335,
-0.04167487844824791,
-0.023559818044304848,
-0.007446322124451399,
0.014514822512865067,
0.04699968174099922,
-0.02931487001478672,
-0.08196702599525452,
0.04135069623589516,
0.018446318805217743,
-0.04030705988407135,
-0.020624686032533646,
1.3104088680375347e-33,
0.07335425168275833,
-0.02120056562125683,
0.04710138589143753,
0.061518553644418716,
0.04897443577647209,
-0.03715670108795166,
-0.021168509498238564,
0.0285856481641531,
-0.0359698086977005,
0.0006959006423130631,
0.01606220193207264,
0.0053551532328128815,
-0.07717422395944595,
0.12399224191904068,
0.022310877218842506,
-0.0450596958398819,
-0.04600755125284195,
-0.00743502750992775,
-0.02024039439857006,
0.015488076955080032,
-0.0048451730981469154,
0.02855989895761013,
-0.05963323265314102,
0.023104878142476082,
0.01829199120402336,
0.008590011857450008,
0.019248222932219505,
-0.10424922406673431,
-0.024864526465535164,
-0.0025793465320020914,
-0.1160980761051178,
0.03812135010957718,
-0.0034793971572071314,
-0.036203984171152115,
0.008635208010673523,
-0.08116050064563751,
-0.020238494500517845,
-0.03354667127132416,
0.02712245285511017,
-0.013082013465464115,
-0.06891284137964249,
0.03343750908970833,
-0.07551886141300201,
-0.06317666918039322,
-0.06017233431339264,
0.008138194680213928,
0.06488120555877686,
0.06853098422288895,
-0.0579184852540493,
-0.03985486179590225,
0.05182911455631256,
0.018718529492616653,
-0.003121746703982353,
-0.0011650642845779657,
-0.10484892129898071,
0.0023040466476231813,
0.01787891983985901,
-0.06923177093267441,
-0.03107145056128502,
-0.09425609558820724,
-0.006880262866616249,
0.07112652063369751,
-0.03085656650364399,
0.07104694843292236,
-0.0017213138053193688,
0.06522878259420395,
0.03499026224017143,
0.009934056550264359,
0.06447603553533554,
-0.016267038881778717,
-0.03803621977567673,
0.06076175719499588,
-0.06362641602754593,
-0.1203208863735199,
0.12108323723077774,
0.014606126584112644,
0.050991155207157135,
-0.006025106180459261,
0.06462311744689941,
0.0764460489153862,
-0.009341668337583542,
-0.05195842310786247,
-0.005481149069964886,
-0.05680520087480545,
-0.050670407712459564,
-0.0229503121227026,
0.03239254653453827,
-0.035044457763433456,
-0.021408816799521446,
-0.0652768537402153,
-0.0015934051480144262,
-0.06749780476093292,
-0.03843402862548828,
0.005482196342200041,
-0.02191382460296154,
-1.676276433778683e-33,
-0.04012800008058548,
0.09187006950378418,
-0.0085062887519598,
0.005426942370831966,
-0.042983390390872955,
0.04330108314752579,
0.046983104199171066,
0.025029899552464485,
0.09100092202425003,
-0.06262872368097305,
0.04959361255168915,
0.0003624598030000925,
-0.020116891711950302,
-0.002510473597794771,
0.025182116776704788,
0.029873237013816833,
-0.006648438982665539,
0.05155603960156441,
-0.058728933334350586,
0.04754504933953285,
0.0496283620595932,
0.10956108570098877,
-0.036660026758909225,
0.022971022874116898,
0.021417394280433655,
0.08534005284309387,
-0.04097087308764458,
0.02804998680949211,
0.0018606141675263643,
-0.007053058128803968,
0.08533936738967896,
0.016018029302358627,
0.019954068586230278,
-0.08252108097076416,
-0.11556577682495117,
0.04299947991967201,
0.04118852689862251,
-0.017287952825427055,
-0.02158544212579727,
0.09315654635429382,
0.06834906339645386,
-0.08589968830347061,
-0.14091283082962036,
0.0425659604370594,
-0.04185936227440834,
-0.10491594672203064,
-0.06513700634241104,
-0.0016023798380047083,
0.039894573390483856,
0.0771634578704834,
-0.09707944840192795,
0.07079582661390305,
-0.10202369838953018,
0.011482012458145618,
-0.0011992615181952715,
-0.10987880080938339,
0.013295454904437065,
-0.02043340913951397,
0.04464858025312424,
0.04034443944692612,
-0.012427268549799919,
0.04008699208498001,
-0.026553817093372345,
-0.044704630970954895,
-0.02020394802093506,
0.005046657752245665,
-0.03587207943201065,
-0.04303746297955513,
0.022606393322348595,
-0.005546197760850191,
0.08267099410295486,
0.09458793699741364,
-0.03354044258594513,
-0.08597608655691147,
0.011732256039977074,
0.044729605317115784,
0.01455980446189642,
0.03615952283143997,
-0.017703188583254814,
0.01075721625238657,
0.016755789518356323,
-0.04879290238022804,
0.020395372062921524,
0.05126927047967911,
0.11786852777004242,
0.05419526994228363,
0.03781428188085556,
-0.00987159926444292,
0.07921557873487473,
0.010165221989154816,
-0.0480530671775341,
-0.032452985644340515,
0.04542884975671768,
0.03696150332689285,
-0.10734064131975174,
-3.781426372029273e-8,
-0.07581868022680283,
-0.01341073028743267,
0.013639223761856556,
0.03023916855454445,
0.003591950051486492,
-0.05671126767992973,
-0.016571734100580215,
0.11349956691265106,
-0.0322086475789547,
0.021291563287377357,
0.12401216477155685,
0.027675341814756393,
-0.05253218114376068,
-0.07013868540525436,
0.034226588904857635,
0.09241464734077454,
-0.00869358703494072,
-0.022651394829154015,
-0.005030816420912743,
0.045129742473363876,
0.014092940837144852,
-0.038480792194604874,
-0.03778735548257828,
0.003695774357765913,
0.013026654720306396,
-0.10637795925140381,
0.012465814128518105,
-0.027911994606256485,
0.12986230850219727,
-0.07411013543605804,
-0.025445355102419853,
0.029706286266446114,
0.0571691058576107,
-0.05998468026518822,
0.02119535207748413,
0.09163115918636322,
0.04779454693198204,
0.04609207063913345,
-0.07658800482749939,
-0.01332854200154543,
-0.1061076894402504,
0.011033155024051666,
0.020759744569659233,
0.02077222429215908,
-0.02779580093920231,
-0.013781744986772537,
0.02226291410624981,
-0.07413546741008759,
0.047132913023233414,
0.02200818993151188,
-0.09121253341436386,
0.039350297302007675,
-0.011762011796236038,
0.0656675472855568,
0.015508590266108513,
-0.024380821734666824,
0.028798529878258705,
0.013218357227742672,
0.024471784010529518,
0.05597693845629692,
0.010115442797541618,
0.010310866869986057,
-0.012996634468436241,
0.04521877318620682
] |
Deep Networks Learn Features From Local Discontinuities in the Label Function | https://openreview.net/forum?id=52UtL8uA35 | [
"Prithaj Banerjee",
"Harish Guruprasad Ramaswamy",
"Mahesh Lorik Yadav",
"Chandra Shekar Lakshminarayanan"
] | Poster | Deep neural networks outperform kernel machines on several datasets due to feature learning that happens during gradient descent training. In this paper, we analyze the mechanism through which feature learning happens and use a notion of features that corresponds to discontinuities in the true label function. We hypothesize that the core feature learning mechanism is label function discontinuities attracting model function discontinuities during training. To test this hypothesis, we perform experiments on classification data where the true label function is given by an oblique decision tree. This setup allows easy enumeration of label function discontinuities, while still remaining intractable for static kernel/linear methods. We then design/construct a novel deep architecture called a Deep Linearly Gated Network (DLGN), whose discontinuities in the input space can be easily enumerated. In this setup, we provide supporting evidence demonstrating the movement of model function discontinuities towards the label function discontinuities during training. The easy enumerability of discontinuities in the DLGN also enables greater mechanistic interpretability. We demonstrate this by extracting the parameters of a high-accuracy decision tree from the parameters of a DLGN. We also show that the DLGN is competitive with ReLU networks and other tree-learning algorithms on several real-world tabular datasets. | Deep Learning, Feature learning, Interpretable, Local Discontinuities, Deep learning theory, Deep neural architectures, Supervised learning | Deep neural networks outperform kernel machines by learning features through discontinuities in label functions during gradient descent training, showing better performance and offering greater interpretability compared to ReLU networks. | 1,259 | null | [
-0.07878987491130829,
-0.07118754833936691,
0.00031097137252800167,
0.03343098983168602,
0.03592173382639885,
0.018565690144896507,
-0.014431391842663288,
-0.03781529515981674,
-0.0027080890722572803,
-0.07200759649276733,
-0.0034893846604973078,
-0.061842482537031174,
-0.03693384677171707,
0.022710608318448067,
-0.03372762352228165,
0.012223199009895325,
0.0671994537115097,
0.10593802481889725,
-0.10651732981204987,
-0.03609141707420349,
0.018443819135427475,
0.08946115523576736,
-0.02370087057352066,
0.08839969336986542,
-0.008372700773179531,
0.027949700132012367,
-0.00746459886431694,
-0.011898667551577091,
-0.017588414251804352,
-0.04422101378440857,
0.003849355038255453,
-0.009057272225618362,
-0.08539529889822006,
0.02105904184281826,
-0.04005269333720207,
0.030997322872281075,
-0.055693984031677246,
0.03794301301240921,
0.04319286346435547,
-0.039952658116817474,
0.029203591868281364,
-0.012954478152096272,
0.008544647134840488,
0.017159195616841316,
0.04573863372206688,
0.05864480137825012,
0.0402500182390213,
-0.06595601886510849,
-0.00012421762221492827,
0.07128933817148209,
0.006903624627739191,
0.019943227991461754,
0.01376681961119175,
0.0734458789229393,
-0.019725030288100243,
-0.00379214808344841,
0.037881795316934586,
0.023360297083854675,
-0.04671890288591385,
0.08982338011264801,
0.00834150705486536,
-0.0641767755150795,
-0.00045686151133850217,
-0.008360533975064754,
-0.0019818057771772146,
-0.0074374740943312645,
0.04488293454051018,
-0.013563680462539196,
-0.037301041185855865,
0.031604669988155365,
0.027001287788152695,
0.12790514528751373,
-0.0789988785982132,
-0.0177717674523592,
0.04546811431646347,
0.048129528760910034,
0.07838587462902069,
0.04284355044364929,
0.05005478486418724,
-0.044121384620666504,
0.03490133211016655,
0.0859735980629921,
0.03515291213989258,
0.010882862843573093,
0.10308437794446945,
-0.001339378417469561,
-0.06514687091112137,
0.05274685472249985,
-0.03290325403213501,
0.0031723654828965664,
-0.05067288130521774,
-0.0658145546913147,
-0.015537128783762455,
-0.06475689262151718,
0.025825290009379387,
-0.021565165370702744,
0.022541161626577377,
-0.08966335654258728,
-0.0742545798420906,
0.06814666092395782,
-0.14155331254005432,
0.02026279643177986,
0.014174415729939938,
-0.04630424827337265,
0.03660237044095993,
-0.024929611012339592,
0.04502249136567116,
0.004392677918076515,
0.11973803490400314,
-0.0759640708565712,
0.0011511504417285323,
0.014074108563363552,
0.0019900191109627485,
0.01660148613154888,
0.04367401823401451,
-0.00795559212565422,
0.017762241885066032,
-0.07488250732421875,
0.04690789431333542,
0.028133520856499672,
-0.1258450150489807,
0.00744848744943738,
-0.028950870037078857,
0.050484005361795425,
-0.02439969964325428,
0.0358036570250988,
-0.0504610575735569,
4.712766237377791e-33,
0.0009904808830469847,
-0.04446433484554291,
0.025972040370106697,
0.02111387625336647,
0.051074668765068054,
-0.08602336049079895,
-0.06771429628133774,
0.02044166438281536,
-0.004114578012377024,
0.08875617384910583,
-0.040690723806619644,
-0.0038155473303049803,
-0.0438876710832119,
0.06569688767194748,
0.06846475601196289,
0.009549720212817192,
0.030606163665652275,
-0.006383655127137899,
-0.015834374353289604,
-0.09664686769247055,
0.003307427978143096,
-0.03499491885304451,
-0.01041145995259285,
-0.06621534377336502,
-0.005284009966999292,
0.016919324174523354,
-0.012736768461763859,
0.03452735021710396,
-0.030045941472053528,
-0.019258780404925346,
-0.104202501475811,
-0.019551562145352364,
0.06940799206495285,
-0.010403529740869999,
0.04768475145101547,
-0.0025100756902247667,
-0.02903246320784092,
-0.014182417653501034,
0.05409187078475952,
-0.012797820381820202,
-0.019509032368659973,
0.04171289876103401,
-0.04120262339711189,
-0.03867605701088905,
-0.04016263037919998,
-0.012256652116775513,
0.06784471124410629,
-0.05951874330639839,
-0.05228918790817261,
-0.06162814050912857,
0.032989803701639175,
0.023728221654891968,
-0.005482665728777647,
-0.09497100859880447,
-0.0012345067225396633,
0.004705574829131365,
0.030853519216179848,
0.05249566212296486,
-0.032705020159482956,
0.08039512485265732,
0.004930824041366577,
0.03274335712194443,
-0.003414121689274907,
0.030736636370420456,
-0.06370145827531815,
0.05264372378587723,
-0.008951650001108646,
0.059128306806087494,
0.052204202860593796,
-0.04431327059864998,
-0.03797612711787224,
0.03583662956953049,
-0.022475456818938255,
-0.09991539269685745,
0.019885726273059845,
0.05233549699187279,
-0.006168085150420666,
-0.17703062295913696,
0.020530609413981438,
0.029470674693584442,
0.0658266469836235,
0.020427191630005836,
-0.000722891534678638,
0.03113107942044735,
-0.01226130872964859,
-0.03660479187965393,
0.0136690279468894,
-0.08494884520769119,
-0.007170307915657759,
-0.014128588140010834,
-0.11941830813884735,
-0.02371855638921261,
0.04415959119796753,
0.0016834461130201817,
0.05376610904932022,
-5.8422822286936185e-33,
-0.08866298198699951,
0.07222820818424225,
-0.045370057225227356,
0.004268281161785126,
-0.032675571739673615,
0.02533823810517788,
-0.005644494201987982,
0.05833856016397476,
0.004864577203989029,
-0.01226979959756136,
0.04214095324277878,
0.014784883707761765,
-0.01885438710451126,
-0.007531688082963228,
-0.004008118528872728,
-0.006654335185885429,
-0.09092319756746292,
0.014733138494193554,
0.019497092813253403,
0.029609039425849915,
0.051865216344594955,
0.10994173586368561,
-0.12172494828701019,
0.04816874861717224,
-0.0461123026907444,
0.04423784464597702,
-0.08612161874771118,
0.04925036057829857,
0.032092638313770294,
-0.0005277445307001472,
-0.08564925193786621,
-0.012754024937748909,
-0.033791281282901764,
-0.08141817152500153,
-0.02190478891134262,
-0.007181096356362104,
-0.01374661922454834,
-0.004489807412028313,
-0.05152023583650589,
0.043304890394210815,
0.04756186902523041,
0.014989623799920082,
-0.06551768630743027,
0.001490341150201857,
-0.013471143320202827,
-0.022243797779083252,
0.011132342740893364,
-0.030532564967870712,
0.012538689188659191,
0.06332338601350784,
-0.02416139654815197,
-0.0025945056695491076,
-0.07234311103820801,
0.099128857254982,
-0.12666140496730804,
-0.029530251398682594,
-0.024022981524467468,
0.034278105944395065,
0.12065795063972473,
0.03536034747958183,
-0.0934121161699295,
-0.009304868057370186,
-0.030544113367795944,
-0.0014035564381629229,
-0.011868524365127087,
-0.04550782963633537,
-0.0007539524231106043,
0.027277443557977676,
0.025118105113506317,
0.02395196445286274,
0.0810353010892868,
0.039284538477659225,
0.03902232274413109,
0.045324649661779404,
-0.0539323091506958,
-0.05791906639933586,
-0.018687838688492775,
-0.12911048531532288,
-0.0035545940045267344,
0.04580763727426529,
0.09808161109685898,
0.019809145480394363,
-0.009538890793919563,
0.08664063364267349,
0.08557527512311935,
-0.020565753802657127,
-0.016706356778740883,
0.011301112361252308,
0.0358370840549469,
-0.08168408274650574,
0.013848413713276386,
-0.0006010499782860279,
-0.0405782014131546,
0.06946876645088196,
-0.082261823117733,
-4.634931016767041e-8,
-0.0702379047870636,
0.030250543728470802,
-0.0333264134824276,
-0.035977743566036224,
0.05798324942588806,
-0.05555325001478195,
0.020617803558707237,
0.0977029949426651,
0.0057272110134363174,
-0.03808583691716194,
0.05518164113163948,
-0.012341739609837532,
-0.037206508219242096,
-0.06696443259716034,
-0.041852496564388275,
0.07087571918964386,
0.055074017494916916,
0.052094101905822754,
0.06487695872783661,
0.021390633657574654,
-0.003333779750391841,
-0.027397720143198967,
0.050698522478342056,
-0.023224659264087677,
0.011254426091909409,
-0.12915243208408356,
-0.029079221189022064,
0.11041150242090225,
0.067240871489048,
0.0785871297121048,
-0.008269269950687885,
0.04272777587175369,
0.0564434714615345,
0.02140498347580433,
-0.006500987336039543,
0.1551632285118103,
-0.02363828383386135,
-0.0028982062358409166,
-0.02539149299263954,
-0.024854375049471855,
-0.03015688620507717,
0.08302100002765656,
-0.036397360265254974,
0.01932472363114357,
-0.05051526799798012,
-0.005462095607072115,
0.061406832188367844,
0.0015345755964517593,
0.012391884811222553,
-0.03918207064270973,
-0.0034002417232841253,
0.043566226959228516,
0.038004424422979355,
0.11049355566501617,
0.02490023523569107,
-0.023095039650797844,
-0.06315451860427856,
-0.07464956492185593,
-0.08465061336755753,
0.0993015319108963,
0.011851765215396881,
0.0018783478299155831,
-0.039813391864299774,
-0.024489551782608032
] |
Data Pruning by Information Maximization | https://openreview.net/forum?id=93XT0lKOct | [
"Haoru Tan",
"Sitong Wu",
"Wei Huang",
"Shizhen Zhao",
"XIAOJUAN QI"
] | Poster | In this paper, we present InfoMax, a novel data pruning method, also known as coreset selection, designed to maximize the information content of selected samples while minimizing redundancy. By doing so, InfoMax enhances the overall informativeness of the coreset. The information of individual samples is measured by importance scores, which capture their influence or difficulty in model learning. To quantify redundancy, we use pairwise sample similarities, based on the premise that similar samples contribute similarly to the learning process.
We formalize the coreset selection problem as a discrete quadratic programming (DQP) task, with the objective of maximizing the total information content, represented as the sum of individual sample contributions minus the redundancies introduced by similar samples within the coreset.
To ensure practical scalability, we introduce an efficient gradient-based solver, complemented by sparsification techniques applied to the similarity matrix and dataset partitioning strategies.
This enables InfoMax to seamlessly scale to datasets with millions of samples.
Extensive experiments demonstrate the superior performance of InfoMax in various data pruning tasks, including image classification, vision-language pre-training, and instruction tuning for large language models. | Data Pruning, Deep Learning | null | 1,258 | null | [
-0.032339584082365036,
-0.029890766367316246,
-0.031651630997657776,
-0.018987976014614105,
0.06850995123386383,
-0.0046723634004592896,
0.014722849242389202,
0.025370832532644272,
-0.020998230203986168,
-0.027182534337043762,
0.015866801142692566,
-0.012795723043382168,
0.05533699318766594,
0.06223361939191818,
-0.027423210442066193,
0.02893959917128086,
0.07065573334693909,
0.07829774171113968,
-0.14916905760765076,
-0.08191215991973877,
-0.03593297302722931,
-0.049551837146282196,
-0.02241671271622181,
0.005320392549037933,
0.07566239684820175,
-0.015274913050234318,
0.03364504873752594,
0.011888588778674603,
0.011919909156858921,
-0.0046847425401210785,
0.04966225102543831,
-0.04554430767893791,
0.05943068116903305,
-0.01567843183875084,
-0.07027918845415115,
0.11390490084886551,
-0.024124661460518837,
0.03115862049162388,
-0.012463189661502838,
-0.03570203110575676,
-0.0007108560530468822,
0.03040951117873192,
-0.026872606948018074,
0.0346669964492321,
0.045905616134405136,
0.08914455771446228,
-0.026329299435019493,
-0.0587751679122448,
0.04140256717801094,
0.03158024698495865,
-0.04602893069386482,
0.025471126660704613,
-0.03764915093779564,
0.007594681810587645,
0.012095116078853607,
-0.025411823764443398,
0.00915201660245657,
-0.009315172210335732,
-0.06443466991186142,
-0.02426937222480774,
-0.03592600300908089,
-0.10809288173913956,
0.002026425674557686,
-0.02226969040930271,
-0.032504964619874954,
-0.05567195266485214,
0.07315155118703842,
0.005221141967922449,
-0.015175034292042255,
0.02803516574203968,
0.01460366602987051,
0.08022724092006683,
-0.04390304535627365,
0.07398740202188492,
0.013469118624925613,
0.015241861343383789,
0.06323503702878952,
-0.015795480459928513,
0.004851293750107288,
-0.05136976018548012,
-0.020273543894290924,
0.0751713365316391,
-0.016561008989810944,
0.038089364767074585,
0.12373676151037216,
-0.05278536304831505,
0.03062650002539158,
0.02535368502140045,
0.031797558069229126,
-0.048560261726379395,
-0.12942256033420563,
0.018485287204384804,
0.0327693447470665,
0.011884910054504871,
0.009903162717819214,
0.0056503149680793285,
-0.08020597696304321,
-0.047098562121391296,
-0.041297826915979385,
0.053069498389959335,
-0.04702479764819145,
0.03425339609384537,
0.016340050846338272,
-0.041028521955013275,
0.010460121557116508,
-0.07726261764764786,
-0.006968215107917786,
0.08438421785831451,
0.037652041763067245,
-0.1389721930027008,
0.03701742738485336,
0.0563415065407753,
-0.02478642761707306,
0.04558632895350456,
0.05015752837061882,
0.0011288197711110115,
0.052819110453128815,
0.01164718996733427,
-0.0005394581239670515,
0.09068511426448822,
0.033928874880075455,
0.027755338698625565,
-0.003718293970450759,
-0.003628365695476532,
0.0021129988599568605,
-0.07335133850574493,
-0.02812470868229866,
3.204663923280414e-33,
0.013157078996300697,
0.05304167419672012,
0.09805424511432648,
-0.015987642109394073,
0.06525939702987671,
-0.057829782366752625,
0.010402696207165718,
0.040228817611932755,
-0.02836616151034832,
0.0644214004278183,
-0.03179897740483284,
-0.020973775535821915,
-0.05583319813013077,
0.11452873796224594,
0.006759177427738905,
-0.01649552769958973,
-0.056858889758586884,
0.09441369771957397,
-0.04124714434146881,
-0.0881117656826973,
0.06359162926673889,
-0.01875936985015869,
-0.0024171213153749704,
-0.04963918775320053,
0.05135786533355713,
-0.029889341443777084,
0.06751338392496109,
-0.04194791987538338,
-0.04069368913769722,
-0.025754889473319054,
-0.07422623038291931,
-0.005704974755644798,
-0.0004449105472303927,
-0.017990093678236008,
0.07727212458848953,
0.013543793931603432,
0.04238646477460861,
-0.026673588901758194,
0.003846754552796483,
0.04224996641278267,
0.019091377034783363,
0.030309654772281647,
0.02420097403228283,
-0.020223291590809822,
-0.030596382915973663,
-0.002758887829259038,
-0.012385527603328228,
0.029100222513079643,
0.00898157712072134,
-0.06390269100666046,
0.06897186487913132,
-0.05401312932372093,
0.03547290340065956,
0.020222194492816925,
-0.04991968721151352,
0.017195550724864006,
-0.016270941123366356,
-0.01147126778960228,
0.04115833714604378,
0.02988417074084282,
-0.024143317714333534,
-0.00018959795124828815,
-0.0016661997651681304,
0.10010175406932831,
0.02712278999388218,
0.03426414728164673,
0.008904985152184963,
0.054661206901073456,
0.06407619267702103,
0.020064115524291992,
-0.04679909721016884,
0.04217735677957535,
-0.03309636935591698,
-0.06906628608703613,
0.061952751129865646,
-0.0015395970549434423,
0.14370381832122803,
-0.057263024151325226,
-0.005711376201361418,
-0.0012672827579081059,
-0.03031211905181408,
0.01089002750813961,
0.0413394570350647,
-0.15728610754013062,
-0.06831484287977219,
-0.009185390546917915,
0.10706835240125656,
-0.03790101036429405,
-0.027137653902173042,
-0.036486923694610596,
-0.05324655771255493,
0.045650091022253036,
-0.08988244086503983,
-0.05758840963244438,
0.011644194833934307,
-2.7001738903311434e-33,
-0.09028185158967972,
0.0005581897567026317,
-0.004173580091446638,
0.028018131852149963,
0.032291438430547714,
-0.04540328308939934,
0.009958450682461262,
-0.008310902863740921,
-0.07140228897333145,
-0.08388085663318634,
0.00201947009190917,
-0.09640736132860184,
0.04604852944612503,
-0.04817793518304825,
-0.06461843103170395,
0.12914764881134033,
-0.07053060829639435,
0.08014047145843506,
-0.02065739966928959,
0.04681330546736717,
-0.021563859656453133,
0.06398565322160721,
-0.05470339208841324,
0.0073831891641020775,
-0.023005610331892967,
0.02851526439189911,
-0.030813731253147125,
0.019734209403395653,
0.05130047723650932,
-0.05913117900490761,
-0.020102854818105698,
-0.006617140024900436,
-0.10782570391893387,
-0.09291835129261017,
-0.06621396541595459,
0.017957253381609917,
-0.015946106985211372,
0.007564632222056389,
-0.0389753095805645,
0.1567792445421219,
0.018129220232367516,
0.021066006273031235,
-0.06365098059177399,
0.05819901451468468,
-0.010322348214685917,
-0.024331601336598396,
-0.00043667369754984975,
-0.10182330012321472,
-0.004549876786768436,
0.0018101109890267253,
-0.024572711437940598,
0.03320061415433884,
-0.04245742782950401,
0.10260167717933655,
0.015861159190535545,
-0.06154175475239754,
0.002550594974309206,
0.014678790234029293,
0.05033232644200325,
-0.03341509774327278,
0.007621308788657188,
-0.020924877375364304,
-0.021974775940179825,
0.046459633857011795,
0.025558067485690117,
-0.07948573678731918,
0.06273635476827621,
0.0335346944630146,
-0.043220117688179016,
0.0453314483165741,
-0.0221934262663126,
0.013078903779387474,
0.035028018057346344,
-0.04441225528717041,
-0.03669672831892967,
0.019619300961494446,
0.01270558312535286,
0.00961125548928976,
-0.022313643246889114,
0.016153579577803612,
0.06957616657018661,
-0.03178897872567177,
0.022279096767306328,
0.018170831725001335,
0.049580298364162445,
0.12733332812786102,
0.12883591651916504,
0.08102815598249435,
0.08375784754753113,
-0.054400961846113205,
-0.06795962154865265,
-0.01192593201994896,
0.04445301741361618,
-0.0183755811303854,
-0.09020695835351944,
-4.574154033321065e-8,
-0.004720933269709349,
0.01569542847573757,
-0.09713675826787949,
-0.0153323570266366,
-0.003938684705644846,
-0.02691279724240303,
-0.08434028923511505,
0.11892352253198624,
-0.0722663402557373,
-0.040081936866045,
0.03435148671269417,
-0.0000462085627077613,
-0.02417108789086342,
-0.0433182530105114,
0.0027694390155375004,
0.03010762669146061,
0.04668961837887764,
-0.05356914550065994,
0.04915870353579521,
0.022516945376992226,
-0.04053652286529541,
0.007616776507347822,
-0.043031081557273865,
0.054544344544410706,
0.027026375755667686,
-0.023297052830457687,
0.01276517752557993,
0.09814267605543137,
0.15766198933124542,
0.036372214555740356,
-0.019719747826457024,
-0.01588868349790573,
0.033522993326187134,
-0.047603707760572433,
0.0402064211666584,
0.05324297770857811,
0.05987057089805603,
-0.004102677572518587,
-0.09692804515361786,
-0.07060354948043823,
0.006687565241008997,
0.0022546774707734585,
0.04939744248986244,
0.020612096413969994,
0.06935637444257736,
0.016488509252667427,
-0.005779437720775604,
-0.05514497309923172,
0.03261205181479454,
-0.05569717660546303,
0.023543862625956535,
-0.055914804339408875,
0.0069651310332119465,
0.05718954652547836,
0.03245611861348152,
-0.003863150253891945,
-0.060940247029066086,
-0.0012326696887612343,
0.0621400885283947,
-0.0340767540037632,
0.04344400390982628,
0.002710001775994897,
-0.08923596888780594,
0.0429028756916523
] |
Mufu: Multilingual Fused Learning for Low-Resource Translation with LLM | https://openreview.net/forum?id=0eMsrRMmCw | [
"Zheng Wei Lim",
"Nitish Gupta",
"Honglin Yu",
"Trevor Cohn"
] | Poster | Multilingual large language models (LLMs) are great translators, but this is largely limited to high-resource languages. For many LLMs, translating in and out of low-resource languages remains a challenging task. To maximize data efficiency in this low-resource setting, we introduce Mufu, which includes a selection of automatically generated multilingual candidates and an instruction to correct inaccurate translations in the prompt. Mufu prompts turn a translation task into a postediting one, and seek to harness the LLM’s reasoning capability with auxiliary translation candidates, from which the model is required to assess the input quality, align the semantics cross-lingually, copy from relevant inputs and override instances that are incorrect. Our experiments on En-XX translations over the Flores-200 dataset show LLMs finetuned against Mufu-style prompts are robust to poor quality auxiliary translation candidates, achieving performance superior to NLLB 1.3B distilled model in 64% of low- and very-low-resource language pairs. We then distill these models to reduce inference cost, while maintaining on average 3.1 chrF improvement over finetune-only baseline in low-resource translations. | translation, low-resource, large language model | We finetune LLMs for translation by prompting with multilingual context in order to harness their cross-lingual reasoning, and substantially improve the models' performance in low-resource translation. | 1,255 | 2409.13949 | [
0.0012401932617649436,
-0.08868101984262466,
0.054975010454654694,
0.007466504350304604,
0.04578839987516403,
-0.06357938051223755,
-0.022288497537374496,
0.02682432346045971,
0.042535945773124695,
-0.03344181180000305,
-0.022448206320405006,
-0.08908264338970184,
0.04743199050426483,
0.005051509477198124,
0.04871204122900963,
-0.008573714643716812,
0.10398638248443604,
0.08152418583631516,
-0.0860910639166832,
-0.08310634642839432,
0.06406556069850922,
-0.0010861698538064957,
0.08003822714090347,
0.045455168932676315,
0.06086932495236397,
-0.046159420162439346,
-0.05455255135893822,
-0.03125064820051193,
0.04641285166144371,
-0.04216203838586807,
-0.02156015671789646,
0.05408204346895218,
-0.0628865510225296,
0.04974529892206192,
-0.013926224783062935,
0.08990725129842758,
-0.05429740250110626,
-0.02541414275765419,
0.029818644747138023,
0.0018961020978167653,
-0.06856643408536911,
0.007376181893050671,
0.007470403332263231,
-0.056322161108255386,
0.049821637570858,
0.007149255834519863,
0.018199162557721138,
0.002512425882741809,
-0.018240395933389664,
0.04120552912354469,
-0.10074581205844879,
-0.014234142377972603,
0.0001665038289502263,
0.025283727794885635,
0.021558493375778198,
-0.05809692665934563,
0.014384230598807335,
0.020107464864850044,
0.0012463609455153346,
-0.032345645129680634,
-0.12693239748477936,
-0.0638878121972084,
-0.021099239587783813,
0.040853291749954224,
-0.025317123159766197,
-0.04812962934374809,
-0.011167650111019611,
0.10338088870048523,
-0.04497981071472168,
0.019672144204378128,
-0.03040355257689953,
0.03454830124974251,
-0.03740842267870903,
0.0476360023021698,
-0.03211214020848274,
0.05465806648135185,
0.017109669744968414,
-0.0012198872864246368,
0.030837256461381912,
-0.0382244773209095,
0.03555678948760033,
0.00939357653260231,
0.04010777175426483,
-0.044924408197402954,
0.1322440803050995,
-0.015539242886006832,
0.009103596210479736,
0.05240173637866974,
0.020105581730604172,
0.028153836727142334,
0.018091123551130295,
-0.004177690949290991,
0.061717111617326736,
0.1075606644153595,
0.0631895661354065,
0.09615433216094971,
0.0023564749862998724,
-0.04329182952642441,
-0.033596478402614594,
0.09208577126264572,
0.024404795840382576,
0.13417308032512665,
0.03617696091532707,
-0.05936465784907341,
-0.05251830071210861,
-0.0587381012737751,
0.04605456814169884,
0.044755227863788605,
0.003106110030785203,
-0.15003426373004913,
0.0691404864192009,
0.07579808682203293,
-0.0013760150177404284,
0.046145759522914886,
0.010368838906288147,
-0.037101902067661285,
0.04532884433865547,
-0.05171884596347809,
0.037415437400341034,
0.024154113605618477,
-0.0829647034406662,
0.014469546265900135,
0.0314943864941597,
0.009131868369877338,
-0.029737193137407303,
-0.08762679994106293,
-0.06408628821372986,
1.6631378966853438e-33,
0.09493386745452881,
0.11918582022190094,
-0.00491625489667058,
0.012863028794527054,
-0.010488308034837246,
-0.014313864521682262,
0.0008232055697590113,
0.06544534862041473,
-0.08155586570501328,
-0.0319591760635376,
0.0011699971510097384,
0.0035683601163327694,
-0.059175871312618256,
0.04886121675372124,
-0.04169297590851784,
-0.05290824919939041,
0.03362850099802017,
0.05466702952980995,
-0.009768345393240452,
0.0012618275359272957,
0.048124320805072784,
0.014670269563794136,
0.0028385280165821314,
-0.029727252200245857,
0.03492747247219086,
0.06695935875177383,
0.06912066787481308,
-0.06749298423528671,
-0.010606708936393261,
0.031723640859127045,
-0.14176596701145172,
-0.12666743993759155,
0.036501411348581314,
0.06319613009691238,
-0.010274859145283699,
-0.10402630269527435,
-0.06610938906669617,
-0.02531023696064949,
-0.019997313618659973,
-0.02638329565525055,
0.027848074212670326,
0.019928930327296257,
0.012190784327685833,
0.023452147841453552,
-0.040791723877191544,
-0.02949225716292858,
-0.03019832633435726,
0.002640282269567251,
-0.028914202004671097,
-0.01485193707048893,
0.005715271458029747,
0.0015813320642337203,
-0.059393126517534256,
-0.03676933795213699,
0.007594863418489695,
0.03649989515542984,
0.03304775431752205,
0.014964299276471138,
0.025378307327628136,
0.022661477327346802,
-0.058605361729860306,
0.020615115761756897,
-0.038841091096401215,
0.08735279738903046,
0.06180818751454353,
0.020771076902747154,
-0.08639770746231079,
0.017961544916033745,
0.07084614038467407,
-0.05106210708618164,
-0.06492911279201508,
-0.05292411893606186,
-0.018029380589723587,
-0.01435678731650114,
0.06228494271636009,
-0.020666636526584625,
0.003583438228815794,
-0.10054221004247665,
0.08202754706144333,
-0.001620460650883615,
0.04954567551612854,
-0.0029902781825512648,
-0.011389014311134815,
-0.047086451202631,
-0.02205038256943226,
-0.030796604230999947,
0.017049534246325493,
-0.024684660136699677,
0.04066877067089081,
0.02772488445043564,
0.02251497469842434,
-0.04692627117037773,
-0.039129432290792465,
-0.04627526178956032,
0.008593639358878136,
-1.2639176991205286e-34,
0.005364482756704092,
0.01093958504498005,
-0.008761608973145485,
0.13366861641407013,
-0.09571737796068192,
-0.0447564572095871,
0.045721735805273056,
0.03136735036969185,
0.004269272554665804,
-0.07580015063285828,
-0.03418167680501938,
-0.045631859451532364,
0.024761337786912918,
-0.02003326639533043,
-0.010399474762380123,
-0.0050347731448709965,
0.041598111391067505,
0.04755581170320511,
0.02489096112549305,
0.09643352031707764,
-0.025152407586574554,
0.0739884003996849,
-0.07715877890586853,
0.05239878222346306,
-0.028827479109168053,
0.07793082296848297,
-0.08090656250715256,
0.02931925095617771,
-0.0009509254596196115,
0.04709118977189064,
-0.016856586560606956,
-0.047872722148895264,
-0.03231623396277428,
0.0025272781495004892,
-0.09841836243867874,
0.011813294142484665,
0.025949129834771156,
0.05400727689266205,
-0.04808904230594635,
0.07535215467214584,
0.032429926097393036,
-0.007460723165422678,
-0.05065705254673958,
0.006087056826800108,
-0.05797303467988968,
0.03608902916312218,
-0.12871359288692474,
-0.04786347225308418,
0.02577582560479641,
0.006255790591239929,
-0.03306848555803299,
-0.04725474491715431,
-0.024804305285215378,
-0.020945198833942413,
0.00987214408814907,
-0.14172165095806122,
-0.01983068883419037,
-0.05807812139391899,
-0.04321330040693283,
-0.09663788229227066,
-0.027675986289978027,
-0.008803851902484894,
0.0605381615459919,
-0.06700433045625687,
0.05203747749328613,
0.005404848139733076,
-0.010788360610604286,
-0.007005656138062477,
0.050675127655267715,
-0.017263151705265045,
0.07004843652248383,
-0.043212536722421646,
-0.014455365017056465,
-0.04296968877315521,
0.0343824103474617,
0.02643037959933281,
-0.012805456295609474,
-0.08346833288669586,
0.0034708050079643726,
-0.020952468737959862,
-0.01699734851717949,
-0.005305551458150148,
-0.010617411695420742,
0.024288222193717957,
0.00021513273532036692,
0.027532361447811127,
0.02675008215010166,
0.11665550619363785,
0.03622546046972275,
0.05788546800613403,
-0.021334433928132057,
0.09712914377450943,
0.005931456107646227,
0.05489996448159218,
-0.04129542037844658,
-4.0241051380007775e-8,
-0.07583384960889816,
0.05440065264701843,
0.0031998902559280396,
0.09710463136434555,
0.038382332772016525,
-0.06989001482725143,
-0.04771394282579422,
0.08683586120605469,
0.03315139561891556,
0.0012287631398066878,
-0.02151026949286461,
0.01965230517089367,
-0.02074410207569599,
-0.08713117241859436,
0.02361946366727352,
0.09421417862176895,
0.04673376306891441,
0.03134457767009735,
0.010009469464421272,
0.019232794642448425,
0.06506979465484619,
0.07427863031625748,
0.008651437237858772,
0.025485070422291756,
0.08502937853336334,
-0.044625699520111084,
-0.07833386212587357,
0.09405724704265594,
0.08040884882211685,
-0.06049384921789169,
-0.03225826472043991,
-0.0367760956287384,
-0.002822639886289835,
-0.04386154189705849,
-0.027162138372659683,
0.04799254238605499,
-0.014926458708941936,
0.02064354158937931,
0.014783412218093872,
-0.0007180472603067756,
0.10387781262397766,
0.042375288903713226,
-0.08601497113704681,
0.03961979225277901,
-0.0013144079130142927,
-0.04863309487700462,
-0.050841785967350006,
-0.09036994725465775,
-0.009526089765131474,
-0.0630730390548706,
0.06881806999444962,
-0.034506384283304214,
-0.014304890297353268,
0.07992963492870331,
-0.009525874629616737,
0.05441443994641304,
-0.044422514736652374,
-0.0178949236869812,
0.03670833259820938,
0.008334948681294918,
0.022504976019263268,
0.009756217710673809,
-0.02644151821732521,
-0.0024631405249238014
] |
Rethinking LLM Unlearning Objectives: A Gradient Perspective and Go Beyond | https://openreview.net/forum?id=huo8MqVH6t | [
"Qizhou Wang",
"Jin Peng Zhou",
"Zhanke Zhou",
"Saebyeol Shin",
"Bo Han",
"Kilian Q Weinberger"
] | Poster | Large language models (LLMs) should undergo rigorous audits to identify potential risks, such as copyright and privacy infringements. Once these risks emerge, timely updates are crucial to remove undesirable responses, ensuring legal and safe model usage. It has spurred recent research into LLM unlearning, focusing on erasing targeted undesirable knowledge without compromising the integrity of other, non-targeted responses. Existing studies have introduced various unlearning objectives to pursue LLM unlearning without necessitating complete retraining. However, each of these objectives has unique properties, and no unified framework is currently available to comprehend them thoroughly. To fill the gap, we propose the metric of the G-effect, quantifying the impacts of unlearning objectives on model performance from a gradient lens. A significant advantage of our metric is its broad ability to detail the unlearning impacts from various aspects across instances, updating steps, and LLM layers. Accordingly, the G-effect offers new insights into identifying drawbacks of existing unlearning objectives, further motivating us to explore a series of candidate solutions for their mitigation and improvements. Finally, we outline promising directions that merit further studies, aiming at contributing to the community to advance this critical field. | LLM Unlearning | null | 1,253 | 2502.19301 | [
-0.08992422372102737,
-0.04076520353555679,
-0.007268964778631926,
0.0162399522960186,
0.1138574555516243,
0.051747873425483704,
-0.03233548253774643,
-0.06624341011047363,
0.08440671116113663,
-0.012529443949460983,
-0.03159010037779808,
-0.018304497003555298,
0.04592449963092804,
-0.018202833831310272,
-0.019116198644042015,
-0.016991104930639267,
0.09221620112657547,
0.07781688123941422,
-0.12441461533308029,
-0.06815209239721298,
0.023681985214352608,
0.03779225051403046,
0.049079980701208115,
0.06549894064664841,
0.006892908830195665,
-0.05086193233728409,
-0.001999689033254981,
-0.02244681492447853,
-0.003497528377920389,
-0.058520492166280746,
0.021125689148902893,
0.05367421358823776,
0.012119977734982967,
0.059710800647735596,
-0.05254250764846802,
0.04399299994111061,
-0.03871709480881691,
-0.004761397372931242,
-0.007796470075845718,
-0.07617272436618805,
-0.04721732437610626,
-0.05191049352288246,
-0.05932700261473656,
0.024811111390590668,
0.06475929170846939,
-0.07164370268583298,
0.048975616693496704,
0.024027729406952858,
-0.11461696028709412,
0.02556547522544861,
-0.03999219834804535,
-0.13046251237392426,
-0.07115714251995087,
-0.013574120588600636,
-0.009159543551504612,
-0.06057780981063843,
-0.017269819974899292,
0.0231736171990633,
-0.04457725211977959,
-0.062371786683797836,
-0.0822378620505333,
-0.08707688003778458,
-0.07785094529390335,
0.05644594132900238,
-0.005828664638102055,
0.00844742264598608,
0.009503309614956379,
0.06881425529718399,
-0.03232662379741669,
0.06373141705989838,
0.028882121667265892,
0.015814898535609245,
-0.019177306443452835,
-0.013679773546755314,
-0.025484509766101837,
0.07406875491142273,
0.010521414689719677,
0.04329824820160866,
0.047742679715156555,
-0.11125703901052475,
0.03923409804701805,
0.011167749762535095,
0.06829309463500977,
0.06600642204284668,
0.018170734867453575,
-0.04526896774768829,
0.0475018136203289,
-0.006657617166638374,
0.04882938787341118,
0.013237642124295235,
-0.019434379413723946,
-0.06628893315792084,
0.09657073765993118,
0.014352414757013321,
0.05440540984272957,
-0.02240951918065548,
-0.032800301909446716,
-0.09292563050985336,
0.021615397185087204,
0.06161194667220116,
-0.005937344394624233,
0.10820603370666504,
-0.0054340665228664875,
-0.07803961634635925,
-0.00712263397872448,
-0.0031268044840544462,
0.010335691273212433,
0.007909169420599937,
0.024868493899703026,
-0.015329579822719097,
0.0031637833453714848,
0.11340229213237762,
0.005218164529651403,
-0.009175980463624,
0.058789633214473724,
-0.022072050720453262,
0.01928485557436943,
-0.02713107131421566,
-0.011431053280830383,
0.07485821843147278,
-0.023687487468123436,
0.012872213497757912,
0.05909451097249985,
0.010184818878769875,
-0.07601732015609741,
-0.002647161018103361,
-0.08396302908658981,
5.115860515348031e-33,
0.04376762732863426,
0.03885861113667488,
-0.05029330775141716,
0.028601882979273796,
0.026473432779312134,
0.025088218972086906,
0.009119316004216671,
0.07370542734861374,
-0.048723962157964706,
-0.004473501350730658,
-0.010629264637827873,
0.09517809003591537,
-0.030290130525827408,
0.06391516327857971,
0.0646137073636055,
0.043126996606588364,
-0.062084589153528214,
0.10338786989450455,
-0.006382285617291927,
-0.031144872307777405,
0.08581619709730148,
-0.0558064840734005,
0.04136257991194725,
-0.07698193192481995,
-0.00969423446804285,
-0.011763382703065872,
0.04472620412707329,
-0.07008591294288635,
-0.04663646221160889,
0.049927275627851486,
-0.05822623148560524,
-0.02615232579410076,
0.06501401960849762,
0.016886495053768158,
0.022673487663269043,
0.03013588860630989,
-0.0709424614906311,
-0.07719823718070984,
0.08906199038028717,
-0.002530458150431514,
-0.040481165051460266,
-0.029097001999616623,
-0.024950526654720306,
-0.05673109367489815,
-0.039261382073163986,
0.011119741015136242,
-0.0008914358913898468,
-0.043547097593545914,
0.0065769231878221035,
-0.044697634875774384,
0.11804884672164917,
0.027012182399630547,
-0.04837779700756073,
-0.03546597436070442,
-0.05535304546356201,
0.038597241044044495,
-0.03215763717889786,
-0.041408490389585495,
-0.03370274603366852,
0.031993310898542404,
-0.0038352867122739553,
0.1047653928399086,
-0.014756309799849987,
0.09141034632921219,
-0.0034236141946166754,
0.028524717316031456,
-0.02899395115673542,
0.03206357732415199,
0.05548899248242378,
-0.020613765344023705,
-0.028009261935949326,
-0.04623360559344292,
-0.05186335742473602,
-0.03196239471435547,
-0.02806544117629528,
-0.07460764050483704,
0.022099897265434265,
-0.005461963824927807,
0.08027578145265579,
-0.0023742401972413063,
0.014623017981648445,
0.005281904712319374,
0.017286820337176323,
-0.06630613654851913,
-0.09207817912101746,
-0.046349529176950455,
0.04552914574742317,
-0.053311221301555634,
-0.059016451239585876,
0.049464695155620575,
-0.010229922831058502,
-0.044750116765499115,
-0.06605324149131775,
0.03913791850209236,
0.027806149795651436,
-7.157410982056193e-33,
-0.11292912811040878,
0.00020838830096181482,
-0.04115084558725357,
0.07439330220222473,
-0.036168280988931656,
-0.08167126029729843,
-0.06471601128578186,
0.04545993357896805,
-0.006691646762192249,
-0.027164757251739502,
-0.029049266129732132,
-0.0371595062315464,
-0.007655446417629719,
0.02619396522641182,
0.026344789192080498,
-0.038014963269233704,
-0.010408543050289154,
-0.011778639629483223,
-0.024838943034410477,
0.060159072279930115,
0.03709433600306511,
0.036535825580358505,
-0.09883120656013489,
0.034885384142398834,
0.00851356890052557,
-0.011269626207649708,
-0.07073365896940231,
-0.026095449924468994,
0.06782946735620499,
0.020551981404423714,
0.00636263145133853,
0.051510732620954514,
-0.10127346962690353,
-0.05371404439210892,
-0.03927198797464371,
0.009060085751116276,
0.07450352609157562,
-0.00003443587411311455,
-0.002704367507249117,
0.04252486303448677,
0.08339985460042953,
0.008294147439301014,
-0.09786462038755417,
-0.03985144942998886,
0.02464355155825615,
-0.01227857917547226,
-0.06766807287931442,
0.00881460215896368,
0.10595057904720306,
-0.03392445668578148,
0.008431745693087578,
-0.017331253737211227,
0.020402027294039726,
0.006180217955261469,
-0.048656173050403595,
-0.017327656969428062,
0.03546792268753052,
-0.0911884605884552,
0.027835022658109665,
0.06598355621099472,
-0.024220669642090797,
0.03509848192334175,
-0.06083960458636284,
-0.048566728830337524,
0.017142390832304955,
-0.05792625993490219,
0.023518797010183334,
0.061437755823135376,
-0.014517576433718204,
-0.004904936999082565,
0.031069036573171616,
-0.0317961648106575,
-0.05626470968127251,
-0.0711318626999855,
-0.019639303907752037,
-0.0024018711410462856,
0.016746189445257187,
-0.0837734118103981,
-0.06568160653114319,
-0.0614929124712944,
0.12048079818487167,
0.04294401407241821,
0.0314839705824852,
0.022974221035838127,
0.0500146746635437,
-0.02644348330795765,
0.03339828550815582,
0.08508133888244629,
0.0006840855348855257,
0.04266698658466339,
-0.08024772256612778,
-0.0418296754360199,
-0.09010732918977737,
0.046634282916784286,
-0.05399547517299652,
-5.112795165018724e-8,
-0.028248323127627373,
0.04377914220094681,
0.037349097430706024,
0.08398416638374329,
0.025233399122953415,
-0.044410284608602524,
-0.09235726296901703,
0.09593260288238525,
-0.026891643181443214,
-0.04013992100954056,
0.04998130723834038,
-0.02032197080552578,
-0.10894850641489029,
-0.002498987130820751,
-0.04475187510251999,
0.07032736390829086,
-0.020493589341640472,
0.039313677698373795,
-0.02477036602795124,
-0.005683491472154856,
0.022970933467149734,
0.043120961636304855,
-0.0023914878256618977,
-0.038133930414915085,
0.06498555839061737,
-0.07441107928752899,
0.017033647745847702,
0.07940466701984406,
0.08156401664018631,
0.07274942100048065,
0.012843014672398567,
-0.025456752628087997,
-0.003065968630835414,
0.04993981122970581,
-0.033698324114084244,
0.06215433031320572,
0.020754579454660416,
-0.016610343009233475,
0.06065516173839569,
0.008106077089905739,
0.021133694797754288,
0.10057925432920456,
0.04476907476782799,
0.08010122925043106,
-0.050988320261240005,
-0.06389372795820236,
-0.06124977022409439,
-0.08716309070587158,
0.005296438932418823,
-0.004018702544271946,
-0.026407113298773766,
-0.0440794974565506,
-0.005154906772077084,
0.1072530746459961,
0.0789758637547493,
0.02239184081554413,
0.007774564903229475,
-0.0014461078681051731,
0.05594811588525772,
0.020574692636728287,
0.1005670577287674,
-0.040079452097415924,
-0.021056804805994034,
0.006015604827553034
] |
HiSplat: Hierarchical 3D Gaussian Splatting for Generalizable Sparse-View Reconstruction | https://openreview.net/forum?id=SBzIbJojs8 | [
"Shengji Tang",
"Weicai Ye",
"Peng Ye",
"Weihao Lin",
"Yang Zhou",
"Tao Chen",
"Wanli Ouyang"
] | Poster | Reconstructing 3D scenes from multiple viewpoints is a fundamental task in stereo vision. Recently, advances in generalizable 3D Gaussian Splatting have enabled high-quality novel view synthesis for unseen scenes from sparse input views by feed-forward predicting per-pixel Gaussian parameters without extra optimization. However, existing methods typically generate single-scale 3D Gaussians, which lack representation of both large-scale structure and texture details, resulting in mislocation and artefacts. In this paper, we propose a novel framework, HiSplat, which introduces a hierarchical manner in generalizable 3D Gaussian Splatting to construct hierarchical 3D Gaussians via a coarse-to-fine strategy. Specifically, HiSplat generates large coarse-grained Gaussians to capture large-scale structures, followed by fine-grained Gaussians to enhance delicate texture details. To promote inter-scale interactions, we propose an Error Aware Module for Gaussian compensation and a Modulating Fusion Module for Gaussian repair. Our method achieves joint optimization of hierarchical representations, allowing for novel view synthesis using only two-view reference images. Comprehensive experiments on various datasets demonstrate that HiSplat significantly enhances reconstruction quality and cross-dataset generalization compared to prior single-scale methods. The corresponding ablation study and analysis of different-scale 3D Gaussians reveal the mechanism behind the effectiveness. Code is at https://github.com/Open3DVLab/HiSplat. | 3D reconstruction, Gaussian Splatting, Generalizable Multi-View Reconstruction | A hierachical 3D gaussian splatting model for generalizable multi-view resconstruction | 1,252 | 2410.06245 | [
-0.029241476207971573,
-0.05945611745119095,
0.04822631925344467,
-0.04372549057006836,
0.06793481856584549,
-0.014886237680912018,
0.0071008955128490925,
-0.06636004149913788,
-0.06414522230625153,
-0.10884816944599152,
-0.009766021743416786,
-0.06039942800998688,
0.03945581987500191,
0.05724917724728584,
-0.03805173560976982,
-0.011094337329268456,
0.09882736951112747,
0.1327292025089264,
-0.06147144362330437,
-0.023393141105771065,
-0.01394406333565712,
-0.017355207353830338,
-0.07736416906118393,
-0.034400518983602524,
0.08029069751501083,
0.034439828246831894,
0.061559632420539856,
-0.05148836225271225,
-0.035878367722034454,
-0.031082898378372192,
0.024323508143424988,
-0.015937069430947304,
-0.003068787045776844,
-0.03635086491703987,
-0.05294634774327278,
0.0629710927605629,
0.012090422213077545,
0.012812874279916286,
0.011680465191602707,
-0.05431041494011879,
0.0257794838398695,
0.07167007029056549,
-0.056665755808353424,
0.036857008934020996,
0.04265989363193512,
0.046984560787677765,
0.08712369948625565,
-0.10364475101232529,
0.059813618659973145,
-0.019235141575336456,
-0.030661877244710922,
0.016232682392001152,
0.01586218550801277,
-0.01286397222429514,
0.03902279958128929,
0.04408280923962593,
0.022933514788746834,
-0.08766075223684311,
0.06323769688606262,
0.030876057222485542,
0.04011901840567589,
-0.012472091242671013,
0.019005922600626945,
-0.011412887834012508,
0.006462987512350082,
-0.04877910763025284,
0.11339596658945084,
-0.09431149065494537,
-0.043925873935222626,
0.07979822903871536,
-0.08365806192159653,
-0.007998659275472164,
-0.0050423406064510345,
0.02815149910748005,
-0.00842189509421587,
0.0022364575415849686,
-0.07392633706331253,
-0.09767859429121017,
0.0033580095041543245,
0.021594233810901642,
0.002388406777754426,
0.051423173397779465,
-0.004730345215648413,
0.005109037272632122,
-0.012410962954163551,
0.03087327629327774,
-0.020930960774421692,
0.11875946819782257,
0.062359750270843506,
0.03001547046005726,
-0.03165470063686371,
-0.012319399043917656,
-0.06571907550096512,
-0.06420857459306717,
-0.025634804740548134,
-0.05382228642702103,
-0.029726240783929825,
-0.0860530436038971,
0.06987328082323074,
-0.0053006126545369625,
0.08008594065904617,
-0.06796683371067047,
0.03868350014090538,
0.0050878035835921764,
0.031972065567970276,
-0.004561615642160177,
0.01810753345489502,
0.08502406626939774,
0.02505175583064556,
0.028513461351394653,
-0.02960011176764965,
-0.018750837072730064,
0.019999278709292412,
-0.015723560005426407,
-0.0515068955719471,
0.07808969914913177,
0.022250540554523468,
0.02712942101061344,
-0.010627136565744877,
-0.00865404587239027,
0.05491703376173973,
-0.010971281677484512,
-0.023339807987213135,
0.01704448275268078,
0.005465905647724867,
-0.05229536443948746,
-0.07275603711605072,
1.4362764868552658e-33,
0.04974688962101936,
0.036782339215278625,
0.027259960770606995,
0.0702483132481575,
-0.002352474257349968,
0.07235119491815567,
-0.028856873512268066,
0.022887498140335083,
-0.0026192497462034225,
0.052617743611335754,
-0.042519863694906235,
-0.03049447014927864,
-0.024808315560221672,
0.08309017866849899,
-0.09711537510156631,
0.08658846467733383,
-0.02617371827363968,
-0.015926184132695198,
-0.03324052691459656,
0.06275860220193863,
-0.07564029097557068,
-0.017372827976942062,
-0.006372672040015459,
-0.00711846211925149,
-0.04205916076898575,
-0.039071936160326004,
0.05706246942281723,
-0.043112222105264664,
0.001766352099366486,
-0.027600225061178207,
-0.02695649489760399,
0.009597864001989365,
0.0004870269331149757,
0.04136744141578674,
-0.01458095945417881,
-0.05795532464981079,
0.04308639466762543,
-0.05797579512000084,
-0.040009815245866776,
-0.004788725171238184,
-0.053535252809524536,
0.09908508509397507,
0.005747791845351458,
-0.014478908851742744,
-0.05446162819862366,
0.03821102902293205,
0.1101926937699318,
0.11413523554801941,
-0.01347858551889658,
-0.0290475282818079,
0.0982121005654335,
-0.0060124630108475685,
-0.02126910537481308,
-0.0086653558537364,
-0.05163811519742012,
0.017765535041689873,
0.035353049635887146,
-0.005700690671801567,
0.05622899904847145,
-0.009419679641723633,
-0.025947628542780876,
0.03023410029709339,
-0.047479741275310516,
0.06627264618873596,
-0.0413450188934803,
0.017358139157295227,
0.05301303416490555,
0.08649948984384537,
-0.00851697288453579,
0.023894701153039932,
-0.0410429984331131,
0.01630820333957672,
-0.0056794174015522,
0.013107996433973312,
-0.03260422125458717,
-0.02357848547399044,
0.0088468873873353,
-0.04552961513400078,
0.047588884830474854,
0.050154171884059906,
-0.06576935201883316,
0.03809204697608948,
0.03514580428600311,
-0.037955619394779205,
-0.11252328008413315,
0.0034233881160616875,
0.006533347070217133,
-0.010431671515107155,
-0.042976319789886475,
-0.05883321166038513,
-0.08204938471317291,
-0.09009669721126556,
-0.04632746800780296,
-0.07314431667327881,
-0.025275783613324165,
-1.8251452627996407e-33,
0.06772033870220184,
-0.03386596217751503,
0.015358345583081245,
0.038484323769807816,
0.013579337857663631,
-0.05520346015691757,
0.027568332850933075,
0.005940122529864311,
0.10512391477823257,
-0.0259327944368124,
-0.03767751157283783,
0.019535904750227928,
0.03373830392956734,
0.043332844972610474,
-0.012821043841540813,
0.0034158702474087477,
-0.060601893812417984,
-0.009405052289366722,
-0.10214146971702576,
-0.03752783685922623,
0.004861743655055761,
0.026917900890111923,
0.023197103291749954,
-0.00007385634671663865,
-0.07074422389268875,
0.12886077165603638,
-0.09523802995681763,
-0.003877525217831135,
-0.007936370559036732,
0.0365489237010479,
-0.010655415244400501,
-0.012727648951113224,
0.013968074694275856,
-0.083755262196064,
-0.007859673351049423,
0.05093371123075485,
0.043624397367239,
0.04315563663840294,
-0.07376117259263992,
-0.03523780778050423,
-0.0008481637923978269,
0.015819162130355835,
-0.009095264598727226,
0.06595360487699509,
0.020443584769964218,
0.05010644719004631,
0.017812516540288925,
-0.05369235947728157,
-0.028565052896738052,
0.12664258480072021,
-0.10487473011016846,
0.11200031638145447,
-0.075045645236969,
0.013546615839004517,
0.03434222936630249,
-0.07752759009599686,
-0.11415508389472961,
-0.0033937799744307995,
0.12238844484090805,
-0.003287698607891798,
-0.0045259310863912106,
-0.035354480147361755,
-0.07791167497634888,
-0.011344199068844318,
-0.014433345757424831,
0.020059270784258842,
-0.007928598672151566,
-0.05035168677568436,
0.048831067979335785,
0.10108409076929092,
-0.034589990973472595,
-0.0014227393548935652,
-0.007742184214293957,
-0.01744762621819973,
-0.015580880455672741,
0.003935555461794138,
-0.012980428524315357,
-0.020144300535321236,
0.07320450991392136,
0.07552120089530945,
0.1067538633942604,
-0.037381403148174286,
0.08347593247890472,
0.12454023212194443,
0.1039956733584404,
0.04026517644524574,
-0.016850104555487633,
0.06253858655691147,
0.08774376660585403,
-0.0054955133236944675,
-0.0040329559706151485,
-0.011630530469119549,
0.015148749575018883,
-0.001358577748760581,
-0.02324390411376953,
-3.861537933858017e-8,
0.023135844618082047,
0.06114466115832329,
-0.019230835139751434,
-0.048733245581388474,
-0.03758732229471207,
-0.052069634199142456,
-0.013398106209933758,
0.03582832217216492,
0.015883808955550194,
-0.12256518006324768,
-0.012867945246398449,
-0.09509287029504776,
-0.027166221290826797,
-0.04843468964099884,
0.040259480476379395,
0.04243095591664314,
-0.02435237169265747,
0.01869909279048443,
-0.01923379860818386,
0.03264716640114784,
-0.06497956812381744,
0.023404566571116447,
0.03503326326608658,
0.04501764848828316,
-0.057448260486125946,
0.02993839420378208,
-0.03977775201201439,
0.09484875947237015,
0.1321740299463272,
-0.05449704825878143,
0.010526984930038452,
-0.04110920801758766,
0.044831667095422745,
-0.016602957621216774,
-0.037829626351594925,
0.011925062164664268,
-0.01647418923676014,
0.02087121270596981,
0.0788532942533493,
-0.12681853771209717,
-0.059320878237485886,
-0.022562330588698387,
0.019398901611566544,
0.012534625828266144,
-0.02166086807847023,
0.06336578726768494,
0.04161443933844566,
-0.10077177733182907,
-0.0011761535424739122,
0.01411912590265274,
-0.0036511938087642193,
0.042310722172260284,
0.0002659018791746348,
0.0757158175110817,
0.053679440170526505,
-0.02888028509914875,
0.0450451634824276,
-0.030969450250267982,
0.06854221969842911,
-0.044639211148023605,
-0.08475252985954285,
-0.003854847513139248,
-0.03481106087565422,
-0.00282701151445508
] |
OBI-Bench: Can LMMs Aid in Study of Ancient Script on Oracle Bones? | https://openreview.net/forum?id=hL5jone2Oh | [
"Zijian Chen",
"tingzhu chen",
"Wenjun Zhang",
"Guangtao Zhai"
] | Poster | We introduce OBI-Bench, a holistic benchmark crafted to systematically evaluate large multi-modal models (LMMs) on whole-process oracle bone inscriptions (OBI) processing tasks demanding expert-level domain knowledge and deliberate cognition. OBI-Bench includes 5,523 meticulously collected diverse-sourced images, covering five key domain problems: recognition, rejoining, classification, retrieval, and deciphering. These images span centuries of archaeological findings and years of research by front-line scholars, comprising multi-stage font appearances from excavation to synthesis, such as original oracle bone, inked rubbings, oracle bone fragments, cropped single characters, and handprinted characters. Unlike existing benchmarks, OBI-Bench focuses on advanced visual perception and reasoning with OBI-specific knowledge, challenging LMMs to perform tasks akin to those faced by experts. The evaluation of 6 proprietary LMMs as well as 17 open-source LMMs highlights the substantial challenges and demands posed by OBI-Bench. Even the latest versions of GPT-4o, Gemini 1.5 Pro, and Qwen-VL-Max are still far from public-level humans in some fine-grained perception tasks. However, they perform at a level comparable to untrained humans in deciphering tasks, indicating remarkable capabilities in offering new interpretative perspectives and generating creative guesses. We hope OBI-Bench can facilitate the community to develop domain-specific multi-modal foundation models towards ancient language research and delve deeper to discover and enhance these untapped potentials of LMMs. | oracle bone inscriptions, ancient character deciphering, large multi-modal models, benchmark | Using large multi-modal models to solve problems in the oracle bone inscription area | 1,237 | null | [
-0.09998046606779099,
0.027064111083745956,
0.03968517854809761,
-0.09515873342752457,
-0.015238150022923946,
-0.06689843535423279,
-0.06492938846349716,
0.026940232142806053,
-0.03536231070756912,
0.023327581584453583,
-0.05407612398266792,
-0.016010751947760582,
0.026741037145256996,
0.02403082884848118,
-0.07281467318534851,
0.01879533752799034,
-0.05274797976016998,
0.013200636953115463,
-0.022313343361020088,
0.005868856329470873,
0.00851499568670988,
0.0033381280954927206,
0.05195464566349983,
-0.06300182640552521,
0.1010233536362648,
0.08664865791797638,
-0.03993062302470207,
-0.05338020995259285,
0.06907666474580765,
-0.05517030879855156,
-0.018190214410424232,
0.06662113219499588,
0.006666592787951231,
-0.05208495259284973,
-0.01724669523537159,
0.09297643601894379,
-0.022951604798436165,
-0.022393159568309784,
0.014604233205318451,
-0.049949537962675095,
-0.005503651220351458,
-0.0012865415774285793,
-0.0530867725610733,
0.03900078311562538,
0.0665501207113266,
0.0004975101328454912,
-0.026081370189785957,
-0.028799468651413918,
-0.008031623438000679,
0.03363652527332306,
-0.050976380705833435,
-0.03721882775425911,
0.0011595480609685183,
-0.009272321127355099,
-0.044294439256191254,
0.04667162150144577,
0.03233307600021362,
-0.055894166231155396,
-0.0501534603536129,
-0.019342150539159775,
-0.043602339923381805,
0.07064411789178848,
-0.06106020137667656,
0.043665625154972076,
-0.008725034072995186,
0.06756336987018585,
0.07509766519069672,
-0.0977933406829834,
-0.036187417805194855,
-0.05644530430436134,
-0.05437231436371803,
0.015341767109930515,
0.04638722166419029,
-0.0038774001877754927,
-0.04345785826444626,
0.052430205047130585,
0.015166024677455425,
-0.07017260044813156,
-0.016635820269584656,
-0.12755583226680756,
0.02452738583087921,
0.04188177362084389,
0.006828151177614927,
0.011454158462584019,
0.027000542730093002,
0.026981772854924202,
-0.06192662566900253,
0.04699164628982544,
0.043683771044015884,
-0.04216396063566208,
0.04108963534235954,
-0.09877712279558182,
0.0033245233353227377,
-0.021710967645049095,
0.0863855630159378,
0.04619723930954933,
0.0756363645195961,
0.026377268135547638,
0.030942901968955994,
0.10590563714504242,
0.008591127581894398,
0.03776512295007706,
0.038316644728183746,
-0.02963084913790226,
-0.025005683302879333,
-0.019213879480957985,
0.012084529735147953,
-0.029143882915377617,
0.03681812435388565,
-0.09525753557682037,
0.05765621364116669,
-0.023372652009129524,
0.009908793494105339,
-0.008419015444815159,
-0.044814907014369965,
-0.02775174379348755,
-0.043340254575014114,
-0.030076105147600174,
0.038820210844278336,
-0.0530216284096241,
0.0032929410226643085,
-0.06605725735425949,
-0.013908134773373604,
0.02808881364762783,
0.05478735640645027,
-0.03265140950679779,
-0.05375925451517105,
2.5373772899628647e-33,
0.06570159643888474,
0.0715150237083435,
0.018057866021990776,
-0.025629444047808647,
0.012591315433382988,
0.019041286781430244,
-0.08806978911161423,
0.0020196004770696163,
-0.07374020665884018,
0.04876258596777916,
0.01371053047478199,
0.04854829981923103,
-0.041741155087947845,
0.06550075858831406,
0.057607416063547134,
0.08617959916591644,
-0.027527762576937675,
0.012711742892861366,
-0.08959027379751205,
0.017143744975328445,
0.019725123420357704,
-0.05332200601696968,
0.02525324374437332,
-0.030205536633729935,
-0.05655500665307045,
0.07384751737117767,
0.03323793783783913,
-0.06081100180745125,
-0.05199973285198212,
0.015087451785802841,
-0.050536446273326874,
-0.0787166953086853,
0.0054410649463534355,
-0.024392755702137947,
-0.007515226490795612,
-0.05995279923081398,
0.026798050850629807,
-0.1026076152920723,
-0.017884427681565285,
0.02927163802087307,
0.09023641049861908,
0.04819255322217941,
0.02412538416683674,
0.007051955442875624,
-0.040840715169906616,
0.06779942661523819,
0.011625217273831367,
0.05948926880955696,
-0.03686358034610748,
-0.009713455103337765,
0.011106581427156925,
0.07405447214841843,
-0.051456257700920105,
-0.03623804450035095,
-0.002999130869284272,
0.026993514969944954,
-0.027428314089775085,
-0.03912739083170891,
0.03570697829127312,
0.04794250801205635,
0.043831560760736465,
0.04453085735440254,
0.03548001870512962,
0.1236703097820282,
-0.03666325658559799,
0.06221669912338257,
-0.056111279875040054,
0.034876707941293716,
0.0456022284924984,
0.067354217171669,
-0.06427360326051712,
-0.02099360153079033,
0.03424404561519623,
-0.017447343096137047,
0.03958309814333916,
0.016435932368040085,
0.0856594517827034,
-0.040954213589429855,
-0.03831077367067337,
0.010248000733554363,
0.010776860639452934,
0.048306770622730255,
-0.08110461384057999,
-0.05998906120657921,
0.011306286789476871,
-0.035500418394804,
0.09445034712553024,
-0.10061565041542053,
-0.017126237973570824,
-0.058166440576314926,
-0.017597071826457977,
-0.056183844804763794,
-0.00373031641356647,
-0.08387861400842667,
-0.08524583280086517,
-3.2637038617344326e-33,
-0.014696416445076466,
-0.03136732801795006,
-0.03100244328379631,
0.009023021906614304,
-0.04886527732014656,
-0.11074980348348618,
-0.0004114647163078189,
0.028417052701115608,
-0.0038294640835374594,
-0.03108895756304264,
0.03080269694328308,
0.076375812292099,
-0.01396727655082941,
-0.016722893342375755,
0.016932610422372818,
-0.0061941202729940414,
0.0043243286199867725,
0.060545872896909714,
0.00012321348185651004,
0.029827678576111794,
0.07784722745418549,
0.05542425066232681,
-0.06918859481811523,
0.021999318152666092,
0.026213839650154114,
0.07529377192258835,
-0.028954757377505302,
-0.05215729400515556,
-0.10684870183467865,
0.010812154971063137,
-0.09694712609052658,
-0.02991481125354767,
0.011308728717267513,
-0.017994314432144165,
-0.040458641946315765,
0.04482539743185043,
0.06869184225797653,
-0.038255784660577774,
-0.01751287467777729,
0.05691324919462204,
0.022219447419047356,
0.018814075738191605,
-0.01208903081715107,
-0.012840858660638332,
-0.057485003024339676,
-0.03908636048436165,
-0.14968079328536987,
0.03895656391978264,
0.02094191312789917,
0.04350971803069115,
0.020380007103085518,
0.014742904342710972,
0.013366783037781715,
-0.10474088788032532,
0.0021378423552960157,
-0.06021704524755478,
-0.0882670134305954,
-0.020057126879692078,
0.015666795894503593,
0.06360885500907898,
0.04042505845427513,
0.03440957888960838,
-0.043876588344573975,
0.008951094932854176,
0.033705830574035645,
-0.008554765023291111,
0.0481417290866375,
-0.00032681008451618254,
-0.09782955795526505,
0.03382688760757446,
0.006785174831748009,
-0.10905278474092484,
0.024759486317634583,
0.1053057461977005,
0.0810231938958168,
0.06892899423837662,
-0.029041573405265808,
-0.041350554674863815,
-0.02525584027171135,
0.002635553013533354,
0.0005379294161684811,
-0.12275994569063187,
0.023752303794026375,
0.10670869052410126,
0.06760850548744202,
0.137617826461792,
-0.06331544369459152,
0.04526034742593765,
0.022808559238910675,
-0.0005242252955213189,
-0.024952305480837822,
0.05137171223759651,
0.06323301792144775,
0.067938432097435,
0.04844117909669876,
-4.806407005730762e-8,
0.006586347706615925,
0.04429268464446068,
-0.009098154492676258,
-0.04425763711333275,
0.014246299862861633,
-0.056330420076847076,
-0.04709245637059212,
0.04019390419125557,
-0.05693267285823822,
-0.06896226853132248,
0.06446294486522675,
-0.038040462881326675,
-0.07310256361961365,
-0.017468148842453957,
0.05121476948261261,
0.1138807088136673,
0.030522339046001434,
-0.005023670848459005,
-0.09749449044466019,
-0.07544250041246414,
0.0902630090713501,
0.046644870191812515,
0.09597449749708176,
-0.02522270195186138,
-0.06887468695640564,
-0.0006263475515879691,
-0.04205763339996338,
0.0649956613779068,
-0.00727652944624424,
0.027747992426156998,
0.04874043166637421,
0.07655812799930573,
0.05999301001429558,
-0.06309978663921356,
0.062205132097005844,
0.01577012985944748,
0.022431159391999245,
-0.00800088420510292,
-0.007901838049292564,
0.012413771823048592,
0.02825452946126461,
-0.05191303417086601,
-0.00991507712751627,
0.0520407073199749,
0.040717314928770065,
0.025906642898917198,
0.05795655772089958,
-0.011853847652673721,
0.0033843540586531162,
-0.03346658870577812,
0.03807150200009346,
-0.042388927191495895,
-0.022404059767723083,
0.08155903965234756,
-0.024626001715660095,
-0.048209164291620255,
0.03371758759021759,
0.12101081013679504,
0.08309356123209,
-0.013447376899421215,
0.09920977056026459,
-0.009746590629220009,
-0.022977298125624657,
-0.029330192133784294
] |
Generalizable Human Gaussians from Single-View Image | https://openreview.net/forum?id=dQ2xiSIYzp | [
"Jinnan Chen",
"Chen Li",
"Jianfeng Zhang",
"Lingting Zhu",
"Buzhen Huang",
"Hanlin Chen",
"Gim Hee Lee"
] | Poster | In this work, we tackle the task of learning 3D human Gaussians from a single image, focusing on recovering detailed appearance and geometry including unobserved regions. We introduce a single-view generalizable Human Gaussian Model (HGM), which employs a novel generate-then-refine pipeline with the guidance from human body prior and diffusion prior. Our approach uses a ControlNet to refine rendered back-view images from coarse predicted human Gaussians, then uses the refined image along with the input image to reconstruct refined human Gaussians. To mitigate the potential generation of unrealistic human poses and shapes, we incorporate human priors from the SMPL-X model as a dual branch, propagating image features from the SMPL-X volume to the image Gaussians using sparse convolution and attention mechanisms. Given that the initial SMPL-X estimation might be inaccurate, we gradually refine it with our HGM model. We validate our approach on several publicly available datasets. Our method surpasses previous methods in both novel view synthesis and surface reconstruction. Our approach also exhibits strong generalization for cross-dataset evaluation and in-the-wild images. | Human Gaussians, 3D Human Reconstruction | We design a novel approach to learn generalizable 3D human Gaussians from a single image, achieving both high-quality rendering and accurate 3D reconstruction. | 1,236 | 2406.06050 | [
-0.06496933847665787,
-0.10230430960655212,
0.0475059375166893,
0.010150846093893051,
0.10390466451644897,
-0.012078956700861454,
-0.0021809532772749662,
-0.040472742170095444,
-0.07004409283399582,
-0.033100537955760956,
-0.007166237570345402,
-0.04871014878153801,
0.01759415492415428,
0.03364644572138786,
-0.0407458059489727,
-0.05977476388216019,
0.09749959409236908,
0.09290169924497604,
-0.1020187959074974,
0.09523878991603851,
-0.03697774559259415,
0.017089398577809334,
0.044051915407180786,
-0.013598416931927204,
0.0836205706000328,
-0.06672778725624084,
0.06691670417785645,
-0.0684087797999382,
-0.00636332668364048,
-0.016659511253237724,
0.040185894817113876,
-0.04146138206124306,
-0.0027966315392404795,
-0.01767512783408165,
-0.040479566901922226,
0.06303022801876068,
-0.007479566615074873,
0.0600387342274189,
-0.013484267517924309,
-0.017709406092762947,
-0.007718509063124657,
0.013403436169028282,
-0.04706395044922829,
0.06632593274116516,
0.09711544215679169,
0.06403905898332596,
0.04168563708662987,
-0.07759617269039154,
0.08306378871202469,
0.01881246641278267,
0.031160404905676842,
-0.0035841509234160185,
0.010188708081841469,
0.008396664634346962,
-0.000003006477300004917,
0.004836436826735735,
-0.008662085048854351,
-0.08506972342729568,
0.03591630980372429,
-0.002239092020317912,
-0.025902127847075462,
-0.04483585059642792,
0.008758032694458961,
0.009861638769507408,
0.025178121402859688,
-0.08376619964838028,
0.010999520309269428,
-0.05759620666503906,
-0.025767920538783073,
0.04062624275684357,
-0.045430053025484085,
0.011511478573083878,
-0.02842535823583603,
-0.061328426003456116,
0.043724942952394485,
0.033291205763816833,
0.052306000143289566,
-0.05972390994429588,
0.023140911012887955,
-0.05347065627574921,
0.05261315032839775,
0.02088734693825245,
0.008267242461442947,
-0.031635135412216187,
0.008150887675583363,
0.04600760340690613,
-0.030299125239253044,
0.12551109492778778,
-0.006937519647181034,
0.02918098121881485,
-0.07969363033771515,
0.01070380862802267,
-0.018710991367697716,
-0.07761625945568085,
-0.024959541857242584,
-0.025932125747203827,
-0.0006368032190948725,
-0.07904397696256638,
-0.009267701767385006,
0.05251277610659599,
-0.0055674598552286625,
0.030220210552215576,
0.12483983486890793,
-0.03797061741352081,
0.04857686161994934,
-0.06171351298689842,
0.05824914574623108,
0.0497574619948864,
0.009840826503932476,
-0.03280726447701454,
0.09492640942335129,
-0.03895757347345352,
-0.009009160101413727,
0.04018034040927887,
0.06621026992797852,
-0.02930712327361107,
-0.015168306417763233,
-0.02030184492468834,
-0.024222439154982567,
-0.038551054894924164,
-0.015528135932981968,
0.04618879780173302,
0.03245930373668671,
0.06727678328752518,
0.039126694202423096,
-0.0383772999048233,
-0.039292555302381516,
4.3353430111550825e-34,
-0.034894637763500214,
0.0287666916847229,
0.09378707408905029,
0.05524703115224838,
0.0076879761181771755,
0.027300026267766953,
0.0073177395388484,
-0.008914653211832047,
-0.043290622532367706,
0.06926554441452026,
-0.02820473723113537,
-0.008091406896710396,
-0.051650989800691605,
0.14402173459529877,
-0.06017019972205162,
0.055224865674972534,
-0.04715555161237717,
0.036704666912555695,
0.005420369561761618,
0.09707554429769516,
0.007257032673805952,
-0.02842273749411106,
-0.03732921555638313,
-0.07032135128974915,
-0.058828048408031464,
-0.06419770419597626,
0.0642084926366806,
0.014851579442620277,
-0.026665033772587776,
-0.0038164749275892973,
-0.11792265623807907,
0.049766190350055695,
0.0019428030354902148,
0.044341955333948135,
-0.0018816872034221888,
-0.013690152205526829,
0.019050318747758865,
-0.038622334599494934,
-0.013889260590076447,
0.025363849475979805,
0.003652730491012335,
0.10147425532341003,
0.009485481306910515,
-0.02127031236886978,
-0.09399212896823883,
-0.0029750086832791567,
0.03599822148680687,
0.07260707765817642,
-0.04315795376896858,
-0.0443694032728672,
0.09213975071907043,
0.010815102607011795,
-0.02889157086610794,
-0.06011294573545456,
-0.11098411679267883,
-0.019553739577531815,
-0.0012006914475932717,
0.016965456306934357,
-0.006818745285272598,
-0.016841452568769455,
0.014197188429534435,
0.07030761241912842,
-0.06865165382623672,
0.05759517475962639,
-0.10016035288572311,
0.010839840397238731,
0.03436891734600067,
0.07155295461416245,
0.01598871499300003,
0.08516831696033478,
-0.06874769926071167,
0.005184318870306015,
-0.007202987093478441,
-0.0223105289041996,
0.07902573049068451,
0.007909415289759636,
0.05762811750173569,
-0.029071595519781113,
0.047696541994810104,
0.08723613619804382,
-0.037448905408382416,
0.021890785545110703,
-0.05792592838406563,
-0.006859533954411745,
-0.05696466192603111,
-0.04109090194106102,
0.05167838931083679,
-0.025847934186458588,
-0.03332725539803505,
-0.03961124271154404,
-0.06742588430643082,
-0.10371250659227371,
-0.039895620197057724,
-0.03591582924127579,
-0.04217110574245453,
9.475913083564745e-35,
0.051660966128110886,
0.028544805943965912,
0.0061564319767057896,
-0.02127612940967083,
0.01641787774860859,
-0.057467374950647354,
0.008266388438642025,
0.028428109362721443,
0.0020213632378727198,
-0.0881175622344017,
0.003971250262111425,
-0.052746281027793884,
0.0662020817399025,
-0.00045469062752090394,
0.03936082497239113,
-0.0011148324701935053,
-0.026825040578842163,
0.006780779920518398,
-0.0757550522685051,
-0.027138153091073036,
0.003112287260591984,
0.045158833265304565,
-0.016011156141757965,
-0.006788183469325304,
-0.12689520418643951,
0.11032194644212723,
-0.008906914852559566,
0.06975427269935608,
-0.030952926725149155,
-0.020091187208890915,
-0.07336246967315674,
-0.004542844370007515,
0.007612431421875954,
-0.0448710173368454,
-0.1182480975985527,
0.05719602480530739,
0.030441313982009888,
0.06738599389791489,
-0.01570679061114788,
0.03189424052834511,
0.023557286709547043,
-0.00757580203935504,
0.011693199165165424,
0.08909177780151367,
-0.015210296958684921,
0.04131205752491951,
0.022727366536855698,
-0.03469778969883919,
-0.04164780303835869,
0.10072928667068481,
-0.10748331248760223,
0.11017787456512451,
-0.11308876425027847,
-0.00457532424479723,
-0.041804708540439606,
-0.076181560754776,
-0.028842657804489136,
-0.0687621682882309,
0.09378275275230408,
-0.038537170737981796,
-0.04738369584083557,
-0.0465015172958374,
-0.08038214594125748,
0.02177613228559494,
0.003961830399930477,
0.014339853078126907,
0.007815922610461712,
-0.012529213912785053,
-0.012500995770096779,
0.04491473734378815,
0.02957596816122532,
-0.044805627316236496,
0.021599866449832916,
0.01328014675527811,
0.056088607758283615,
-0.03829088434576988,
0.010906107723712921,
0.02197330817580223,
0.0897533968091011,
-0.010060355067253113,
0.10575069487094879,
-0.07357202470302582,
0.019169308245182037,
0.07504548877477646,
0.14313103258609772,
0.04277953878045082,
0.039748240262269974,
0.03897230699658394,
0.025838498026132584,
-0.0167141892015934,
-0.10146154463291168,
0.012565151788294315,
-0.02780616097152233,
0.06517469882965088,
-0.007056587375700474,
-3.8305223881707207e-8,
-0.07884039729833603,
0.06967775523662567,
0.06664807349443436,
-0.0007160379900597036,
-0.040988739579916,
-0.009748335927724838,
-0.009703815914690495,
0.06125878915190697,
-0.04940294101834297,
-0.04645121842622757,
-0.03946274518966675,
-0.024923643097281456,
0.0009217243641614914,
-0.060279421508312225,
0.02315901592373848,
0.08478805422782898,
-0.06949793547391891,
0.04169650003314018,
-0.03797456622123718,
-0.009917314164340496,
-0.05564189702272415,
0.010159911587834358,
0.008760597556829453,
-0.02095399610698223,
0.07719294726848602,
-0.026491621509194374,
-0.06367316842079163,
0.07955983281135559,
0.0633639246225357,
-0.029166078194975853,
0.009876876138150692,
0.007735402323305607,
-0.01664530485868454,
-0.045460838824510574,
-0.01852857880294323,
-0.025128183886408806,
-0.009911200031638145,
0.007648455444723368,
0.06721196323633194,
-0.04015205428004265,
-0.02817997895181179,
0.07551971077919006,
0.03525623306632042,
0.022838372737169266,
0.001623252290301025,
0.05856189504265785,
0.049573514610528946,
-0.044462550431489944,
0.04463744908571243,
0.006837120279669762,
-0.005747349001467228,
-0.06044602021574974,
0.0868515893816948,
0.11259172856807709,
0.055365364998579025,
0.047571007162332535,
0.02847849763929844,
-0.02793256752192974,
0.0688493400812149,
0.04753545671701431,
-0.02742128260433674,
0.02985254116356373,
-0.06929967552423477,
0.015655044466257095
] |
Wavelet Diffusion Neural Operator | https://openreview.net/forum?id=FQhDIGuaJ4 | [
"Peiyan Hu",
"Rui Wang",
"Xiang Zheng",
"Tao Zhang",
"Haodong Feng",
"Ruiqi Feng",
"Long Wei",
"Yue Wang",
"Zhi-Ming Ma",
"Tailin Wu"
] | Poster | Simulating and controlling physical systems described by partial differential equations (PDEs) are crucial tasks across science and engineering. Recently, diffusion generative models have emerged as a competitive class of methods for these tasks due to their ability to capture long-term dependencies and model high-dimensional states. However, diffusion models typically struggle with handling system states with abrupt changes and generalizing to higher resolutions. In this work, we propose Wavelet Diffusion Neural Operator (WDNO), a novel PDE simulation and control framework that enhances the handling of these complexities. WDNO comprises two key innovations. Firstly, WDNO performs diffusion-based generative modeling in the wavelet domain for the entire trajectory to handle abrupt changes and long-term dependencies effectively. Secondly, to address the issue of poor generalization across different resolutions, which is one of the fundamental tasks in modeling physical systems, we introduce multi-resolution training. We validate WDNO on five physical systems, including 1D advection equation, three challenging physical systems with abrupt changes (1D Burgers' equation, 1D compressible Navier-Stokes equation and 2D incompressible fluid), and a real-world dataset ERA5, which demonstrates superior performance on both simulation and control tasks over state-of-the-art methods, with significant improvements in long-term and detail prediction accuracy. Remarkably, in the challenging context of the 2D high-dimensional and indirect control task aimed at reducing smoke leakage, WDNO reduces the leakage by 33.2% compared to the second-best baseline. | PDE, physics, simulation, control, diffusion model, wavelet, abrupt changes, multi-resolution | We propose Wavelet Diffusion Neural Operator (WDNO), a novel method for generative PDE simulation and control, to address diffusion models' challenges of modeling system states with abrupt changes and generalizing to higher resolutions. | 1,235 | 2412.04833 | [
-0.07368649542331696,
-0.12056193500757217,
0.11322004348039627,
0.031753506511449814,
-0.003115872386842966,
-0.04694109782576561,
-0.10859239846467972,
-0.06935566663742065,
-0.0021495006512850523,
-0.02686794474720955,
-0.027232883498072624,
-0.023627860471606255,
-0.08843255788087845,
-0.05487304553389549,
-0.047142334282398224,
-0.034590303897857666,
-0.0004253661318216473,
0.0571141242980957,
-0.0767766535282135,
-0.0582447312772274,
0.05678495392203331,
0.04412924498319626,
-0.053532615303993225,
-0.01529909111559391,
0.031294241547584534,
0.07966950535774231,
0.06132316589355469,
-0.026502884924411774,
0.0005536666722036898,
-0.0027546498458832502,
0.10534381866455078,
-0.014563307166099548,
-0.12258981913328171,
0.024630600586533546,
-0.10018974542617798,
0.040856875479221344,
-0.07212650775909424,
0.009646132588386536,
-0.07167986780405045,
-0.10072685033082962,
0.07993566989898682,
0.0381719134747982,
-0.007984844967722893,
0.04737279936671257,
0.06191292032599449,
-0.021572932600975037,
-0.01857813261449337,
0.02063468098640442,
-0.006113999057561159,
-0.050924770534038544,
0.00633835606276989,
-0.05276772752404213,
-0.005815985146909952,
0.03625226020812988,
0.0243497584015131,
-0.0007579349912703037,
0.04398713260889053,
0.030581219121813774,
0.025485079735517502,
-0.027475375682115555,
-0.08696379512548447,
-0.0023841129150241613,
0.05211611092090607,
-0.036593835800886154,
0.008575302548706532,
0.0011357804760336876,
0.020203392952680588,
0.12193214893341064,
0.011214898899197578,
0.06152266636490822,
-0.0009219787316396832,
0.024765389040112495,
-0.03801366686820984,
-0.02216971106827259,
0.03380468115210533,
-0.0461791530251503,
-0.00969131849706173,
0.020137164741754532,
0.03761708736419678,
0.01842690072953701,
0.0252701248973608,
0.04337669536471367,
0.06407933682203293,
-0.055513929575681686,
0.023009303957223892,
0.024621112272143364,
-0.030306417495012283,
0.05735760182142258,
0.046270132064819336,
-0.011146190576255322,
-0.004043112508952618,
0.037518542259931564,
0.021402541548013687,
-0.009124953299760818,
-0.0004555262275971472,
-0.029514873400330544,
0.00655362056568265,
0.0027864950243383646,
0.022455312311649323,
0.024474114179611206,
0.00043537700548768044,
0.027925146743655205,
-0.002627014648169279,
0.08003940433263779,
0.10503977537155151,
-0.048333652317523956,
0.09313791990280151,
0.02777092717587948,
0.026127571240067482,
-0.003907792270183563,
0.04379682242870331,
0.01530963834375143,
0.004111104179173708,
0.02784704975783825,
0.08358418196439743,
-0.12861627340316772,
-0.010063419118523598,
-0.03808675333857536,
-0.012097357772290707,
-0.03506749868392944,
-0.06425847113132477,
-0.05386479198932648,
-0.07900957018136978,
-0.014450937509536743,
0.007324158679693937,
-0.03173435106873512,
0.0023807515390217304,
7.670594714084756e-34,
0.004851258359849453,
-0.025738663971424103,
-0.06687116622924805,
0.08238358050584793,
0.05909836292266846,
0.011135846376419067,
-0.009294033981859684,
-0.004206161480396986,
0.005120768677443266,
-0.0361456461250782,
-0.0674666091799736,
0.014012747444212437,
-0.14361777901649475,
0.09782377630472183,
0.029459292069077492,
-0.02921099215745926,
0.004275673534721136,
0.03186102211475372,
0.009734971448779106,
0.02709590643644333,
0.06838694214820862,
-0.03208170086145401,
-0.053000833839178085,
-0.06494858115911484,
-0.006756654940545559,
0.06440067291259766,
-0.03736867755651474,
0.01319670956581831,
0.03060026839375496,
0.0121700344607234,
-0.08942610770463943,
-0.02831159345805645,
-0.09517402946949005,
0.056675612926483154,
-0.00879678875207901,
0.028721358627080917,
-0.05478984862565994,
0.004768567159771919,
-0.029163436964154243,
-0.003472523530945182,
-0.051799263805150986,
-0.02284500189125538,
-0.027075914666056633,
0.08881576359272003,
-0.06432728469371796,
-0.08810090273618698,
0.06713744252920151,
0.004099685698747635,
-0.024422965943813324,
-0.06037747114896774,
-0.048479825258255005,
0.002655930584296584,
0.07727673649787903,
-0.13126260042190552,
0.010009276680648327,
0.008682304061949253,
0.11041399091482162,
-0.043722692877054214,
0.013608102686703205,
-0.013704631477594376,
-0.06159781664609909,
0.03855188935995102,
0.026958895847201347,
0.029069742187857628,
0.026289241388440132,
0.04395350441336632,
-0.0338422916829586,
0.08628814667463303,
0.007323862053453922,
0.014846701174974442,
0.039662931114435196,
0.01216970570385456,
0.06367333978414536,
-0.0723985806107521,
0.11407468467950821,
-0.02204660139977932,
0.13264083862304688,
-0.008097123354673386,
-0.09107580035924911,
0.03462772071361542,
-0.006654057651758194,
-0.0175735205411911,
0.02959907054901123,
-0.012266764417290688,
0.050762198865413666,
-0.004767156206071377,
-0.012558316811919212,
-0.042104922235012054,
-0.07925239205360413,
-0.11525004357099533,
-0.08801136910915375,
-0.028247376903891563,
0.05551594868302345,
0.007958333007991314,
0.03427533805370331,
-2.9903083039751465e-33,
0.0022958405315876007,
0.09238458424806595,
0.006130474153906107,
0.007521776482462883,
-0.052175819873809814,
0.03958786651492119,
-0.09018521755933762,
0.06413690000772476,
0.03125837817788124,
-0.11549641191959381,
-0.05314909666776657,
-0.09584386646747589,
-0.022460168227553368,
-0.02443891391158104,
0.025098923593759537,
0.011911515146493912,
0.009552009403705597,
-0.09221424162387848,
0.024729249998927116,
-0.02990284375846386,
-0.03590381145477295,
0.06512890011072159,
-0.054458875209093094,
0.05980892479419708,
-0.08311795443296432,
-0.0016837414586916566,
-0.006270742509514093,
0.05901874229311943,
0.003764679189771414,
-0.057664938271045685,
-0.03367456793785095,
-0.04076707363128662,
0.07262647151947021,
-0.019691674038767815,
-0.040903761982917786,
0.14023055136203766,
0.08926727622747421,
-0.02799125574529171,
-0.015725256875157356,
0.0130600705742836,
0.050652727484703064,
-0.08783725649118423,
0.009538300335407257,
0.05954587087035179,
0.005825825035572052,
0.01685216650366783,
0.010237687267363071,
-0.01965968869626522,
-0.016964765265583992,
0.03670375794172287,
-0.0033114394173026085,
0.06249347701668739,
-0.07660698890686035,
0.028630651533603668,
0.03971176967024803,
-0.019870415329933167,
0.0025047012604773045,
0.004821168724447489,
0.0316917710006237,
-0.05154695734381676,
-0.02251746505498886,
-0.08889587223529816,
0.04345829412341118,
0.03302163630723953,
-0.0703582912683487,
-0.04800192639231682,
-0.0935886949300766,
0.04020095616579056,
0.06569972634315491,
0.01902909390628338,
-0.037850432097911835,
0.02300751581788063,
-0.0132713932543993,
-0.04454079270362854,
0.04785826429724693,
-0.10362817347049713,
-0.004516862332820892,
-0.055183038115501404,
0.007123399991542101,
-0.0005022817058488727,
0.03449719771742821,
-0.0002852940233424306,
0.02732846699655056,
0.031258080154657364,
0.008216440677642822,
-0.01602981425821781,
0.0484497956931591,
0.019099852070212364,
-0.0011855711927637458,
0.019562887027859688,
0.0585392527282238,
0.027625633403658867,
-0.013732731342315674,
0.00001763104955898598,
-0.08951719850301743,
-4.6984457213739006e-8,
-0.026479605585336685,
0.09302347153425217,
0.10070249438285828,
-0.03542301058769226,
0.06023281440138817,
-0.026383008807897568,
0.015185459516942501,
0.01105946023017168,
0.014997358433902264,
-0.00534382089972496,
0.0336184948682785,
-0.06184300407767296,
-0.036788083612918854,
-0.07919473946094513,
0.028450477868318558,
0.018370043486356735,
-0.0032663203310221434,
-0.03019152581691742,
-0.018274448812007904,
-0.016970908269286156,
0.042070042341947556,
0.0639052763581276,
0.0037326347082853317,
-0.03982332721352577,
0.11993597447872162,
-0.031591836363077164,
-0.039473455399274826,
-0.0067232768051326275,
0.00700979121029377,
0.009588973596692085,
-0.07106425613164902,
0.02370436117053032,
0.0028970104176551104,
0.06499705463647842,
0.015395000576972961,
0.08692901581525803,
0.10139760375022888,
0.019360920414328575,
-0.04476721212267876,
-0.033555805683135986,
-0.09356359392404556,
0.07940693944692612,
-0.028841940686106682,
0.04881821200251579,
-0.007303720340132713,
0.06698217242956161,
0.0755932629108429,
-0.059631042182445526,
0.07856008410453796,
0.08200838416814804,
0.017054196447134018,
0.032202158123254776,
0.03595612943172455,
0.01534141506999731,
0.03175157681107521,
0.03762219846248627,
-0.04214402288198471,
-0.05187787488102913,
0.05434040352702141,
0.03031781129539013,
-0.024862835183739662,
0.02154507115483284,
-0.04728353023529053,
0.05426234379410744
] |
MMSearch: Unveiling the Potential of Large Models as Multi-modal Search Engines | https://openreview.net/forum?id=J2Jyp1SZ0n | [
"Dongzhi Jiang",
"Renrui Zhang",
"Ziyu Guo",
"Yanmin Wu",
"jiayi lei",
"Pengshuo Qiu",
"Pan Lu",
"Zehui Chen",
"Guanglu Song",
"Peng Gao",
"Yu Liu",
"Chunyuan Li",
"Hongsheng Li"
] | Poster | The advent of Large Language Models (LLMs) has paved the way for AI search engines, e.g., SearchGPT, showcasing a new paradigm in human-internet interaction. However, most current AI search engines are limited to text-only settings, neglecting the multimodal user queries and the text-image interleaved nature of website information. Recently, Large Multimodal Models (LMMs) have made impressive strides. Yet, whether they can function as AI search engines remains under-explored, leaving the potential of LMMs in multimodal search an open question. To this end, we first design a delicate pipeline, MMSearch-Engine, to empower any LMMs with multimodal search capabilities. On top of this, we introduce MMSearch, a comprehensive evaluation benchmark to assess the multimodal search performance of LMMs. The curated dataset contains 300 manually collected instances spanning 14 subfields, which involves no overlap with the current LMMs' training data, ensuring the correct answer can only be obtained within searching. By using MMSearch-Engine, the LMMs are evaluated by performing three individual tasks (requery, rerank, and summarization), and one challenging end-to-end task with a complete searching process. We conduct extensive experiments on closed-source and open-source LMMs. Among all tested models, GPT-4o with MMSearch-Engine achieves the best results, which surpasses the commercial product, Perplexity Pro, in the end-to-end task, demonstrating the effectiveness of our proposed pipeline. We further present error analysis to unveil current LMMs still struggle to fully grasp the multimodal search tasks, and conduct ablation study to indicate the potential of scaling test-time computation for AI search engine. We hope MMSearch may provide unique insights to guide the future development of multimodal AI search engine. | Large Multimodal Model, AI Search Engine, Benchmark | null | 1,234 | null | [
-0.02726539596915245,
-0.10620545595884323,
0.04253918305039406,
0.04080415889620781,
0.10524323582649231,
-0.001733017386868596,
-0.08020275831222534,
0.01700432598590851,
-0.008731693029403687,
-0.09575773030519485,
-0.011084252968430519,
-0.06400339305400848,
0.06112705171108246,
0.017443373799324036,
0.004996278323233128,
0.004064671695232391,
0.034024596214294434,
0.07140732556581497,
-0.07104820758104324,
-0.016302185133099556,
0.06678889691829681,
0.013763993978500366,
0.07050846517086029,
-0.12098975479602814,
0.04539615660905838,
0.04032359644770622,
-0.03720640391111374,
-0.07690834254026413,
0.02757149748504162,
-0.04697559401392937,
0.012209190987050533,
0.03343873471021652,
0.022703932598233223,
0.09244649112224579,
-0.04817582666873932,
0.014915404841303825,
-0.09219084680080414,
-0.005113998427987099,
0.011895069852471352,
-0.05448183789849281,
-0.050212014466524124,
-0.023974530398845673,
0.041611868888139725,
-0.015741122886538506,
0.05185800418257713,
0.010726440697908401,
-0.0584850087761879,
-0.051667723804712296,
0.047303542494773865,
0.03876076638698578,
-0.13596320152282715,
-0.0681309774518013,
0.0073571158573031425,
0.09791234880685806,
-0.024487311020493507,
-0.0693473294377327,
-0.025133293122053146,
0.004658883903175592,
0.019969843327999115,
-0.031416118144989014,
-0.004351831506937742,
-0.037017591297626495,
-0.05003588646650314,
0.0009649041458033025,
0.0667700320482254,
-0.04924301430583,
0.03628143295645714,
-0.05547831207513809,
0.06420501321554184,
-0.02044074982404709,
0.00549420528113842,
0.01669767126441002,
-0.05205661803483963,
0.04946690797805786,
-0.0016886423109099269,
0.03379387408494949,
0.0197149608284235,
-0.022190799936652184,
0.02695910446345806,
-0.03436543792486191,
0.041414160281419754,
-0.05170969292521477,
0.05910232663154602,
-0.04694933444261551,
0.07135861366987228,
-0.03922455385327339,
-0.008077393285930157,
-0.032841019332408905,
0.023636389523744583,
0.0016145141562446952,
-0.03376219421625137,
-0.04010383412241936,
0.020452437922358513,
-0.02954600751399994,
0.04717268794775009,
0.0070083909668028355,
0.052491117268800735,
-0.05771179124712944,
-0.006176447495818138,
0.05691799893975258,
0.045052241533994675,
0.07107541710138321,
-0.03606313839554787,
-0.04024408757686615,
-0.029476650059223175,
0.04140932857990265,
0.08217822760343552,
0.04347086697816849,
0.05070467293262482,
-0.06181558594107628,
0.00684631010517478,
0.0475635752081871,
0.042852096259593964,
-0.06410907208919525,
0.05833253264427185,
-0.04796589910984039,
0.04808482527732849,
0.012666112743318081,
0.094284787774086,
0.05029825493693352,
-0.027148818597197533,
0.02499953843653202,
0.01452697440981865,
-0.02884380705654621,
0.036160875111818314,
0.026689773425459862,
-0.11176717281341553,
2.9306780469518243e-33,
0.11942202597856522,
-0.0017351586138829589,
0.022297341376543045,
0.06201699748635292,
0.018342506140470505,
-0.0010282084112986922,
-0.01440861914306879,
0.08314231783151627,
-0.120575912296772,
-0.03863467276096344,
-0.03908626735210419,
0.0662350282073021,
-0.08196272701025009,
0.059088192880153656,
0.00017778734036255628,
0.002430469961836934,
-0.010969966650009155,
-0.026649240404367447,
0.003037712536752224,
0.03970229998230934,
0.04362490028142929,
-0.044967181980609894,
0.031600844115018845,
0.016274429857730865,
0.02387436479330063,
-0.009691023267805576,
0.07310713827610016,
-0.11658928543329239,
-0.0048373108729720116,
0.028240330517292023,
-0.14640477299690247,
-0.023654798045754433,
-0.04325784742832184,
-0.0022216858342289925,
0.030056197196245193,
-0.0023952326737344265,
-0.07603511214256287,
-0.015444736927747726,
-0.018285900354385376,
-0.03341320902109146,
-0.07386414706707001,
-0.044538918882608414,
0.005454030353575945,
-0.06837289780378342,
-0.11063570529222488,
-0.023738058283925056,
-0.10113053768873215,
-0.020602956414222717,
-0.0056084319949150085,
-0.023901185020804405,
0.08938205242156982,
0.06014922261238098,
-0.05119907855987549,
-0.012378056533634663,
0.06182737275958061,
0.023422924801707268,
-0.005036478396505117,
0.04769153147935867,
0.04459603503346443,
0.004462765995413065,
-0.009611456654965878,
0.025344369933009148,
0.06807158142328262,
0.08949845284223557,
0.03502029925584793,
0.03396066278219223,
-0.008978883735835552,
0.0553535595536232,
0.06643408536911011,
0.04672667011618614,
-0.017435627058148384,
0.009107059799134731,
0.048534806817770004,
-0.052213847637176514,
-0.03635872155427933,
-0.015389990992844105,
0.03642203286290169,
-0.06650668382644653,
0.05532577633857727,
0.04444761946797371,
0.05372679978609085,
0.0011299415491521358,
0.039062730967998505,
-0.06989167630672455,
0.013410752639174461,
-0.052812252193689346,
-0.018380843102931976,
-0.1207159161567688,
-0.025604259222745895,
-0.038351330906152725,
-0.02601538598537445,
0.035245951265096664,
-0.06476990133523941,
-0.018597513437271118,
-0.04888492077589035,
-2.5099697208184815e-33,
0.018009740859270096,
-0.0439334474503994,
-0.018763020634651184,
0.03941423073410988,
-0.01417914405465126,
-0.048587553203105927,
0.02159956656396389,
0.004437427967786789,
-0.013888681307435036,
0.00008103481377474964,
0.030341841280460358,
0.006319954060018063,
0.011268699541687965,
-0.05668535828590393,
0.007074660621583462,
0.035444725304841995,
0.0867653414607048,
-0.045521367341279984,
-0.01787714473903179,
0.05566047132015228,
-0.027720225974917412,
0.08202730864286423,
-0.08685702830553055,
0.06369700282812119,
-0.11722493171691895,
0.02809840813279152,
-0.07435853779315948,
0.0022271950729191303,
-0.07631813734769821,
-0.039320215582847595,
0.029181228950619698,
-0.04008456692099571,
-0.024684786796569824,
-0.05259377136826515,
-0.03367048129439354,
0.0864613801240921,
0.017658881843090057,
-0.024808019399642944,
-0.010977962985634804,
0.15330685675144196,
0.10589175671339035,
0.040740713477134705,
-0.044609155505895615,
-0.01475462131202221,
0.02662799321115017,
-0.07608553767204285,
-0.11570589244365692,
-0.01654645800590515,
0.057863179594278336,
0.0388685017824173,
-0.01586379110813141,
-0.03707891330122948,
-0.08488668501377106,
-0.05612201616168022,
-0.020445585250854492,
-0.08489149808883667,
-0.06152080371975899,
0.017865242436528206,
0.010082455351948738,
0.06198716163635254,
-0.04262559488415718,
-0.07447411864995956,
-0.08531317859888077,
-0.01656888797879219,
0.002115015871822834,
-0.022769013419747353,
-0.014487440697848797,
-0.04965532571077347,
-0.01625365950167179,
-0.029017455875873566,
0.045234452933073044,
-0.06009091064333916,
0.027636554092168808,
0.06263736635446548,
0.018499065190553665,
0.0648002177476883,
0.08704390376806259,
0.048818059265613556,
0.04113767668604851,
-0.09271807968616486,
-0.010326034389436245,
0.014160441234707832,
0.0195726677775383,
0.03568755462765694,
0.050476137548685074,
0.09206566959619522,
-0.028005652129650116,
0.08347807824611664,
0.021114768460392952,
-0.004303842782974243,
0.035409312695264816,
-0.004401915706694126,
0.019364412873983383,
0.04321928694844246,
0.04080461710691452,
-3.6769360889366e-8,
-0.03758775070309639,
0.07129787653684616,
-0.018212947994470596,
0.007985690608620644,
0.047126904129981995,
0.005049684550613165,
-0.005831163376569748,
0.07490958273410797,
0.013982141390442848,
-0.02327129617333412,
0.0573299303650856,
-0.018868491053581238,
-0.029344040900468826,
0.03983555734157562,
0.03645021840929985,
0.04724840074777603,
-0.0697808638215065,
-0.019135572016239166,
-0.023741265758872032,
-0.03784249350428581,
0.09345272928476334,
0.02548792026937008,
0.029833443462848663,
0.042895153164863586,
0.023239336907863617,
0.019677966833114624,
-0.07777252793312073,
0.04919826611876488,
-0.039139363914728165,
0.038151707500219345,
-0.05995340645313263,
0.036536362022161484,
-0.06425029784440994,
-0.055567532777786255,
0.01008060947060585,
0.023576674982905388,
-0.011081194505095482,
-0.06830262392759323,
-0.07819876819849014,
-0.01494363509118557,
0.08369100838899612,
0.05140935257077217,
-0.020088687539100647,
0.0283231008797884,
0.05089922249317169,
0.0353267677128315,
0.011617670767009258,
-0.08068814873695374,
0.040377769619226456,
-0.02678743377327919,
-0.0351695753633976,
-0.04995918646454811,
0.030243415385484695,
0.1297755241394043,
0.0835568755865097,
-0.04761522263288498,
0.011769119650125504,
0.06871121376752853,
0.11225917190313339,
0.04416289180517197,
0.05049381032586098,
-0.03851613029837608,
-0.05480797588825226,
0.018751803785562515
] |
Coreset Selection via Reducible Loss in Continual Learning | https://openreview.net/forum?id=mAztx8QO3B | [
"Ruilin Tong",
"Yuhang Liu",
"Javen Qinfeng Shi",
"Dong Gong"
] | Poster | Rehearsal-based continual learning (CL) aims to mitigate catastrophic forgetting by maintaining a subset of samples from previous tasks and replaying them. The rehearsal memory can be naturally constructed as a coreset, designed to form a compact subset that enables training with performance comparable to using the full dataset. The coreset selection task can be formulated as bilevel optimization that solves for the subset to minimize the outer objective of the learning task. Existing methods primarily rely on inefficient probabilistic sampling or local gradient-based scoring to approximate sample importance through an iterative process that can be susceptible to ambiguity or noise. Specifically, non-representative samples like ambiguous or noisy samples are difficult to learn and incur high loss values even when training on the full dataset. However, existing methods relying on local gradient tend to highlight these samples in an attempt to minimize the outer loss, leading to a suboptimal coreset. To enhance coreset selection, especially in CL where high-quality samples are essential, we propose a coreset selection method that measures sample importance using reducible loss (ReL) that quantifies the impact of adding a sample to model performance. By leveraging ReL and a process derived from bilevel optimization, we identify and retain samples that yield the highest performance gain. They are shown to be informative and representative. Furthermore, ReL requires only forward computation, making it significantly more efficient than previous methods. To better apply coreset selection in CL, we extend our method to address key challenges such as task interference, streaming data, and knowledge distillation. Experiments on data summarization and continual learning demonstrate the effectiveness and efficiency of our approach. | Continual learning, Coreset selection | null | 1,231 | null | [
0.028762580826878548,
-0.07083132117986679,
0.039292529225349426,
0.05140434950590134,
0.06957114487886429,
0.010463759303092957,
0.01211134996265173,
-0.039763227105140686,
0.0020795317832380533,
-0.0511002354323864,
-0.04253731667995453,
-0.024416448548436165,
-0.01127004623413086,
-0.04120349511504173,
-0.07244851440191269,
-0.015244961716234684,
0.03124930150806904,
0.08349684625864029,
-0.05072528496384621,
-0.08058347553014755,
-0.006410121917724609,
0.009057516232132912,
0.07727892696857452,
0.05859098955988884,
0.01970726251602173,
-0.045261893421411514,
-0.0036993417888879776,
0.00011089551117038354,
0.04829315096139908,
-0.018880151212215424,
0.020648209378123283,
0.04085274413228035,
-0.020839480683207512,
0.002031918615102768,
-0.03750499710440636,
0.04954266548156738,
-0.06776031851768494,
0.06631492078304291,
0.04838842898607254,
-0.002256908221170306,
0.028780963271856308,
0.0474369153380394,
-0.008950658142566681,
0.00895483884960413,
0.01819760352373123,
-0.01758444868028164,
0.014631139114499092,
-0.07155909389257431,
0.03384917601943016,
0.07418985664844513,
-0.0054879761300981045,
-0.011120482347905636,
-0.042626846581697464,
0.03516697511076927,
0.05802702158689499,
0.020090023055672646,
0.05973445996642113,
0.018894493579864502,
-0.06427556276321411,
0.004852559417486191,
-0.014959568157792091,
-0.14979009330272675,
-0.00580199621617794,
-0.07017624378204346,
-0.1031707152724266,
-0.04004796966910362,
0.07977854460477829,
0.0663113072514534,
0.03346571698784828,
0.07315587997436523,
-0.03142603114247322,
0.07482189685106277,
-0.07770173996686935,
0.0377841480076313,
0.015302050858736038,
0.06878748536109924,
0.02846166305243969,
-0.05808088555932045,
0.021179847419261932,
-0.03316526487469673,
-0.003618166781961918,
0.06942278891801834,
0.04031747952103615,
-0.08270744234323502,
0.153885155916214,
-0.009003577753901482,
0.10375150293111801,
-0.03725454583764076,
0.028459714725613594,
-0.022116530686616898,
-0.07848604023456573,
-0.01699921116232872,
-0.007085015531629324,
-0.05996508523821831,
0.05400692671537399,
0.0025657822843641043,
0.0452616810798645,
-0.03988608717918396,
-0.0074011958204209805,
0.05329518765211105,
0.024350769817829132,
0.04877825453877449,
-0.08506596833467484,
-0.0019023222848773003,
-0.02640705555677414,
-0.08896517753601074,
-0.02418465167284012,
0.09805181622505188,
0.03496190533041954,
-0.07218310981988907,
0.036484651267528534,
0.001646155957132578,
0.02450462616980076,
0.03865089267492294,
0.09039156883955002,
0.008683853782713413,
0.0385076105594635,
0.0008201855816878378,
-0.05814315378665924,
0.11922578513622284,
-0.015885494649410248,
0.026452165096998215,
-0.022479571402072906,
0.01626303419470787,
-0.0078895203769207,
-0.033251307904720306,
-0.0023896850179880857,
3.715455891647038e-33,
0.020830651745200157,
0.013187846168875694,
0.012382416054606438,
-0.0410614013671875,
0.02305210381746292,
-0.03868323192000389,
0.017034431919455528,
0.010146625339984894,
0.03631100431084633,
-0.025067618116736412,
0.019254840910434723,
-0.05739200860261917,
-0.0705903097987175,
0.06552271544933319,
0.06400085240602493,
-0.02189297415316105,
-0.04655846580862999,
0.08401289582252502,
-0.025498168542981148,
-0.034635335206985474,
-0.015694301575422287,
0.01711237244307995,
-0.030700452625751495,
-0.09877768158912659,
0.04916961118578911,
-0.02388303354382515,
0.060411304235458374,
0.034744054079055786,
-0.06709661334753036,
0.015599098056554794,
-0.08462720364332199,
-0.017106665298342705,
0.028759820386767387,
0.00003764458961086348,
0.04162457585334778,
-0.004620164632797241,
0.020730935037136078,
-0.04197072982788086,
0.042868148535490036,
-0.037349581718444824,
0.017849929630756378,
0.043050482869148254,
-0.0007482719956897199,
-0.008090600371360779,
-0.0669899582862854,
-0.16891226172447205,
0.027750452980399132,
0.023149477317929268,
-0.009618512354791164,
-0.0754363089799881,
0.015128757804632187,
0.010028536431491375,
0.030176475644111633,
-0.03318149223923683,
-0.042583078145980835,
0.022019393742084503,
0.01291930302977562,
-0.012996826320886612,
0.047130949795246124,
0.029549697414040565,
0.017720375210046768,
0.027684882283210754,
-0.06540131568908691,
0.16003957390785217,
0.020087512210011482,
0.06095895171165466,
-0.037022870033979416,
0.01316128484904766,
-0.008336848579347134,
-0.056505508720874786,
0.022284939885139465,
-0.015465731732547283,
-0.006901418790221214,
-0.1090865358710289,
0.08040458709001541,
-0.029213402420282364,
0.09080284833908081,
-0.007269144989550114,
-0.015767347067594528,
-0.04080220311880112,
0.046057719737291336,
0.0006696484051644802,
-0.054600562900304794,
-0.1287313848733902,
-0.06057320162653923,
-0.011909525841474533,
0.0883217602968216,
-0.08267941325902939,
-0.033171892166137695,
0.023506687954068184,
-0.10497352480888367,
0.03247074410319328,
-0.012466276995837688,
-0.05822159722447395,
-0.07022291421890259,
-3.048938839787293e-33,
-0.04735729098320007,
0.021490540355443954,
-0.009923173114657402,
0.07797158509492874,
0.0003987675590906292,
-0.0062540266662836075,
-0.019687630236148834,
-0.060680367052555084,
-0.08667896687984467,
-0.14435052871704102,
-0.029280846938490868,
-0.06989473849534988,
-0.011090777814388275,
-0.003644681768491864,
-0.05067813768982887,
0.05699484422802925,
-0.055720556527376175,
0.06134350597858429,
-0.021678874269127846,
0.008933632634580135,
0.027174822986125946,
0.005372523330152035,
0.015811122953891754,
0.009706458076834679,
0.007001339457929134,
0.04524989053606987,
-0.06882931292057037,
0.113475002348423,
0.04324137419462204,
-0.09765692800283432,
-0.039435166865587234,
0.006001901347190142,
-0.040304943919181824,
-0.0719689279794693,
-0.027934150770306587,
0.029949015006422997,
0.043629661202430725,
0.031394679099321365,
-0.0678580030798912,
0.15441106259822845,
0.05423101782798767,
0.017945194616913795,
-0.016971256583929062,
0.034191325306892395,
0.024350648745894432,
-0.0877390056848526,
-0.03522634878754616,
0.04845377057790756,
0.018304243683815002,
0.03469991683959961,
0.012971903197467327,
0.007267652545124292,
-0.1275741010904312,
0.040388450026512146,
-0.027556119486689568,
-0.022812852635979652,
-0.03435666859149933,
-0.061521101742982864,
0.004152875859290361,
-0.02040754072368145,
-0.0384427011013031,
-0.029535068199038506,
-0.0014748654793947935,
-0.008620291016995907,
0.06346714496612549,
0.01930522546172142,
0.01521819457411766,
0.030624017119407654,
-0.033097393810749054,
0.061038099229335785,
0.027415383607149124,
0.06448132544755936,
0.008306353352963924,
-0.06361894309520721,
0.011660560965538025,
0.024925384670495987,
-0.014306986704468727,
-0.0267800185829401,
-0.028203103691339493,
-0.0038525769487023354,
-0.014301007613539696,
-0.02000751532614231,
-0.06607576459646225,
0.0530143640935421,
0.042103126645088196,
0.12097804248332977,
-0.01290249265730381,
0.10676456987857819,
-0.012673799879848957,
-0.04522668942809105,
-0.02413334511220455,
-0.03341468796133995,
0.04714534059166908,
-0.050264693796634674,
-0.060616202652454376,
-4.3172136088287516e-8,
-0.025042081251740456,
0.025683511048555374,
-0.00839282013475895,
0.03843245655298233,
0.08720747381448746,
-0.1224617287516594,
-0.03185891732573509,
0.06142839789390564,
-0.01823936216533184,
-0.036636848002672195,
0.09878695756196976,
-0.02617025189101696,
-0.01164504699409008,
-0.04145857319235802,
-0.014891370199620724,
0.07533365488052368,
0.07606428861618042,
0.0686350017786026,
-0.011767472140491009,
-0.014227185398340225,
-0.032330308109521866,
-0.009999864734709263,
0.008453041315078735,
0.015397768467664719,
0.014783471822738647,
-0.05987453833222389,
0.09254775196313858,
0.12224163115024567,
0.0897940993309021,
-0.03011143021285534,
-0.006876102648675442,
0.000013839502571499906,
0.06662511825561523,
-0.02434500865638256,
0.012895544059574604,
0.06411395221948624,
0.036486443132162094,
0.0006766334408894181,
-0.038870926946401596,
0.02196446992456913,
-0.05068838968873024,
0.036572400480508804,
0.005306795705109835,
0.036751214414834976,
0.021717200055718422,
-0.04981144890189171,
-0.0000025992094379034825,
-0.029524361714720726,
-0.008756740018725395,
-0.058063581585884094,
-0.040934864431619644,
-0.010287748649716377,
-0.0008434198098257184,
0.0872388631105423,
0.09075485914945602,
0.05293438583612442,
0.012652900069952011,
0.013652756810188293,
-0.008882835507392883,
0.03783601149916649,
0.06831313669681549,
0.043205875903367996,
-0.08336078375577927,
0.0541570819914341
] |
SegLLM: Multi-round Reasoning Segmentation with Large Language Models | https://openreview.net/forum?id=Pm1NXHgzyf | [
"XuDong Wang",
"Shaolun Zhang",
"Shufan Li",
"Kehan Li",
"Konstantinos Kallidromitis",
"Yusuke Kato",
"Kazuki Kozuka",
"Trevor Darrell"
] | Poster | We present SegLLM, a novel multi-round interactive reasoning segmentation model that enhances LLM-based segmentation by exploiting conversational memory of both visual and textual outputs. By leveraging a mask-aware multimodal LLM, SegLLM re-integrates previous segmentation results into its input stream, enabling it to reason about complex user intentions and segment objects in relation to previously identified entities, including positional, interactional, and hierarchical relationships, across multiple interactions. This capability allows SegLLM to respond to visual and text queries in a chat-like manner. Evaluated on the newly curated MRSeg benchmark, SegLLM outperforms existing methods in multi- round interactive reasoning segmentation by over 20%. Additionally, we observed that training on multi-round reasoning segmentation data enhances performance on standard single-round referring segmentation and localization tasks, resulting in a 5.5% increase in cIoU for referring expression segmentation and a 4.5% improvement in [email protected] for referring expression localization. | LLMs, Reasoning Segmentation, Muiti-round Conversations | null | 1,230 | null | [
0.008182203397154808,
-0.0632166936993599,
0.0430150032043457,
0.06749319285154343,
0.04761804640293121,
-0.08482816070318222,
0.034940436482429504,
0.084632508456707,
0.060119468718767166,
-0.0886145681142807,
-0.1309754103422165,
-0.09715951234102249,
-0.003418441629037261,
0.06870688498020172,
0.06043335050344467,
0.034607917070388794,
0.04404209926724434,
0.008587528020143509,
-0.07239679992198944,
-0.05758989602327347,
0.025226300582289696,
0.010208837687969208,
0.10650753974914551,
-0.013298572041094303,
0.0030064887832850218,
-0.027414951473474503,
-0.019753552973270416,
-0.03124205395579338,
0.04268456622958183,
0.0345587395131588,
0.032388538122177124,
0.09332513064146042,
0.016080252826213837,
0.1019759476184845,
-0.041426654905080795,
0.008146109990775585,
-0.015759652480483055,
0.007447091396898031,
-0.016582785174250603,
-0.05005383491516113,
-0.10451602935791016,
-0.0412011556327343,
0.044542256742715836,
-0.019620880484580994,
0.03657706826925278,
-0.04865637421607971,
-0.04336697980761528,
0.04820307344198227,
0.04956711828708649,
0.053698331117630005,
-0.08899999409914017,
-0.0439789853990078,
-0.02214997261762619,
0.09426718205213547,
-0.057046834379434586,
-0.006622850429266691,
0.03368208557367325,
0.0379657968878746,
0.06900563836097717,
0.023103106766939163,
-0.07802382856607437,
-0.15607723593711853,
0.0074730743654072285,
0.04148408770561218,
-0.14468510448932648,
-0.016030525788664818,
0.03367841616272926,
-0.07426489144563675,
-0.06924765557050705,
0.0033585878554731607,
-0.03164926543831825,
0.022828685119748116,
-0.04235944524407387,
0.022129198536276817,
-0.03355635330080986,
0.07827505469322205,
-0.03551990166306496,
-0.08318929374217987,
0.04084980860352516,
-0.07410319894552231,
0.03126877173781395,
-0.021807294338941574,
0.021944677457213402,
-0.01799759268760681,
0.09662330150604248,
-0.00843382440507412,
0.0023133857175707817,
0.03345918655395508,
-0.024502404034137726,
0.02941768802702427,
-0.005024031735956669,
-0.048756193369627,
0.06313531845808029,
0.028180867433547974,
0.05546378344297409,
0.04482090845704079,
0.02524918131530285,
-0.05544121935963631,
-0.025396209210157394,
0.06644751131534576,
-0.02673276700079441,
0.10164109617471695,
0.004678914323449135,
-0.058261286467313766,
-0.002516060369089246,
0.011654270812869072,
0.05717865750193596,
-0.0004616859368979931,
0.06192750856280327,
-0.07945412397384644,
-0.019665967673063278,
0.02698773518204689,
0.09871523827314377,
0.04660610482096672,
0.023843297734856606,
-0.0712866559624672,
0.045997146517038345,
-0.05483509227633476,
0.08127383142709732,
0.011422088369727135,
0.02763422764837742,
0.004466574173420668,
-0.04912569001317024,
0.047984421253204346,
0.08192035555839539,
-0.02226116694509983,
-0.015461131930351257,
7.089337266201172e-36,
0.09190893173217773,
-0.004130392801016569,
0.08233083039522171,
0.03942469134926796,
0.003954928833991289,
0.011468864046037197,
-0.02344358153641224,
0.07182829082012177,
-0.051641348749399185,
-0.049117378890514374,
-0.0043693711049854755,
-0.04251857101917267,
-0.018979379907250404,
0.07887230068445206,
-0.00013548358401749283,
-0.02366224117577076,
-0.020104480907320976,
0.013051481917500496,
-0.07872051000595093,
-0.030015351250767708,
-0.04591367766261101,
-0.03159462660551071,
-0.0003817472024820745,
0.08274403214454651,
0.09401371330022812,
0.06434943526983261,
0.06718750298023224,
-0.056802310049533844,
-0.012188740074634552,
0.00557836564257741,
-0.1195371076464653,
-0.05794632434844971,
-0.007273595314472914,
0.019735759124159813,
0.04378291592001915,
-0.024892346933484077,
-0.03265520930290222,
-0.030134953558444977,
0.020426921546459198,
0.030245060101151466,
-0.01031634770333767,
-0.037307169288396835,
0.04715762287378311,
0.0011404526885598898,
-0.11656488478183746,
-0.016098475083708763,
-0.06699596345424652,
0.05589393898844719,
-0.06438939273357391,
0.021834049373865128,
0.007769227027893066,
0.10093469172716141,
-0.03004423901438713,
-0.03227381408214569,
0.0017497409135103226,
-0.08021464198827744,
-0.0006221860530786216,
-0.007395461201667786,
0.01769886165857315,
-0.040112029761075974,
-0.03868640959262848,
0.06816169619560242,
-0.05576098710298538,
0.048141904175281525,
0.02513294853270054,
0.02358851209282875,
-0.06098778918385506,
0.0822509303689003,
0.057635169476270676,
0.012521508149802685,
-0.015044993720948696,
0.047198425978422165,
0.014181959442794323,
0.027127576991915703,
-0.020825695246458054,
0.025474315509200096,
-0.013593553565442562,
-0.035004448145627975,
0.059182826429605484,
0.009704379364848137,
0.0494735948741436,
-0.006366070359945297,
-0.057048141956329346,
-0.07411982864141464,
-0.01575864851474762,
-0.06291910260915756,
0.07039349526166916,
-0.13001632690429688,
-0.04067020118236542,
-0.015576227568089962,
-0.02482214756309986,
-0.06580948829650879,
-0.026345618069171906,
0.0010810227831825614,
-0.00462436955422163,
2.737970625484888e-34,
0.045544207096099854,
-0.000599839782807976,
-0.057027678936719894,
0.05336267873644829,
-0.04310684651136398,
-0.04128332808613777,
0.03286563977599144,
0.016683237627148628,
0.01715569570660591,
-0.07397279888391495,
-0.014596241526305676,
0.0378510020673275,
0.014907622709870338,
-0.04664372280240059,
0.06728379428386688,
-0.015509407967329025,
0.031194765120744705,
-0.0026810832787305117,
-0.036063045263290405,
0.07428599148988724,
0.020533651113510132,
0.007444396149367094,
-0.07035671174526215,
-0.02589591033756733,
-0.07579020410776138,
0.057813435792922974,
0.00546995596960187,
-0.02058582939207554,
-0.030262060463428497,
0.05617844685912132,
-0.01863032765686512,
-0.02191990241408348,
0.016202380880713463,
-0.06293985992670059,
-0.06945818662643433,
0.014253262430429459,
0.02975987270474434,
-0.054017022252082825,
-0.010861326940357685,
0.023006660863757133,
0.10919306427240372,
-0.04590923711657524,
-0.020099595189094543,
0.04309723898768425,
-0.030824368819594383,
0.0005560280405916274,
-0.08865579962730408,
0.05822058767080307,
-0.09430364519357681,
0.005414076615124941,
-0.08674383908510208,
0.009647502563893795,
-0.05872124806046486,
-0.04734928533434868,
-0.00324195041321218,
-0.03693477809429169,
-0.05935677886009216,
-0.049324143677949905,
-0.032337818294763565,
-0.06116326525807381,
-0.0726345106959343,
0.061672892421483994,
0.045088548213243484,
-0.04401341825723648,
0.03440955653786659,
-0.01620185188949108,
-0.05302352085709572,
-0.03119179606437683,
0.0002759826020337641,
-0.023283278569579124,
0.06955033540725708,
-0.05771070718765259,
-0.028941895812749863,
-0.006946783512830734,
0.05088474601507187,
0.03866805508732796,
-0.023799661546945572,
-0.026276836171746254,
-0.03283607214689255,
-0.027536820620298386,
0.03143904730677605,
-0.08219005912542343,
0.062227681279182434,
0.10073462873697281,
0.06140679121017456,
0.033224109560251236,
-0.01725848764181137,
0.14822255074977875,
0.02316027134656906,
0.08319694548845291,
-0.021630246192216873,
0.014035362750291824,
0.05112549662590027,
0.09413282573223114,
-0.011355063877999783,
-3.516940338954555e-8,
-0.04143389314413071,
0.031289275735616684,
0.005426367279142141,
0.032203309237957,
0.08736366778612137,
-0.05161106213927269,
-0.06062483415007591,
0.10873993486166,
-0.020816540345549583,
0.029787003993988037,
0.005476112011820078,
-0.05447453260421753,
-0.04010090231895447,
-0.05392651632428169,
0.008467152714729309,
0.0761077031493187,
0.03517839312553406,
-0.027117077261209488,
0.006954491138458252,
0.02218974381685257,
0.055442068725824356,
-0.015600020065903664,
-0.04938792064785957,
0.04539403319358826,
0.042323846369981766,
-0.040869589895009995,
0.0027893821243196726,
0.14864450693130493,
0.028766566887497902,
-0.046226125210523605,
-0.02984420582652092,
0.04056665301322937,
0.013070523738861084,
0.023727484047412872,
0.06787198781967163,
-0.05328531563282013,
0.0020241239108145237,
0.0009132702834904194,
0.007921389304101467,
-0.03636186942458153,
0.0751265212893486,
0.0023123077116906643,
0.0134687265381217,
0.024277962744235992,
0.014041892252862453,
0.08284514397382736,
-0.029210615903139114,
-0.06924537569284439,
0.011436342261731625,
-0.033075738698244095,
-0.03271506726741791,
-0.05727803707122803,
0.025222763419151306,
0.11322293430566788,
0.017086239531636238,
-0.0016978827770799398,
-0.01969192922115326,
0.028948649764060974,
0.039532583206892014,
0.051805317401885986,
0.0010367026552557945,
0.042163364589214325,
-0.07508531957864761,
0.024977907538414
] |
Out-of-distribution Generalization for Total Variation based Invariant Risk Minimization | https://openreview.net/forum?id=c4wEKJOjY3 | [
"Yuanchao Wang",
"Zhao-Rong Lai",
"Tianqi Zhong"
] | Poster | Invariant risk minimization is an important general machine learning framework that has recently been interpreted as a total variation model (IRM-TV). However, how to improve out-of-distribution (OOD) generalization in the IRM-TV setting remains unsolved. In this paper, we extend IRM-TV to a Lagrangian multiplier model named OOD-TV-IRM. We find that the autonomous TV penalty hyperparameter is exactly the Lagrangian multiplier. Thus OOD-TV-IRM is essentially a primal-dual optimization model, where the primal optimization minimizes the entire invariant risk and the dual optimization strengthens the TV penalty. The objective is to reach a semi-Nash equilibrium where the balance between the training loss and OOD generalization is maintained. We also develop a convergent primal-dual algorithm that facilitates an adversarial learning scheme. Experimental results show that OOD-TV-IRM outperforms IRM-TV in most situations. | Out-of-distribution generalization, total variation, invariant risk minimization, primal-dual optimization | We propose an out-of-distribution generalization methodology for the total variation based invariant risk minimization. | 1,225 | 2502.19665 | [
-0.004929953254759312,
-0.0898422971367836,
0.001550947898067534,
0.005288759712129831,
0.07878158241510391,
0.04407718777656555,
-0.0007921197102405131,
0.010115784592926502,
-0.010249197483062744,
-0.020353853702545166,
-0.039183780550956726,
0.009760524146258831,
-0.01760772056877613,
0.03433618322014809,
-0.021897131577134132,
-0.02329918183386326,
0.09262744337320328,
0.1212485283613205,
-0.0855545699596405,
-0.005959133617579937,
0.015160007402300835,
0.015805143862962723,
-0.010316253639757633,
0.0007754776743240654,
0.07470359653234482,
-0.07407328486442566,
0.009186380542814732,
0.04127683490514755,
-0.01991226337850094,
-0.0682961568236351,
0.11532256752252579,
-0.01976427435874939,
-0.011702547781169415,
-0.03699931874871254,
-0.0386323481798172,
0.03146486356854439,
-0.1094367578625679,
0.020068639889359474,
0.024073785170912743,
0.01332278735935688,
0.02794129215180874,
-0.0028784789610654116,
-0.0620911531150341,
-0.022354111075401306,
0.0362408384680748,
-0.007803182117640972,
-0.0017420112853869796,
-0.11734632402658463,
0.05818856507539749,
0.04793282970786095,
0.06987126171588898,
0.016297070309519768,
-0.10153229534626007,
0.06674724072217941,
-0.039401885122060776,
-0.0880679041147232,
0.025694655254483223,
-0.014288917183876038,
-0.01440558023750782,
-0.02024674415588379,
0.011939303949475288,
-0.11499393731355667,
-0.023907586932182312,
0.03472228720784187,
-0.006006974261254072,
-0.06468387693166733,
0.09490787982940674,
-0.004467118065804243,
-0.0022157339844852686,
-0.016140911728143692,
-0.038059573620557785,
0.09199276566505432,
-0.02986082062125206,
-0.005837502423673868,
0.02368619106709957,
0.022076988592743874,
0.06621753424406052,
-0.015436697751283646,
0.051222462207078934,
-0.023410875350236893,
0.05597373843193054,
0.041569534689188004,
0.012321430258452892,
0.0036198492161929607,
0.11346086114645004,
-0.01613772101700306,
-0.06421962380409241,
-0.019487418234348297,
0.0999678298830986,
-0.01536780409514904,
-0.04994329437613487,
0.03654726594686508,
0.01526182796806097,
0.03486618772149086,
0.013739952817559242,
-0.01846388354897499,
-0.024704935029149055,
-0.07854672521352768,
-0.018524089828133583,
0.1366579234600067,
-0.06740102916955948,
-0.029312822967767715,
0.03165242448449135,
-0.022594153881072998,
0.11226251721382141,
-0.1202084943652153,
0.03526165336370468,
0.0035490491427481174,
0.057036109268665314,
-0.0668352022767067,
0.09541517496109009,
0.1066737174987793,
0.06399830430746078,
-0.012530115433037281,
-0.07273966819047928,
0.020802658051252365,
-0.00700728315860033,
-0.005919910501688719,
0.04572869837284088,
0.021769940853118896,
-0.06179747357964516,
-0.056980833411216736,
0.00551277631893754,
0.05882807448506355,
0.022566232830286026,
0.007468115538358688,
-0.049733832478523254,
3.546959062032233e-33,
0.011556049808859825,
0.000557640683837235,
0.054338522255420685,
-0.10919435322284698,
0.03468775376677513,
0.007648143917322159,
0.0362054668366909,
0.03223058581352234,
0.0010762587189674377,
-0.002896781312301755,
0.024420419707894325,
-0.04671614244580269,
-0.08761966973543167,
0.07429538667201996,
0.049359217286109924,
0.013595117256045341,
-0.05887027084827423,
-0.0015932644018903375,
0.03494950756430626,
-0.046894319355487823,
0.06601377576589584,
-0.04576800391077995,
-0.04286424443125725,
0.0027820745017379522,
-0.04279366880655289,
-0.0027636270970106125,
0.06928226351737976,
-0.03884444385766983,
-0.05466080084443092,
0.047465547919273376,
-0.09118824452161789,
0.0027560258749872446,
-0.018627554178237915,
0.06494516879320145,
0.08157345652580261,
-0.031309809535741806,
-0.11146324872970581,
0.006556128151714802,
0.0017373798182234168,
0.015611231327056885,
-0.07830420136451721,
0.025708355009555817,
0.025743061676621437,
-0.08822901546955109,
-0.1027718037366867,
-0.06050163134932518,
0.03233116865158081,
0.07253235578536987,
-0.07680042833089828,
-0.09805804491043091,
-0.049500059336423874,
0.013866771012544632,
-0.016147997230291367,
-0.04800880700349808,
-0.02992723323404789,
-0.01162303239107132,
-0.043886613100767136,
-0.008600566536188126,
0.001542513258755207,
-0.003926288336515427,
-0.009675728157162666,
0.024080421775579453,
0.07384930551052094,
0.051948998123407364,
-0.02007964253425598,
0.034709349274635315,
0.041926220059394836,
0.031199146062135696,
0.021252257749438286,
-0.009754191152751446,
-0.012599176727235317,
0.026338160037994385,
0.040391452610492706,
-0.10440093278884888,
0.02770925685763359,
0.023774079978466034,
0.11997934430837631,
-0.043491292744874954,
0.02463357150554657,
-0.01669546216726303,
-0.014732529409229755,
0.11918468773365021,
0.02157013863325119,
-0.013490221463143826,
-0.030511008575558662,
0.0007747145718894899,
0.028968820348381996,
-0.03999049961566925,
-0.02687925286591053,
-0.06405232846736908,
-0.05415203049778938,
-0.0031865318305790424,
-0.03905150666832924,
-0.031656742095947266,
-0.0016804448096081614,
-2.743374777091801e-33,
-0.07992462813854218,
0.05697835981845856,
-0.06588850915431976,
0.011090564541518688,
0.01304563321173191,
-0.05336126312613487,
-0.04867798462510109,
0.008280438371002674,
-0.07908943295478821,
0.08499602973461151,
-0.05174906924366951,
0.009730852209031582,
0.015206638723611832,
0.014340045861899853,
0.00014182768063619733,
-0.0041381786577403545,
-0.04285019263625145,
0.07155193388462067,
0.021113460883498192,
-0.01442953385412693,
0.05151614174246788,
0.07459542900323868,
-0.022035034373402596,
0.014589931815862656,
0.033449284732341766,
-0.025942254811525345,
-0.10117144882678986,
0.12839031219482422,
-0.00043753033969551325,
0.05131972208619118,
-0.053001195192337036,
0.03940564766526222,
-0.020677249878644943,
-0.037211935967206955,
-0.07707580178976059,
0.07847589999437332,
0.014218413271009922,
0.015745915472507477,
-0.02021968737244606,
0.19182367622852325,
-0.014003626070916653,
-0.034263040870428085,
-0.024784747511148453,
0.007146455347537994,
0.05823845788836479,
0.018157057464122772,
0.06619168072938919,
-0.11064407229423523,
0.066836878657341,
-0.022861959412693977,
-0.018758337944746017,
0.06332051753997803,
-0.05659596994519234,
0.06901654601097107,
-0.05929320678114891,
-0.09856577217578888,
0.03023678809404373,
-0.012148650363087654,
0.045180484652519226,
-0.03347286954522133,
-0.025560647249221802,
-0.022839343175292015,
-0.026611490175127983,
0.02226417139172554,
-0.021441863849759102,
0.013742022216320038,
0.010296110063791275,
-0.048804305493831635,
-0.017562327906489372,
0.017201615497469902,
0.03877360001206398,
0.011180018074810505,
0.024054333567619324,
-0.009220944717526436,
-0.009525791741907597,
0.03172385320067406,
0.03186086192727089,
0.004107336979359388,
-0.024282395839691162,
-0.030483052134513855,
0.07885326445102692,
-0.08386741578578949,
0.06313896924257278,
0.03461887314915657,
0.003315589390695095,
0.006173755507916212,
0.02936408855021,
0.05614515766501427,
-0.013061536476016045,
-0.014118378981947899,
-0.0587814524769783,
0.053831152617931366,
0.030190816149115562,
0.029338883236050606,
-0.028062518686056137,
-4.544714116150317e-8,
-0.03954843431711197,
0.0109397666528821,
0.013288949616253376,
-0.022580036893486977,
0.06943876296281815,
-0.0073388186283409595,
-0.03343311324715614,
0.05214435234665871,
-0.01765584386885166,
-0.06962726265192032,
0.058182280510663986,
0.009387937374413013,
0.015599522739648819,
-0.04618644714355469,
-0.039894070476293564,
0.06751131266355515,
-0.043160080909729004,
0.05615187808871269,
-0.013815417885780334,
0.07180430740118027,
0.013437020592391491,
0.01786239817738533,
0.019992727786302567,
-0.06729675829410553,
0.09159082919359207,
-0.03704254701733589,
0.03351559117436409,
0.02493060939013958,
0.12125060707330704,
-0.0019828984513878822,
-0.05087623372673988,
0.003173431381583214,
-0.011207264848053455,
-0.03256051614880562,
-0.08726181089878082,
0.12405308336019516,
-0.008779370225965977,
-0.004611663054674864,
-0.012456297874450684,
0.001989610493183136,
-0.03015768714249134,
0.01927727833390236,
-0.005792406387627125,
0.04916556552052498,
0.024352816864848137,
0.025705408304929733,
0.05058225616812706,
-0.12984347343444824,
0.019726112484931946,
-0.01631072908639908,
0.06743685156106949,
-0.0003908713988494128,
0.08661521971225739,
0.07115478068590164,
0.059275757521390915,
-0.0640871450304985,
0.01149069145321846,
-0.0739278569817543,
0.024637291207909584,
0.035039909183979034,
0.011824075132608414,
-0.003977831453084946,
-0.10612044483423233,
-0.01192336156964302
] |
Multi-Reward as Condition for Instruction-based Image Editing | https://openreview.net/forum?id=9RFocgIccP | [
"Xin Gu",
"Ming Li",
"Libo Zhang",
"Fan Chen",
"Longyin Wen",
"Tiejian Luo",
"Sijie Zhu"
] | Poster | High-quality training triplets (instruction, original image, edited image) are essential for instruction-based image editing. Predominant training datasets (e.g., InsPix2Pix) are created using text-to-image generative models (e.g., Stable Diffusion, DALL-E) which are not trained for image editing. Accordingly, these datasets suffer from inaccurate instruction following, poor detail preserving, and generation artifacts. In this paper, we propose to address the training data quality issue with multi-perspective reward data instead of refining the ground-truth image quality. 1) we first design a quantitative metric system based on best-in-class LVLM (Large Vision Language Model), i.e., GPT-4o in our case, to evaluate the generation quality from 3 perspectives, namely, instruction following, detail preserving, and generation quality. For each perspective, we collected quantitative score in $0\sim 5$ and text descriptive feedback on the specific failure points in ground-truth edited images, resulting in a high-quality editing reward dataset, i.e., RewardEdit20K. 2) We further proposed a novel training framework to seamlessly integrate the metric output, regarded as multi-reward, into editing models to learn from the imperfect training triplets. During training, the reward scores and text descriptions are encoded as embeddings and fed into both the latent space and the U-Net of the editing models as auxiliary conditions. During inference, we set these additional conditions to the highest score with no text description for failure points, to aim at the best generation outcome. 3) We also build a challenging evaluation benchmark with real-world images/photos and diverse editing instructions, named as Real-Edit. Experiments indicate that our multi-reward conditioned model outperforms its no-reward counterpart on two popular editing pipelines, i.e., InsPix2Pix and SmartEdit. Code is released at https://github.com/bytedance/Multi-Reward-Editing. | Instruction-based Image Editing | null | 1,218 | 2411.04713 | [
-0.012043135240674019,
-0.05422678217291832,
0.04270841181278229,
-0.02187427133321762,
-0.0018552152905613184,
-0.01726454682648182,
-0.07228775322437286,
-0.03424285352230072,
0.07037512958049774,
0.007729196920990944,
0.0059200469404459,
-0.032549288123846054,
0.08561751246452332,
0.021236853674054146,
-0.04538879171013832,
-0.013086195103824139,
0.06110963225364685,
0.055348627269268036,
-0.1500927060842514,
-0.09510693699121475,
0.07967400550842285,
0.03824272006750107,
0.07442111521959305,
-0.0014333706349134445,
0.026920123025774956,
0.014803115278482437,
0.020908530801534653,
-0.04824428632855415,
0.05805712193250656,
-0.08596225082874298,
-0.02622135914862156,
0.03489604592323303,
0.04073987528681755,
0.0021667787805199623,
-0.06378477066755295,
0.12543697655200958,
-0.009064744226634502,
0.012192429043352604,
0.018832780420780182,
-0.01599762961268425,
-0.03395799174904823,
0.042962390929460526,
0.03856472298502922,
-0.006101545877754688,
0.09158232808113098,
0.030428435653448105,
-0.032311104238033295,
-0.0999663770198822,
-0.03607712313532829,
0.03223340958356857,
-0.03160380944609642,
-0.02766358107328415,
-0.05530920997262001,
-0.0037482166662812233,
-0.055156394839286804,
0.0586823932826519,
0.04502476006746292,
0.031230805441737175,
0.014145336113870144,
-0.03784867376089096,
-0.00942478608340025,
-0.012987220659852028,
-0.04436073824763298,
-0.02484610117971897,
-0.0015437083784490824,
-0.04401914402842522,
0.0488409586250782,
-0.028773754835128784,
-0.04249443858861923,
-0.028362523764371872,
-0.0286125298589468,
-0.0019459897885099053,
-0.053681399673223495,
0.014567586593329906,
-0.09282643347978592,
0.09432993829250336,
0.023667048662900925,
-0.045275866985321045,
-0.0060361106880009174,
-0.045653294771909714,
0.08629238605499268,
-0.013901464641094208,
0.06326599419116974,
-0.006241057999432087,
0.03291303291916847,
-0.0467584989964962,
0.07399431616067886,
0.06002071499824524,
0.03449283167719841,
0.02517259679734707,
-0.07067319750785828,
0.03525257110595703,
-0.004334558732807636,
-0.0009076530113816261,
0.019892815500497818,
0.028224173933267593,
-0.021805385127663612,
-0.11740688979625702,
-0.03128582611680031,
0.060546375811100006,
0.05220990628004074,
-0.00016395971761085093,
-0.06928575783967972,
-0.025444185361266136,
0.1147928535938263,
-0.009374809451401234,
0.09024236351251602,
0.03663500025868416,
0.01796714961528778,
-0.08646661788225174,
0.06262469291687012,
0.01990610733628273,
-0.0373469702899456,
-0.01970590651035309,
0.020415252074599266,
0.0021612998098134995,
0.013551644049584866,
0.001607642974704504,
-0.014107158407568932,
-0.004216330591589212,
-0.019004713743925095,
-0.0494135245680809,
0.03466468304395676,
-0.003521355800330639,
-0.030122606083750725,
-0.15176445245742798,
-0.0027773345354944468,
-8.964623894863047e-34,
0.056036558002233505,
0.05651035159826279,
0.04224018007516861,
-0.006634038407355547,
0.09549695998430252,
0.04148688167333603,
-0.03235290199518204,
0.0007669520564377308,
-0.06686842441558838,
-0.04283858463168144,
0.003078897250816226,
-0.03211849927902222,
-0.0771288052201271,
0.1618148535490036,
0.10644173622131348,
0.038058798760175705,
-0.04629794880747795,
0.01504990179091692,
-0.020571326836943626,
0.05261674523353577,
0.0544474720954895,
-0.037469107657670975,
0.024753829464316368,
-0.02299344353377819,
-0.01274392381310463,
0.03947058692574501,
0.06784117221832275,
-0.04301079735159874,
-0.05472268536686897,
-0.051495879888534546,
-0.07385415583848953,
-0.03663866966962814,
0.1064949780702591,
0.030086234211921692,
-0.019823523238301277,
-0.03462949022650719,
-0.008019690401852131,
-0.014842326752841473,
0.039253734052181244,
-0.024183005094528198,
-0.0076416307128965855,
0.08787372708320618,
-0.007129733916372061,
-0.028665535151958466,
-0.051261745393276215,
0.046296779066324234,
0.05504964664578438,
0.028311949223279953,
-0.08120699971914291,
-0.0391874685883522,
0.031059319153428078,
0.022715449333190918,
0.01724393106997013,
-0.04359671473503113,
-0.08777531236410141,
0.021666759625077248,
0.08419660478830338,
0.044206056743860245,
0.039911381900310516,
0.020544668659567833,
0.053580258041620255,
0.07783167064189911,
-0.015705345198512077,
0.0696190893650055,
-0.015058926306664944,
0.03954268991947174,
0.025917425751686096,
0.009750367142260075,
0.021318841725587845,
-0.013415439054369926,
-0.059998173266649246,
-0.009930333122611046,
-0.005617637652903795,
-0.07191722840070724,
0.06942056119441986,
-0.03336697071790695,
0.013796022161841393,
-0.013595941476523876,
0.03387141600251198,
0.0375993475317955,
-0.03394316881895065,
-0.02566242590546608,
-0.0020575455855578184,
-0.14991992712020874,
-0.02572396956384182,
-0.023023998364806175,
-0.0011237449944019318,
-0.025066496804356575,
-0.01991424523293972,
0.0022235033102333546,
-0.03495827689766884,
-0.012584935873746872,
0.008378581143915653,
-0.011530385352671146,
0.01653108559548855,
1.5356309224256964e-33,
0.0037359660491347313,
0.0861719474196434,
0.012140515260398388,
0.11872950196266174,
-0.03895758092403412,
-0.03517364338040352,
0.004349255468696356,
0.07228169590234756,
0.033188898116350174,
-0.05884493514895439,
0.004261239431798458,
0.015086316503584385,
-0.010430139489471912,
-0.035217687487602234,
-0.04846575856208801,
-0.08734411001205444,
-0.014377759769558907,
-0.02810685895383358,
-0.07165764272212982,
0.02481027878820896,
0.09161744266748428,
0.09741362929344177,
-0.06287255138158798,
0.043714165687561035,
-0.014007851481437683,
0.07740835100412369,
-0.017386529594659805,
0.049603499472141266,
0.00653372285887599,
0.009086226113140583,
0.03223598375916481,
-0.06153739616274834,
0.04223859682679176,
-0.016021963208913803,
-0.056679755449295044,
-0.036271609365940094,
0.11026369035243988,
-0.01733255386352539,
-0.1116611585021019,
0.15941181778907776,
0.02751743420958519,
-0.044697392731904984,
-0.08191736042499542,
0.05503709986805916,
-0.00902738980948925,
-0.03440733253955841,
-0.07221540808677673,
-0.02983466163277626,
0.03898550942540169,
0.043696630746126175,
-0.048377107828855515,
0.023679811507463455,
-0.09518952667713165,
0.1160004660487175,
0.044994112104177475,
-0.13802483677864075,
-0.11474594473838806,
-0.029944417998194695,
0.06059988588094711,
-0.030352579429745674,
-0.06681213527917862,
-0.03021170198917389,
-0.01441210601478815,
0.007438535802066326,
0.013789728283882141,
0.018822135403752327,
-0.05231760814785957,
0.028695033863186836,
0.00043151990394108,
0.10434655845165253,
0.020216453820466995,
0.0136325154453516,
0.059276483952999115,
-0.05704566463828087,
0.026683295145630836,
-0.006470188032835722,
0.021007748320698738,
0.029793255031108856,
0.0466657392680645,
-0.0595247745513916,
0.015705656260252,
-0.027686096727848053,
0.0680491179227829,
0.08038380742073059,
0.046350765973329544,
0.02341284602880478,
0.03493079915642738,
0.03602380305528641,
0.01572270132601261,
0.05285865440964699,
-0.08653794974088669,
0.026825740933418274,
0.042913660407066345,
0.01091443095356226,
-0.04641173407435417,
-4.38546443604082e-8,
-0.019365428015589714,
0.023764215409755707,
-0.029088446870446205,
0.00788936298340559,
-0.013647319748997688,
-0.06504608690738678,
-0.05190592631697655,
0.0837668851017952,
0.047365717589855194,
-0.03921946510672569,
0.039985645562410355,
-0.0053834798745810986,
-0.11307280510663986,
-0.10960469394922256,
-0.012974203564226627,
0.0567755363881588,
-0.007629373576492071,
0.10271558165550232,
0.05925356596708298,
0.004781045485287905,
0.02152193710207939,
0.02027822844684124,
0.03191946819424629,
-0.06647361814975739,
-0.013664549216628075,
-0.024802885949611664,
0.016698794439435005,
-0.0003062599280383438,
0.03492698445916176,
0.0010252034990116954,
0.0977795422077179,
-0.05064759775996208,
0.1158360168337822,
-0.009137513116002083,
0.023708660155534744,
-0.01959078013896942,
0.02207241579890251,
0.026325900107622147,
0.005057177040725946,
-0.03424400836229324,
-0.004572620615363121,
0.08240630477666855,
-0.011236988008022308,
0.009697594679892063,
0.0315990187227726,
0.0000072487350735173095,
0.034809209406375885,
-0.09614308178424835,
0.004784237127751112,
-0.0354798398911953,
-0.015704354271292686,
0.011453182436525822,
-0.0628436878323555,
0.028218435123562813,
0.04876449331641197,
-0.024248365312814713,
0.03799007460474968,
-0.03764956071972847,
0.04974270612001419,
0.033317748457193375,
0.038313474506139755,
-0.013450877740979195,
-0.03209095448255539,
-0.03758101910352707
] |
Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations | https://openreview.net/forum?id=Hu0FSOSEyS | [
"Litu Rout",
"Yujia Chen",
"Nataniel Ruiz",
"Constantine Caramanis",
"Sanjay Shakkottai",
"Wen-Sheng Chu"
] | Poster | Generative models transform random noise into images, while their inversion aims to reconstruct structured noise for recovery and editing.
This paper addresses two key tasks: (i) *inversion* and (ii) *editing* of real images using stochastic equivalents of rectified flow models (e.g., Flux).
While Diffusion Models (DMs) dominate the field of generative modeling for images, their inversion suffers from faithfulness and editability challenges due to nonlinear drift and diffusion.
Existing DM inversion methods require costly training of additional parameters or test-time optimization of latent variables.
Rectified Flows (RFs) offer a promising alternative to DMs, yet their inversion remains underexplored.
We propose RF inversion using dynamic optimal control derived via a linear quadratic regulator, and prove that the resulting vector field is equivalent to a rectified stochastic differential equation.
We further extend our framework to design a stochastic sampler for Flux.
Our method achieves state-of-the-art performance in zero-shot inversion and editing, surpassing prior works in stroke-to-image synthesis and semantic image editing, with large-scale human evaluations confirming user preference.
See our project page https://rf-inversion.github.io/ for code and demo. | Inverse Problems, Generative Modeling, Diffusion Models, Rectified Flows, Posterior Sampling, Optimal Control | We present an efficient inversion method for rectified flow models, including Flux, that requires no additional parameter training, latent variable optimization, prompt tuning, or complex attention processors. | 1,202 | 2410.10792 | [
-0.05012395232915878,
-0.07549096643924713,
0.09271751344203949,
0.07117832452058792,
0.032141562551259995,
-0.07304563373327255,
-0.07403251528739929,
-0.11746329069137573,
0.08045630156993866,
0.02968817576766014,
-0.015113106928765774,
-0.0001774036791175604,
0.012416098266839981,
-0.0015593954594805837,
-0.03547120839357376,
0.02955424226820469,
0.0486438162624836,
0.12790055572986603,
-0.042783502489328384,
-0.0315430648624897,
-0.04766124114394188,
0.029934698715806007,
0.022977307438850403,
-0.06004172936081886,
0.06845197826623917,
0.009240640327334404,
0.018594468012452126,
-0.033091817051172256,
-0.03524461016058922,
-0.05629236623644829,
0.0830952599644661,
0.012841244228184223,
-0.016524415463209152,
-0.004701949190348387,
-0.04941193014383316,
0.07999936491250992,
-0.011886674910783768,
0.029511982575058937,
-0.009043700993061066,
0.03142886236310005,
0.018061859533190727,
0.011578836478292942,
-0.01497146300971508,
0.0152811324223876,
0.06750353425741196,
-0.0336252897977829,
-0.013782422058284283,
-0.09207519888877869,
0.05234738066792488,
0.09509477019309998,
0.0549636110663414,
0.016889240592718124,
-0.11129052937030792,
0.02517082914710045,
0.037320852279663086,
0.05887069180607796,
0.10184337198734283,
-0.05433114990592003,
-0.004404481500387192,
-0.027623265981674194,
-0.00830653216689825,
-0.06285729259252548,
0.012435073964297771,
-0.022497767582535744,
-0.04039785638451576,
-0.03337467089295387,
0.0023269555531442165,
0.0086438013240695,
0.001518220524303615,
-0.00985865481197834,
-0.034840814769268036,
0.030484426766633987,
-0.05811256915330887,
-0.03276529908180237,
-0.030035734176635742,
-0.054609641432762146,
0.03886938840150833,
-0.03764421120285988,
0.04964649677276611,
-0.05852382257580757,
0.07984206825494766,
-0.005274015013128519,
0.0659557655453682,
-0.020040569826960564,
0.05097617954015732,
0.041331175714731216,
0.016894202679395676,
0.06672403961420059,
0.058158598840236664,
0.033638011664152145,
-0.08121112734079361,
0.005700272973626852,
-0.08408608287572861,
-0.02300151437520981,
0.012194252572953701,
-0.005360994953662157,
0.04162119701504707,
-0.05822741240262985,
0.03447018563747406,
0.07643367350101471,
0.064637191593647,
-0.015002358704805374,
-0.05326107516884804,
0.017799856141209602,
0.00517160864546895,
-0.06284166127443314,
0.09558742493391037,
0.042589470744132996,
-0.025601301342248917,
-0.03343389183282852,
0.0781385600566864,
-0.017766065895557404,
0.05462035536766052,
0.010410313494503498,
0.026820290833711624,
-0.05479689687490463,
0.03027583472430706,
-0.03962942585349083,
-0.061031416058540344,
0.017797600477933884,
-0.04182972013950348,
0.04193595051765442,
-0.08833722770214081,
-0.010576188564300537,
0.03309324383735657,
-0.056914445012807846,
0.0380743071436882,
2.050932748337647e-33,
0.0134958541020751,
0.02914639376103878,
0.05515281856060028,
0.04412542283535004,
0.07818111777305603,
0.028556915000081062,
0.009505491703748703,
-0.047085076570510864,
-0.03364144638180733,
-0.010955065488815308,
0.007025386672466993,
-0.07739940285682678,
-0.11683760583400726,
0.06657737493515015,
0.029489310458302498,
-0.0557599775493145,
-0.05420803278684616,
0.047304704785346985,
0.00958321150392294,
0.019322790205478668,
0.055993273854255676,
-0.033733658492565155,
-0.10241632163524628,
-0.039987962692976,
-0.021992672234773636,
-0.018487006425857544,
0.022736720740795135,
-0.0002729404077399522,
0.0647880882024765,
-0.008745933882892132,
-0.1652372181415558,
-0.030597886070609093,
0.022874366492033005,
0.023257482796907425,
-0.054924268275499344,
-0.03775227069854736,
0.04214296117424965,
-0.03841853514313698,
-0.035345155745744705,
-0.06922636926174164,
0.008873989805579185,
0.06486009806394577,
-0.0811290517449379,
-0.03925475478172302,
-0.12276042252779007,
-0.010819447226822376,
0.07037446647882462,
0.024088528007268906,
0.013856318779289722,
-0.035945404320955276,
0.049832992255687714,
0.03814121335744858,
-0.010147890076041222,
-0.04527999088168144,
-0.026826558634638786,
0.02622642181813717,
0.0890158861875534,
-0.05922101065516472,
0.008296506479382515,
0.040646880865097046,
0.0297126155346632,
0.09821612387895584,
-0.03572709485888481,
0.041841018944978714,
0.06711748987436295,
0.06459473818540573,
0.04679730162024498,
0.06851927191019058,
0.020953277125954628,
-0.02653832919895649,
-0.04555400088429451,
0.006956356577575207,
-0.06961532682180405,
-0.05539918318390846,
0.025771893560886383,
-0.0208191629499197,
0.059068936854600906,
0.018011484295129776,
0.027183331549167633,
0.019997918978333473,
-0.08872513473033905,
-0.007120263762772083,
-0.06665866822004318,
-0.009418157860636711,
-0.006208893842995167,
-0.021865930408239365,
-0.0031906072981655598,
-0.07092530280351639,
-0.030302515253424644,
-0.03802031651139259,
-0.027358580380678177,
-0.04250491410493851,
0.011095589958131313,
-0.013452086597681046,
0.011850657872855663,
-2.0924479027308287e-33,
-0.03641688823699951,
0.032772742211818695,
-0.011256680823862553,
0.08688773959875107,
-0.02868734300136566,
-0.054228946566581726,
0.01134903822094202,
0.07873114198446274,
0.10691609978675842,
-0.10097291320562363,
-0.08239449560642242,
0.020175619050860405,
-0.0447930246591568,
-0.05284830182790756,
0.005682629998773336,
-0.03426952287554741,
-0.032574500888586044,
0.013289554975926876,
-0.07809785008430481,
0.0021740966476500034,
-0.000381767371436581,
0.14569705724716187,
0.0711575299501419,
0.030046194791793823,
-0.06336936354637146,
0.1051800325512886,
-0.03668024018406868,
0.08772066980600357,
0.02694535441696644,
-0.03372783213853836,
-0.05632481351494789,
-0.018223706632852554,
0.03114978037774563,
0.008528448641300201,
-0.05656231567263603,
0.07110341638326645,
0.024193165823817253,
-0.00007637801172677428,
-0.03195401281118393,
0.10731729865074158,
0.02658318169414997,
-0.007285945117473602,
-0.0255744606256485,
0.08435142040252686,
0.005781617481261492,
-0.010692433454096317,
-0.02738324925303459,
-0.05434328690171242,
0.08048653602600098,
0.031755633652210236,
-0.06445950269699097,
0.04303434118628502,
-0.1246817335486412,
0.04335132613778114,
-0.023610634729266167,
-0.07328314334154129,
-0.05207633227109909,
-0.09118202328681946,
0.12015552073717117,
0.021380925551056862,
-0.06054609268903732,
-0.032203614711761475,
0.0017949935281649232,
-0.04660353064537048,
-0.03950749710202217,
-0.020544329658150673,
-0.0497465506196022,
-0.014072743244469166,
0.020285898819565773,
0.06630971282720566,
0.037488386034965515,
0.006079508922994137,
0.015514974482357502,
0.03406527638435364,
0.09952903538942337,
-0.09496928751468658,
0.06153068691492081,
0.04294760525226593,
0.037236664444208145,
-0.05987368896603584,
0.025863543152809143,
0.016126032918691635,
0.016961991786956787,
0.0012691252632066607,
0.02239985205233097,
0.012198315933346748,
-0.03752287104725838,
-0.10041867941617966,
0.024998990818858147,
0.0033469574991613626,
-0.03998314216732979,
-0.020857855677604675,
0.07815851271152496,
0.03146607056260109,
-0.10649247467517853,
-3.9129655959868614e-8,
-0.08408220112323761,
0.06576741486787796,
0.07192997634410858,
-0.0047657000832259655,
0.02372981421649456,
-0.04587813839316368,
0.05405018478631973,
0.006734174210578203,
0.08843988925218582,
-0.08191957324743271,
0.087244413793087,
-0.005172558594495058,
-0.034944336861371994,
-0.07802126556634903,
-0.05818845331668854,
0.06469818204641342,
-0.02244851551949978,
0.05060305446386337,
0.024432385340332985,
-0.02163589932024479,
-0.024400772526860237,
0.0490846112370491,
-0.040555890649557114,
0.009605059400200844,
0.08816129714250565,
0.02283487655222416,
-0.00484898267313838,
-0.0053902301006019115,
0.10363318771123886,
-0.007423374801874161,
-0.0031397731509059668,
0.00528733991086483,
0.08428672701120377,
0.0832817405462265,
-0.010847237892448902,
-0.03507988527417183,
0.049046650528907776,
0.017745818942785263,
-0.04642832279205322,
-0.05008544772863388,
-0.05871459096670151,
0.03759828954935074,
-0.021863998845219612,
-0.013864217326045036,
-0.07171554118394852,
0.03464018926024437,
0.03635759651660919,
-0.050656095147132874,
0.029309282079339027,
-0.01689320243895054,
0.04272110387682915,
-0.031762104481458664,
0.02123895101249218,
0.033312443643808365,
0.0623677596449852,
-0.001534632290713489,
0.030803998932242393,
0.029672279953956604,
0.006204609293490648,
0.01974509283900261,
-0.02118133381009102,
0.057536978274583817,
-0.0021496154367923737,
-0.041884586215019226
] |
Vector-ICL: In-context Learning with Continuous Vector Representations | https://openreview.net/forum?id=xing7dDGh3 | [
"Yufan Zhuang",
"Chandan Singh",
"Liyuan Liu",
"Jingbo Shang",
"Jianfeng Gao"
] | Poster | Large language models (LLMs) have shown remarkable in-context learning (ICL) capabilities on textual data. We explore whether these capabilities can be extended to continuous vectors from diverse domains, obtained from black-box pretrained encoders. By aligning input data with an LLM's embedding space through lightweight projectors, we observe that LLMs can effectively process and learn from these projected vectors, which we term Vector-ICL. In particular, we find that pretraining projectors with general language modeling objectives enables Vector-ICL, while task-specific finetuning further enhances performance. In our experiments across various tasks and modalities, including text reconstruction, numerical function regression, text classification, summarization, molecule captioning, time-series classification, graph classification, and fMRI decoding, Vector-ICL often surpasses both few-shot ICL and domain-specific model or tuning. We further conduct analyses and case studies, indicating the potential of LLMs to process vector representations beyond traditional token-based paradigms. | large language models, in-context learning | We discover that large language models can effectively process and in-context learn from continuous representations from various domains, often outperforming regular ICL and domain-specific models across diverse tasks and modalities. | 1,196 | null | [
-0.02766658551990986,
-0.08128524571657181,
-0.002652745461091399,
0.03169742599129677,
0.07237546145915985,
0.034609828144311905,
0.006543878465890884,
0.008564083836972713,
0.08272171765565872,
-0.017490100115537643,
-0.10826246440410614,
-0.012524896301329136,
0.01931724324822426,
0.10325410962104797,
-0.0140313021838665,
-0.02821747399866581,
0.05050390213727951,
0.08180340379476547,
-0.08375168591737747,
-0.01828872784972191,
0.045448411256074905,
0.05815945565700531,
0.04878242686390877,
-0.011216500774025917,
0.026498669758439064,
0.06066764146089554,
0.029351938515901566,
-0.04778321459889412,
0.01706971600651741,
0.015210151672363281,
0.03191584721207619,
0.05829625204205513,
-0.02374330349266529,
0.0499342642724514,
-0.07086173444986343,
0.091219462454319,
-0.07173093408346176,
0.06144364923238754,
-0.007818757556378841,
-0.05954786762595177,
0.02762015350162983,
-0.023390477523207664,
-0.0028505537193268538,
0.061009883880615234,
0.07751456648111343,
0.005972752347588539,
0.02534293569624424,
-0.0888427272439003,
0.004871428944170475,
0.020629476755857468,
-0.07972347736358643,
0.013501638546586037,
-0.01730717532336712,
0.022182878106832504,
-0.015540525317192078,
0.030018050223588943,
0.015397942624986172,
-0.015956610441207886,
-0.0040373350493609905,
0.015440835617482662,
-0.05346939340233803,
-0.09002125263214111,
0.062472254037857056,
-0.005821569822728634,
-0.029983196407556534,
0.011901328340172768,
0.00006604847294511274,
0.1120627298951149,
-0.05515565350651741,
0.062202733010053635,
-0.04218703508377075,
0.05657248571515083,
-0.024270115420222282,
0.06157400831580162,
0.04632379114627838,
0.06431383639574051,
0.08473427593708038,
-0.03177105262875557,
0.13891135156154633,
-0.04978117346763611,
0.035665564239025116,
0.08793264627456665,
0.07466073334217072,
-0.034547507762908936,
0.09525632113218307,
0.07568756490945816,
0.027421560138463974,
0.04348147287964821,
-0.004630531184375286,
0.05009835585951805,
-0.0029500245582312346,
-0.08782565593719482,
0.08044691383838654,
-0.004658601712435484,
0.022374937310814857,
-0.061230096966028214,
0.04137440398335457,
-0.0314335934817791,
-0.014911530539393425,
0.06974567472934723,
0.032968681305646896,
0.02101798728108406,
-0.008135058917105198,
-0.047854069620370865,
-0.05638163164258003,
-0.04863668978214264,
0.03747234493494034,
0.042074691504240036,
0.011533440090715885,
-0.029080256819725037,
-0.005748255178332329,
0.03922916576266289,
-0.02064961940050125,
0.03964453563094139,
0.043714024126529694,
-0.06160159409046173,
0.020311960950493813,
-0.08085785806179047,
0.12547551095485687,
0.11676772683858871,
-0.1026487872004509,
-0.011152050457894802,
-0.08347159624099731,
0.04467053338885307,
-0.057556021958589554,
-0.05169966444373131,
-0.037331774830818176,
1.5442213984073072e-33,
-0.018967796117067337,
0.020105548202991486,
0.033748093992471695,
0.08065074682235718,
-0.01277544442564249,
-0.02889682725071907,
-0.003954659216105938,
0.08251378685235977,
0.009499641135334969,
0.006050400901585817,
-0.029720233753323555,
0.02170487679541111,
-0.05189313367009163,
0.13669399917125702,
0.023443644866347313,
0.009971060790121555,
-0.059329550713300705,
0.02275879867374897,
-0.03474055975675583,
-0.05741075426340103,
-0.03811272233724594,
-0.06787097454071045,
0.016896143555641174,
-0.042859192937612534,
0.031139323487877846,
0.051188163459300995,
0.07206638902425766,
-0.07929714024066925,
-0.06764998286962509,
-0.018158553168177605,
-0.14036232233047485,
-0.01277620904147625,
0.04695671796798706,
0.05846647545695305,
-0.0169202983379364,
0.02933373674750328,
0.030170369893312454,
0.015716440975666046,
-0.02754661813378334,
-0.008426484651863575,
-0.011677171103656292,
0.05025621876120567,
-0.02757313661277294,
-0.020734647288918495,
-0.10323908925056458,
-0.02540656365454197,
-0.016139592975378036,
0.009699782356619835,
-0.0657924935221672,
-0.03402923047542572,
0.021300042048096657,
0.02659566141664982,
-0.06043627858161926,
-0.056459691375494,
0.032803118228912354,
-0.019313015043735504,
-0.019519886001944542,
-0.0013737819390371442,
0.022686682641506195,
0.024192526936531067,
0.007144724018871784,
0.009769407100975513,
0.009382382035255432,
0.09559094905853271,
0.01890162192285061,
0.0702739730477333,
-0.08206438273191452,
0.018948856741189957,
0.03279367461800575,
0.010640185326337814,
-0.011568344198167324,
0.08708208799362183,
-0.004343836568295956,
-0.056866224855184555,
0.05452762171626091,
0.038558803498744965,
-0.008667363785207272,
-0.1155560240149498,
-0.04110383987426758,
0.11789394915103912,
0.05103262886404991,
-0.0486835241317749,
-0.00927343126386404,
-0.03634463623166084,
-0.0703263133764267,
-0.08901120722293854,
0.017445962876081467,
-0.09678225964307785,
-0.01579531468451023,
-0.05834722891449928,
-0.04519601911306381,
-0.06074821576476097,
-0.010210545733571053,
0.02303903177380562,
-0.05730700492858887,
-1.4625416224274436e-33,
-0.03910447657108307,
0.01044115237891674,
-0.05402238667011261,
-0.0009774299105629325,
-0.003853532951325178,
-0.013240847736597061,
0.003824563231319189,
-0.008350263349711895,
-0.01909635029733181,
-0.023295089602470398,
0.03622101619839668,
-0.0874842032790184,
0.030771197751164436,
0.02275417558848858,
0.008028258569538593,
-0.02481485903263092,
-0.08161135762929916,
0.03919529542326927,
0.000036692494177259505,
0.05242205038666725,
0.04475990682840347,
0.05932195112109184,
-0.042681872844696045,
0.013105847872793674,
-0.02424655482172966,
0.05931155011057854,
-0.014053111895918846,
0.05515645816922188,
-0.015186837874352932,
-0.025393353775143623,
-0.022226138040423393,
0.0004101736703887582,
0.012208876200020313,
-0.041747432202100754,
-0.06776470690965652,
0.05659317225217819,
0.009915788657963276,
-0.06572730094194412,
-0.05752330273389816,
0.03844256326556206,
0.094997338950634,
0.0016005425713956356,
-0.07721386849880219,
0.03018389642238617,
-0.04347353056073189,
0.02135825715959072,
-0.05099048092961311,
-0.04801984876394272,
0.11222805082798004,
0.056388068944215775,
-0.04304536432027817,
0.010947168804705143,
-0.07009018212556839,
-0.030374102294445038,
-0.04434244707226753,
-0.11548963934183121,
0.011640425771474838,
-0.0809500589966774,
0.008006901480257511,
-0.04663126915693283,
-0.05194913595914841,
-0.021917732432484627,
0.05417158827185631,
-0.08751163631677628,
-0.006331051234155893,
-0.08862236142158508,
0.00013986500562168658,
0.013820528984069824,
-0.014868374913930893,
0.003935311920940876,
0.01492566429078579,
-0.0016579318325966597,
-0.0015797254163771868,
0.010085183195769787,
0.013198374770581722,
-0.05498809367418289,
0.03982499614357948,
-0.06206769123673439,
-0.043937161564826965,
-0.058101341128349304,
0.027293676510453224,
-0.05509576201438904,
0.03619743138551712,
0.10757790505886078,
0.03012632206082344,
0.12829338014125824,
0.013670976273715496,
0.056318096816539764,
0.03582252934575081,
0.044345222413539886,
-0.039135705679655075,
0.050715502351522446,
-0.026116155087947845,
0.06845234334468842,
-0.007266244385391474,
-4.0148439239828804e-8,
-0.12211989611387253,
0.03482675924897194,
-0.07130211591720581,
-0.003975569736212492,
-0.04245174303650856,
-0.02874746546149254,
-0.02603679522871971,
0.0433594174683094,
-0.0011627351632341743,
0.01995190419256687,
0.007207581773400307,
-0.032084591686725616,
-0.03424597531557083,
-0.09400128573179245,
0.01114292349666357,
0.05329137295484543,
0.05265618860721588,
0.009385183453559875,
0.07728097587823868,
-0.05233160778880119,
0.04442404955625534,
0.05941183120012283,
0.03014877438545227,
-0.02826681360602379,
0.05032207444310188,
-0.10386382043361664,
-0.02799571305513382,
0.0802575871348381,
0.07144250720739365,
-0.0590532124042511,
-0.06205781176686287,
0.09255325049161911,
0.004618357867002487,
-0.02205367386341095,
-0.03298935666680336,
0.05822135880589485,
0.05209747701883316,
-0.033234983682632446,
-0.0433751605451107,
0.025701506063342094,
-0.005364709068089724,
0.0775994211435318,
-0.0430428571999073,
0.0071966503746807575,
-0.033675748854875565,
0.032477572560310364,
0.005937875248491764,
-0.05348023399710655,
0.03243759274482727,
-0.08407054841518402,
0.0067014847882092,
0.020813914015889168,
-0.02262263558804989,
0.03885293006896973,
0.09061915427446365,
0.07731834799051285,
-0.006801661103963852,
-0.00032001399085856974,
0.013450362719595432,
0.04922901466488838,
-0.02394159324467182,
0.010003024712204933,
-0.0910043939948082,
-0.023314254358410835
] |
Understanding and Mitigating Bottlenecks of State Space Models through the Lens of Recency and Over-smoothing | https://openreview.net/forum?id=pymXpl4qvi | [
"Peihao Wang",
"Ruisi Cai",
"Yuehao Wang",
"Jiajun Zhu",
"Pragya Srivastava",
"Zhangyang Wang",
"Pan Li"
] | Poster | Structured State Space Models (SSMs) have emerged as alternatives to transformers.
While SSMs are often regarded as effective in capturing long-sequence dependencies, we rigorously demonstrate that they are inherently limited by strong recency bias.
Our empirical studies also reveal that this bias impairs the models' ability to recall distant information and introduces robustness issues. Our scaling experiments then discovered that deeper structures in SSMs can facilitate the learning of long contexts.
However, subsequent theoretical analysis reveals that as SSMs increase in depth, they exhibit another inevitable tendency toward over-smoothing, e.g., token representations becoming increasingly indistinguishable.
This *fundamental dilemma* between recency and over-smoothing hinders the scalability of existing SSMs.
Inspired by our theoretical findings, we propose to *polarize* two channels of the state transition matrices in SSMs, setting them to zero and one, respectively, simultaneously addressing recency bias and over-smoothing.
Experiments demonstrate that our polarization technique consistently enhances the associative recall accuracy of long-range tokens and unlocks SSMs to benefit further from deeper architectures.
All source codes are released at https://github.com/VITA-Group/SSM-Bottleneck. | State Space Models, Long-context LLMs, Over-smoothing | We theoretically and empirically reveal and tackle locality and over-smoothing bottlenecks for state space models. | 1,195 | 2501.00658 | [
-0.06601346284151077,
-0.12445291131734848,
-0.00421274546533823,
0.048433009535074234,
0.011988901533186436,
0.022742845118045807,
-0.0480547733604908,
-0.035054951906204224,
0.064057856798172,
-0.05276858061552048,
-0.02907070517539978,
0.021238887682557106,
0.0009752523037604988,
0.009595921263098717,
0.022863183170557022,
-0.03492327407002449,
0.07361658662557602,
0.07481884956359863,
-0.09415936470031738,
-0.03122047521173954,
0.06988617032766342,
0.0030746720731258392,
-0.020586520433425903,
0.012094065546989441,
0.05091743543744087,
0.0676518976688385,
-0.0021701508667320013,
-0.03682991862297058,
0.03036010079085827,
-0.028316199779510498,
0.041454073041677475,
0.0436418391764164,
-0.04566636309027672,
-0.012137668207287788,
-0.00952575821429491,
0.0712243840098381,
-0.024338915944099426,
0.01208033412694931,
-0.003305077785626054,
-0.04343455284833908,
0.015534591861069202,
0.0609981007874012,
0.0244662556797266,
0.1095886155962944,
0.03826306387782097,
-0.018328355625271797,
0.06225273385643959,
-0.07117285579442978,
-0.02559363842010498,
-0.023013755679130554,
-0.06668392568826675,
0.0546967089176178,
-0.04508088529109955,
0.09916714578866959,
-0.031446244567632675,
0.04676119610667229,
0.05064896494150162,
0.08641353994607925,
-0.06370475143194199,
-0.06502161175012589,
-0.10769432038068771,
-0.12951314449310303,
-0.036207281053066254,
0.0005738760810345411,
-0.0781661719083786,
0.0053162039257586,
-0.008087201975286007,
0.043824970722198486,
-0.020261459052562714,
0.06638745963573456,
-0.049896206706762314,
0.043721187859773636,
-0.020443325862288475,
0.026621269062161446,
0.054683927446603775,
0.021212853491306305,
0.05909198150038719,
-0.004595139529556036,
0.08049731701612473,
-0.07088920474052429,
0.05918115749955177,
-0.030168404802680016,
0.009622537530958652,
0.03389216214418411,
0.08989717811346054,
0.004933428019285202,
0.00593620166182518,
-0.04540441930294037,
0.04617872089147568,
0.039494000375270844,
0.015371426939964294,
-0.10604144632816315,
0.051726218312978745,
-0.040090449154376984,
0.05090875178575516,
0.01156216487288475,
0.014292705804109573,
-0.03698471561074257,
0.017874063923954964,
0.09999158978462219,
0.007505184505134821,
0.06983380019664764,
0.015982631593942642,
-0.022680439054965973,
0.01151344459503889,
-0.02443116530776024,
-0.020950863137841225,
0.04968198761343956,
0.021775685250759125,
-0.051715604960918427,
0.002833201549947262,
0.03803545609116554,
0.00041076578781940043,
0.05881974846124649,
-0.025267131626605988,
-0.027051866054534912,
0.02634439989924431,
-0.0869147852063179,
-0.017462829127907753,
0.04846939817070961,
-0.003194355173036456,
-0.04483542963862419,
-0.034397538751363754,
0.06788507103919983,
-0.08225830644369125,
0.002816862193867564,
-0.06958648562431335,
3.270904131975204e-33,
-0.029007157310843468,
0.03446761146187782,
-0.006106236018240452,
0.05657830461859703,
-0.00989099033176899,
0.04498175159096718,
0.009763327427208424,
0.017937568947672844,
-0.04716179147362709,
0.03292851522564888,
0.028438474982976913,
0.10577019304037094,
-0.0018394752405583858,
0.0423196479678154,
0.032650239765644073,
-0.030092408880591393,
-0.07569930702447891,
0.06604772061109543,
-0.03729903697967529,
-0.0647294744849205,
-0.032495465129613876,
-0.01073412410914898,
0.005125649273395538,
-0.09959981590509415,
0.01696447841823101,
-0.01948612928390503,
0.046361591666936874,
0.0351673923432827,
-0.06183505430817604,
0.008010338991880417,
-0.09247549623250961,
0.07926345616579056,
-0.014080488122999668,
0.04380566254258156,
0.07635148614645004,
-0.0074552204459905624,
0.039679814130067825,
-0.09471159428358078,
0.008915041573345661,
-0.09219870716333389,
0.007220251020044088,
0.03175099939107895,
-0.0334046296775341,
0.04178234562277794,
-0.08475112915039062,
-0.11984883248806,
0.04457150027155876,
-0.08862517774105072,
-0.07122117280960083,
-0.019744569435715675,
0.05431097745895386,
0.05149902403354645,
-0.06663794070482254,
-0.01914454996585846,
0.03678826615214348,
-0.04401697218418121,
0.01779227703809738,
0.0735485851764679,
0.0030902819707989693,
0.07588151842355728,
-0.06055460125207901,
0.028076566755771637,
0.047294504940509796,
0.0632789134979248,
0.0411359928548336,
0.07160836458206177,
-0.043154407292604446,
0.13587841391563416,
0.03519558161497116,
-0.013651488348841667,
0.00942181795835495,
0.014218322932720184,
-0.07579576969146729,
-0.04668789729475975,
0.029636478051543236,
-0.054490018635988235,
0.09622229635715485,
-0.042509790509939194,
-0.0662551075220108,
-0.018896985799074173,
0.022871756926178932,
-0.05565453693270683,
-0.047031424939632416,
0.030454672873020172,
-0.06482461094856262,
-0.07590179145336151,
0.017077334225177765,
-0.06908343732357025,
-0.072281114757061,
-0.042319562286138535,
-0.043380554765462875,
-0.05229981243610382,
0.04991063103079796,
0.041373878717422485,
-0.06940484046936035,
-4.2209188298426064e-33,
-0.03036721982061863,
-0.04852019250392914,
-0.03198471665382385,
0.03382017835974693,
-0.09556298702955246,
-0.10670086741447449,
0.012350451201200485,
0.01041350420564413,
-0.03619471564888954,
-0.06802915781736374,
-0.06179054453969002,
-0.0537540428340435,
0.05719810351729393,
-0.004704898688942194,
0.05546291172504425,
0.01139808353036642,
0.034246187657117844,
-0.039792537689208984,
0.04609077051281929,
0.013506187126040459,
0.09155779331922531,
-0.008186155930161476,
-0.10564272850751877,
0.045263100415468216,
-0.04247371852397919,
0.09493742883205414,
-0.0888248085975647,
0.041783466935157776,
0.03247204050421715,
-0.033188868314027786,
-0.047299183905124664,
-0.0608602911233902,
-0.030014848336577415,
0.015845343470573425,
-0.04533207789063454,
0.06923994421958923,
0.000454185763373971,
0.0033544537145644426,
-0.007375743240118027,
0.021046340465545654,
0.03359892964363098,
-0.012511728331446648,
-0.0250437892973423,
0.06646682322025299,
0.023517344146966934,
-0.03192536532878876,
-0.011865598149597645,
0.0598159059882164,
-0.039990149438381195,
-0.02515695057809353,
-0.035518743097782135,
0.06490761041641235,
-0.07920987904071808,
0.04305015504360199,
-0.07986611127853394,
-0.04044459015130997,
-0.03813452273607254,
-0.0028171243611723185,
0.05365480110049248,
-0.02150283195078373,
-0.0037556716706603765,
-0.04912472143769264,
-0.011894033290445805,
-0.07246725261211395,
0.014449527487158775,
-0.04916238412261009,
0.025287669152021408,
-0.05229370296001434,
0.02212069369852543,
0.008519038558006287,
0.002394681330770254,
-0.024858001619577408,
-0.08893895149230957,
-0.0252037663012743,
0.01269996166229248,
-0.028142049908638,
0.07667286694049835,
-0.08497335016727448,
-0.015511637553572655,
0.0011414469918236136,
-0.03636353090405464,
0.041393689811229706,
-0.0032697971910238266,
0.07017302513122559,
0.050622958689928055,
0.12993866205215454,
-0.0028528275433927774,
0.08719058334827423,
0.03061412274837494,
0.004603540990501642,
-0.020213380455970764,
0.014313232153654099,
-0.002757682465016842,
0.06580370664596558,
-0.03347392380237579,
-5.00275127990335e-8,
-0.07720029354095459,
0.02466362528502941,
0.03545606881380081,
0.04108833894133568,
0.08776894211769104,
-0.05334397777915001,
-0.010025039315223694,
0.0903259739279747,
-0.04132688790559769,
-0.05764668807387352,
0.040401097387075424,
0.0016444091452285647,
-0.07138972729444504,
-0.05356718599796295,
0.04628174751996994,
0.07641152292490005,
0.03208541125059128,
0.01105447392910719,
0.0019079182529821992,
0.05599160119891167,
-0.005194421857595444,
0.02242448180913925,
-0.019594261422753334,
0.01358036883175373,
-0.0037168345879763365,
-0.0663142129778862,
0.028588324785232544,
0.10967586189508438,
0.045435041189193726,
-0.02333723194897175,
-0.00608784006908536,
0.032984670251607895,
0.01817106269299984,
-0.026954205706715584,
-0.035073019564151764,
0.05046279355883598,
0.06934010982513428,
-0.022375164553523064,
0.06121945381164551,
-0.010103260166943073,
0.015255825594067574,
0.045788757503032684,
-0.08408552408218384,
0.09451179951429367,
-0.06254612654447556,
0.06688304990530014,
0.0328325517475605,
-0.04527252912521362,
-0.012861110270023346,
-0.035985905677080154,
0.015227455645799637,
-0.017052164301276207,
-0.06787947565317154,
0.05885683745145798,
0.02052212320268154,
0.0449090339243412,
-0.0019950612913817167,
-0.03605405241250992,
-0.019926713779568672,
0.09978337585926056,
-0.023075886070728302,
0.045735858380794525,
-0.14843086898326874,
0.0004929418209940195
] |
LLaMaFlex: Many-in-one LLMs via Generalized Pruning and Weight Sharing | https://openreview.net/forum?id=AyC4uxx2HW | [
"Ruisi Cai",
"Saurav Muralidharan",
"Hongxu Yin",
"Zhangyang Wang",
"Jan Kautz",
"Pavlo Molchanov"
] | Poster | Large Language Model (LLM) providers typically train a family of models, each of a different size targeting a specific deployment scenario. Models in the family are all trained from scratch, making the process extremely resource intensive.
Recent work has successfully reduced the cost of training model families through a combination of structured pruning and knowledge distillation; here, only the largest model in the family is trained from scratch, and smaller models are obtained via pruning. We observe that while effective, this strategy must still perform pruning and distillation with hundreds of billions of training tokens for every new model, keeping overall training costs high.
In this work, we introduce a novel nested weight-shared architecture named LLaMaFlex that can be pruned across both width and depth dimensions in a zero-shot manner to instantly yield a large number of highly accurate compressed models.
LLaMaFlex starts from a pretrained model, and only requires a single continued training phase consisting of ~60B tokens, which trains the elastic network and an end-to-end Gumbel Softmax-based router; this router is able to interpolate smoothly across model sizes, enabling the "train once, deploy many'' paradigm.
We train LLaMaFlex on Llama 3.1 8B and use it to zero-shot generate a family of compressed models that achieves accuracy on par with or better than state-of-the-art pruned, elastic/flexible, and trained-from-scratch models. | large language models, elastic networks, training efficiency, inference efficiency | null | 1,194 | null | [
-0.054577115923166275,
-0.055584799498319626,
0.04909680411219597,
0.024695919826626778,
0.007125332485884428,
-0.05359254032373428,
-0.09791061282157898,
-0.018553752452135086,
0.013264262117445469,
-0.04465411603450775,
-0.027671266347169876,
-0.004217457491904497,
0.07504242658615112,
0.04716339707374573,
0.002494116546586156,
0.03806130960583687,
0.052516695111989975,
0.11235973238945007,
-0.12160786986351013,
-0.0004217368259560317,
-0.005739712622016668,
-0.04372101277112961,
-0.014046337455511093,
0.014552300795912743,
0.06567521393299103,
-0.0166082251816988,
0.008316071704030037,
-0.03987547755241394,
0.05340121313929558,
-0.07830561697483063,
0.05310437083244324,
-0.03598186373710632,
-0.009465419687330723,
0.0369083434343338,
-0.028071613982319832,
0.04557080194354057,
-0.03652322664856911,
0.012260441668331623,
0.015422952361404896,
-0.04994739592075348,
0.02007465995848179,
0.04640573263168335,
-0.02341599389910698,
0.05153346806764603,
0.059088610112667084,
-0.009904017671942711,
0.028004176914691925,
0.007247996050864458,
0.010463161394000053,
0.008539105765521526,
-0.014456851407885551,
-0.09231863915920258,
0.0074273524805903435,
0.0695548728108406,
-0.05848852917551994,
-0.000997695024125278,
-0.00952838733792305,
-0.00585402175784111,
-0.05610230937600136,
0.000695570488460362,
-0.032376114279031754,
-0.06246617063879967,
-0.039294976741075516,
-0.0074696349911391735,
-0.049085378646850586,
-0.0476650670170784,
0.07592964172363281,
0.11721030622720718,
0.053670596331357956,
-0.02067405730485916,
0.04342076927423477,
0.04067905247211456,
-0.017731565982103348,
0.11220279335975647,
0.024609815329313278,
0.053281351923942566,
0.10463311523199081,
0.048484738916158676,
0.05875835567712784,
-0.0014768237015232444,
-0.020430630072951317,
-0.03368029370903969,
0.029909394681453705,
-0.02501327358186245,
0.07200606912374496,
-0.0557280071079731,
0.004997935611754656,
0.06280406564474106,
0.039268575608730316,
-0.024403931573033333,
-0.08427390456199646,
0.017538422718644142,
0.040280699729919434,
0.027605213224887848,
-0.010167993605136871,
0.007919865660369396,
-0.03152856230735779,
-0.02749491110444069,
-0.10407615453004837,
0.10209213942289352,
-0.02686571702361107,
0.0344884917140007,
0.11648038774728775,
-0.039966244250535965,
0.008054335601627827,
-0.022759102284908295,
0.001985896145924926,
0.06526622176170349,
0.03012770414352417,
-0.10023117065429688,
0.022275740280747414,
0.08480088412761688,
-0.011243418790400028,
0.04615618661046028,
0.04402676597237587,
0.007095761131495237,
-0.05958837270736694,
-0.060148611664772034,
0.011367673985660076,
0.057607680559158325,
-0.06961904466152191,
0.00912134163081646,
-0.031120605766773224,
0.010583007708191872,
0.03208450600504875,
-0.035394445061683655,
-0.060707513242959976,
3.646259682001915e-33,
0.06918776035308838,
0.12332998961210251,
-0.01892770454287529,
0.0019545613322407007,
0.0971868634223938,
0.0004178849922027439,
0.046956971287727356,
0.08873755484819412,
-0.09240199625492096,
0.0033987858332693577,
-0.08905386179685593,
0.0002293290162924677,
-0.050893448293209076,
0.04902520403265953,
0.08393153548240662,
-0.02425430156290531,
-0.03828316554427147,
0.06020951271057129,
0.02832726389169693,
-0.025970863178372383,
0.04233447462320328,
0.005640835966914892,
-0.010198786854743958,
-0.06349843740463257,
0.026461588218808174,
0.023075038567185402,
0.13921605050563812,
-0.024659134447574615,
-0.08660281449556351,
0.02563854493200779,
-0.03777734935283661,
-0.04122890904545784,
0.015550106763839722,
0.018457477912306786,
0.028593672439455986,
-0.03745068982243538,
-0.03414567559957504,
-0.06469441950321198,
0.012372990138828754,
-0.04320477694272995,
-0.011328821070492268,
0.030178252607584,
-0.042691394686698914,
-0.0725119337439537,
-0.10586071759462357,
-0.0656156912446022,
-0.00036783749237656593,
0.06200103089213371,
-0.05314987897872925,
-0.014829238876700401,
0.09342353790998459,
0.013649544678628445,
-0.025136087089776993,
0.015873437747359276,
-0.07363574206829071,
-0.028060758486390114,
0.02206440456211567,
0.011440128087997437,
0.07251553237438202,
0.06254233419895172,
-0.012926979921758175,
0.010755350813269615,
-0.03052333928644657,
0.158049076795578,
0.0547693707048893,
0.006433655042201281,
0.00445131678134203,
0.05626725032925606,
0.05128495767712593,
-0.029846275225281715,
-0.08405278623104095,
-0.020267708227038383,
0.017174381762742996,
-0.07871701568365097,
0.050540074706077576,
-0.07179410010576248,
0.08880554884672165,
-0.07973207533359528,
0.017656773328781128,
0.028358224779367447,
-0.000301307940389961,
0.03320573642849922,
-0.03304304927587509,
-0.022339731454849243,
-0.06080620363354683,
-0.027134139090776443,
0.016388801857829094,
-0.038688916712999344,
-0.033485691994428635,
0.014619388617575169,
-0.04015399515628815,
-0.08630456030368805,
-0.023337189108133316,
-0.004922181833535433,
0.022334884852170944,
-1.681832297625249e-33,
-0.012220426462590694,
-0.011911844834685326,
-0.046131476759910583,
0.08780109137296677,
0.019700929522514343,
-0.09653440862894058,
0.0021328399889171124,
0.0635678619146347,
-0.07978369295597076,
-0.023383161053061485,
-0.03579492121934891,
-0.05309121310710907,
0.045127540826797485,
-0.015306197106838226,
-0.027678996324539185,
0.003963708411902189,
-0.011119097471237183,
-0.06381416320800781,
0.03144773468375206,
-0.023128926753997803,
-0.016217375174164772,
0.08423805981874466,
-0.045259151607751846,
0.014722202904522419,
-0.014308718964457512,
0.04183796048164368,
-0.018033921718597412,
0.07759146392345428,
-0.002649219473823905,
0.03637665510177612,
-0.030253764241933823,
-0.03661298006772995,
0.03616398572921753,
-0.09527530521154404,
-0.10106004029512405,
0.05133751779794693,
0.03613040968775749,
0.1009637713432312,
0.06194488704204559,
0.06992252171039581,
0.06438952684402466,
-0.07318936288356781,
-0.09332368522882462,
0.03749797120690346,
-0.021133415400981903,
0.0014832752058282495,
-0.07499711215496063,
-0.09663897007703781,
0.0051575759425759315,
-0.005889804568141699,
-0.015081014484167099,
0.01571604609489441,
0.02661328762769699,
0.011673924513161182,
0.01933140680193901,
-0.11173861473798752,
-0.06546837836503983,
0.012984402477741241,
-0.00865041185170412,
-0.023705979809165,
-0.030197003856301308,
-0.051368068903684616,
-0.014272910542786121,
-0.01679469458758831,
-0.011293557472527027,
-0.048710647970438004,
0.01111616101115942,
0.03938191756606102,
-0.05017109960317612,
0.026846062391996384,
-0.021475696936249733,
0.014185814186930656,
0.00007489504059776664,
0.0518951416015625,
-0.0574524849653244,
0.01538735069334507,
-0.0049837310798466206,
-0.054593998938798904,
0.02050795964896679,
-0.002989385509863496,
-0.015271652489900589,
-0.010135121643543243,
0.02935774251818657,
0.006677640601992607,
0.12096071988344193,
0.02248421683907509,
0.08687283843755722,
0.09893310815095901,
0.008104386739432812,
0.04970681667327881,
-0.02311129868030548,
0.06961525231599808,
0.07888421416282654,
0.023329325020313263,
0.006336386315524578,
-4.480989801436408e-8,
-0.0803842693567276,
0.08253497630357742,
-0.025428039953112602,
0.07301979511976242,
0.03109780140221119,
-0.03420665115118027,
-0.02086338773369789,
0.1353142261505127,
0.0005778258200734854,
0.01510717160999775,
0.08693484216928482,
-0.0033165314234793186,
-0.10324804484844208,
-0.0076358141377568245,
-0.002107609063386917,
0.07439267635345459,
-0.0084916353225708,
-0.02544117160141468,
-0.02157111093401909,
-0.00619336636736989,
0.010191576555371284,
0.0568046011030674,
0.045986708253622055,
0.02288629487156868,
0.09307830035686493,
-0.08578450977802277,
-0.052787669003009796,
0.11209844797849655,
0.0752154067158699,
0.036288756877183914,
-0.08218292891979218,
-0.009571458213031292,
-0.012702614068984985,
-0.04230600595474243,
-0.03771207109093666,
0.07778321951627731,
0.005471532698720694,
0.028711583465337753,
-0.009787008166313171,
-0.047539252787828445,
0.024664143100380898,
0.023568101227283478,
-0.02736440673470497,
0.017268579453229904,
-0.0059534767642617226,
-0.05689568072557449,
-0.06712689250707626,
-0.0696869045495987,
0.05783356353640556,
0.0008059581741690636,
0.023962736129760742,
-0.015984611585736275,
-0.05262031778693199,
0.039223358035087585,
-0.006049972027540207,
0.06800064444541931,
-0.04834723845124245,
-0.038486406207084656,
0.04734152927994728,
0.019095301628112793,
0.03486085683107376,
-0.09514182060956955,
-0.07220146059989929,
0.0035311365500092506
] |
Generative Inbetweening: Adapting Image-to-Video Models for Keyframe Interpolation | https://openreview.net/forum?id=ykD8a9gJvy | [
"Xiaojuan Wang",
"Boyang Zhou",
"Brian Curless",
"Ira Kemelmacher-Shlizerman",
"Aleksander Holynski",
"Steve Seitz"
] | Poster | We present a method for generating video sequences with coherent motion between a pair of input keyframes. We adapt a pretrained large-scale image-to-video diffusion model (originally trained to generate videos moving forward in time from a single input image) for keyframe interpolation, i.e., to produce a video between two input frames. We accomplish this adaptation through a lightweight fine-tuning technique that produces a version of the model that instead predicts videos moving backwards in time from a single input image. This model (along with the original forward-moving model) is subsequently used in a dual-directional diffusion sampling process that combines the overlapping model estimates starting from each of the two keyframes. Our experiments shows that our method outperforms both existing diffusion-based methods and traditional frame interpolation techniques. | generative keyframe interpolation, image-to-video diffusion models | null | 1,189 | 2408.15239 | [
-0.0704270452260971,
-0.14673881232738495,
0.050241872668266296,
0.005738519597798586,
0.10408411175012589,
-0.007919528521597385,
-0.06071312725543976,
-0.063478022813797,
0.06005425006151199,
-0.08611278980970383,
-0.04605720937252045,
-0.0006426576292142272,
-0.0536121241748333,
0.025447990745306015,
-0.04943414032459259,
-0.06594157218933105,
-0.015906617045402527,
0.06828174740076065,
-0.08135438710451126,
-0.06563713401556015,
-0.03574234992265701,
-0.05160079151391983,
0.0014657049905508757,
-0.03696700558066368,
0.06925278156995773,
0.06288914382457733,
-0.0021419767290353775,
-0.036487702280282974,
0.044545818120241165,
0.00010979888611473143,
0.04875127226114273,
-0.03584985435009003,
-0.08375933021306992,
-0.008918861858546734,
-0.09470422565937042,
0.01217435859143734,
-0.02876613475382328,
0.04097997024655342,
-0.02801140770316124,
-0.028928272426128387,
0.029408728703856468,
0.08831758052110672,
0.045207567512989044,
-0.0029321785550564528,
0.005921490956097841,
0.03501347824931145,
0.002377944765612483,
-0.05217890441417694,
0.0010868163080886006,
0.04537349194288254,
-0.0928952693939209,
0.04106809198856354,
-0.04159048944711685,
0.06726543605327606,
-0.04012364149093628,
0.024913165718317032,
0.09366601705551147,
-0.00564997922629118,
0.09971662610769272,
-0.046716611832380295,
-0.03610248118638992,
-0.0248428788036108,
0.010502738878130913,
0.03709488734602928,
-0.04526955261826515,
-0.06622689962387085,
0.07399661093950272,
0.0550074428319931,
-0.026567980647087097,
-0.002791428240016103,
-0.06332365423440933,
0.037403713911771774,
-0.03107236512005329,
0.011828313581645489,
-0.03731169179081917,
0.024553677067160606,
0.0456620417535305,
-0.024423914030194283,
-0.024338090792298317,
-0.0709087997674942,
0.10440892726182938,
-0.039579764008522034,
0.08065245300531387,
-0.027713656425476074,
0.04051097109913826,
-0.002147737191990018,
0.010476344265043736,
0.06152355670928955,
-0.0017268627416342497,
0.019123677164316177,
-0.07820621877908707,
0.06521379947662354,
-0.02848271280527115,
-0.013397810980677605,
0.003350045531988144,
0.022934485226869583,
-0.030485443770885468,
-0.05591791868209839,
0.045320719480514526,
0.08498670905828476,
0.014965321868658066,
-0.006842195987701416,
0.0009037923882715404,
0.008165244944393635,
0.09247566759586334,
-0.031533367931842804,
0.07398509979248047,
0.026316070929169655,
-0.0031651481986045837,
0.011872584000229836,
0.048570651561021805,
0.011722362600266933,
0.05805090069770813,
0.012112399563193321,
0.0762079507112503,
-0.0531679205596447,
-0.00241727102547884,
0.016032686457037926,
0.024003231897950172,
0.046532221138477325,
-0.013792011886835098,
-0.007897808216512203,
-0.06393784284591675,
-0.010192382149398327,
-0.008598008193075657,
-0.05782226473093033,
0.006804381497204304,
-8.527762669228224e-35,
-0.0052549527026712894,
-0.04137292876839638,
0.10173584520816803,
0.05296593904495239,
0.06521187722682953,
0.04265367612242699,
-0.0027748423162847757,
-0.03825853019952774,
-0.025234030559659004,
-0.01603776402771473,
-0.03470536321401596,
-0.05526834726333618,
-0.05806741863489151,
0.041504260152578354,
-0.022102147340774536,
-0.038139741867780685,
-0.04063498228788376,
0.0492137111723423,
0.07728710770606995,
0.11201812326908112,
0.017181485891342163,
0.0026302186306566,
-0.04475098103284836,
-0.011093725450336933,
-0.03294862061738968,
0.049176909029483795,
0.044283125549554825,
-0.05865241587162018,
0.07071907818317413,
0.003644012613222003,
-0.1194753497838974,
-0.022956524044275284,
-0.09121778607368469,
-0.04288112372159958,
0.07410994172096252,
-0.017695089802145958,
-0.0013624209677800536,
-0.004204527009278536,
-0.056481365114450455,
-0.017478352412581444,
0.026410747319459915,
0.04602019488811493,
0.010009869001805782,
-0.021774059161543846,
-0.14103932678699493,
-0.031242938712239265,
0.056569553911685944,
0.07338590919971466,
-0.035695020109415054,
0.017341570928692818,
0.04587513208389282,
-0.013374056667089462,
-0.01617191731929779,
-0.10720403492450714,
-0.049723245203495026,
0.019945377483963966,
0.10826035588979721,
-0.04571453109383583,
0.033400874584913254,
0.0034462513867765665,
0.003687930991873145,
0.05111939087510109,
-0.0542505607008934,
0.015598311088979244,
0.03622442111372948,
-0.050825245678424835,
-0.036773987114429474,
0.05732348561286926,
0.01749788410961628,
-0.008875923231244087,
-0.038839876651763916,
-0.014372416771948338,
-0.030949821695685387,
-0.06013541296124458,
0.0437636636197567,
0.002219438785687089,
-0.01716686598956585,
0.004761775955557823,
-0.011811458505690098,
0.047165002673864365,
-0.06750863045454025,
-0.024468479678034782,
-0.002495960798114538,
-0.06600864976644516,
0.013713955879211426,
0.013194681145250797,
-0.04135081171989441,
-0.024852506816387177,
-0.03891695663332939,
-0.10485347360372543,
-0.047855161130428314,
-0.008818441070616245,
-0.034139953553676605,
0.02629169449210167,
0.013619815930724144,
-5.7211328421619965e-34,
0.05739642307162285,
0.056831587105989456,
0.0480605885386467,
0.10959923267364502,
0.004285553004592657,
0.008986776694655418,
0.02767416089773178,
-0.014844627119600773,
0.11838209629058838,
-0.06859959661960602,
-0.05255761370062828,
-0.05132889002561569,
0.0060895998030900955,
-0.059676747769117355,
-0.028436757624149323,
-0.06659089028835297,
0.05733788013458252,
-0.016534851863980293,
-0.06173323094844818,
-0.0011316196760162711,
0.008520415052771568,
0.00886248517781496,
-0.01456470787525177,
0.03688127174973488,
-0.06532590091228485,
0.08016454428434372,
0.03555208444595337,
0.07330138236284256,
-0.03139067813754082,
-0.05662119388580322,
0.01760447397828102,
-0.03405001014471054,
0.07687100023031235,
-0.07135763764381409,
-0.14701057970523834,
0.07397850602865219,
0.04849176108837128,
-0.035121578723192215,
-0.07919301837682724,
0.08383035659790039,
0.07174451649188995,
0.018132980912923813,
-0.018351154401898384,
0.10229510068893433,
-0.02137570083141327,
0.002498920541256666,
-0.0849098265171051,
0.010482980869710445,
-0.023067694157361984,
0.02670344151556492,
-0.04834948480129242,
0.0908837839961052,
-0.1170637458562851,
0.010632995516061783,
0.00622523482888937,
-0.09949174523353577,
-0.09108209609985352,
-0.0014685970963910222,
0.0767405778169632,
-0.018456483259797096,
0.015367139130830765,
-0.04250001162290573,
-0.01892150565981865,
-0.06972748786211014,
0.04178900644183159,
0.01066657155752182,
-0.06053505465388298,
-0.09281987696886063,
0.04219626635313034,
0.03742605447769165,
0.043186526745557785,
-0.008060702122747898,
-0.008459124714136124,
0.020757688209414482,
0.02739071100950241,
-0.09098239243030548,
0.1043143942952156,
-0.05865808203816414,
0.03731340914964676,
-0.06630856543779373,
0.014015207067131996,
-0.044639382511377335,
0.054966140538454056,
0.010336966253817081,
-0.0017306263325735927,
0.042304567992687225,
0.06043550372123718,
0.016366198658943176,
0.03993019089102745,
0.04476453363895416,
-0.020344579592347145,
0.044101566076278687,
0.05011653155088425,
-0.003068184945732355,
-0.03529183566570282,
-3.7282863019072465e-8,
-0.060495078563690186,
0.07800565659999847,
0.024102138355374336,
0.05857529118657112,
-0.0746290385723114,
-0.05930689349770546,
0.07588160783052444,
0.026446891948580742,
0.07616951316595078,
-0.1417934149503708,
0.05895506963133812,
0.03212938457727432,
0.039138492196798325,
0.018272031098604202,
-0.004105773754417896,
0.07032705098390579,
0.013858562335371971,
-0.003504845080897212,
0.013699820265173912,
0.10242699831724167,
-0.052896495908498764,
0.03754447400569916,
-0.010959313251078129,
-0.052583012729883194,
0.012448824942111969,
0.05032854899764061,
-0.006413098890334368,
-0.021257800981402397,
0.1027970016002655,
-0.018944114446640015,
-0.03987020254135132,
0.013076127506792545,
-0.008367251604795456,
-0.06428924947977066,
0.0011004952248185873,
0.01223067007958889,
0.012302225455641747,
0.012150351889431477,
-0.014938467182219028,
-0.11297693848609924,
-0.04777088016271591,
0.09845223277807236,
-0.024429041892290115,
-0.004942728206515312,
-0.030816668644547462,
0.05701686069369316,
0.053892120718955994,
-0.009030275046825409,
0.013853148557245731,
-0.007417392451316118,
0.0347791388630867,
-0.016476081684231758,
-0.0033580365125089884,
0.020906412973999977,
0.11077570915222168,
-0.003488758113235235,
0.04426167532801628,
-0.016028936952352524,
0.05616900697350502,
0.04762022942304611,
-0.045562565326690674,
-0.030609339475631714,
0.006346839014440775,
0.04214388504624367
] |
Uncertainty-Aware Decoding with Minimum Bayes Risk | https://openreview.net/forum?id=hPpyUv1XyQ | [
"Nico Daheim",
"Clara Meister",
"Thomas Möllenhoff",
"Iryna Gurevych"
] | Poster | Despite their outstanding performance in the majority of scenarios, contemporary language models still occasionally generate undesirable outputs, for example, hallucinated text. While such behaviors have previously been linked to uncertainty, there is a notable lack of methods that actively consider uncertainty during text generation. In this work, we show how Minimum Bayes Risk (MBR) decoding, which selects model generations according to an expected risk, can be generalized into a principled uncertainty-aware decoding method. In short, we account for model uncertainty during decoding by incorporating a posterior over model parameters into MBR’s computation of expected risk. We show that this modified expected risk is useful for both choosing outputs and deciding when to abstain from generation and can provide improvements without incurring overhead. We benchmark different methods for learning posteriors and show that performance improves with prediction diversity. We release our code publicly. | mbr, uncertainty, llms, decoding, machine translation, language generation, variational learning | We generalize MBR into an uncertainty-aware decoding method. | 1,172 | 2503.05318 | [
-0.04329269379377365,
-0.0681304782629013,
-0.010571147315204144,
0.06600598990917206,
0.05973254144191742,
0.024280959740281105,
-0.028324777260422707,
-0.005353291053324938,
0.09019205719232559,
-0.05124719813466072,
-0.04409518092870712,
-0.047931868582963943,
0.1224992573261261,
-0.05564241483807564,
-0.08919040858745575,
-0.009010525420308113,
0.015748750418424606,
0.06000889092683792,
-0.05726892128586769,
-0.05877136439085007,
0.03797038272023201,
0.06977294385433197,
-0.009580553509294987,
0.013360725715756416,
0.0860239639878273,
-0.03542264550924301,
0.013887239620089531,
0.0597686842083931,
0.036288876086473465,
0.027652977034449577,
0.04671729356050491,
0.013335763476788998,
0.017733216285705566,
-0.06989938020706177,
0.004787388723343611,
0.01346120797097683,
-0.08611422777175903,
-0.05469455569982529,
0.05197490006685257,
-0.029591752216219902,
-0.0591764971613884,
0.022851144894957542,
0.038716595619916916,
0.045257434248924255,
0.028805896639823914,
-0.049303945153951645,
-0.01847568340599537,
-0.09325408935546875,
-0.00442093051970005,
-0.020566517487168312,
-0.005119659937918186,
-0.019198069348931313,
-0.006693600676953793,
-0.04082508385181427,
-0.05859929695725441,
-0.013056378811597824,
-0.011260158382356167,
-0.036621563136577606,
-0.05161290988326073,
-0.04846658185124397,
-0.10856770724058151,
-0.09514179080724716,
-0.029082490131258965,
-0.011471371166408062,
-0.0011362386867403984,
0.030246181413531303,
0.04891487956047058,
0.10180496424436569,
-0.04192940890789032,
0.05756838247179985,
-0.056395094841718674,
0.02722162753343582,
-0.07624302059412003,
0.0989660695195198,
0.004961045458912849,
0.040970612317323685,
0.020891744643449783,
-0.09042195230722427,
-0.014649770222604275,
-0.0022931229323148727,
-0.021618515253067017,
0.006072082556784153,
0.07505093514919281,
-0.0024006671737879515,
0.09548540413379669,
0.01645342819392681,
0.062435001134872437,
0.04203243553638458,
0.03985860198736191,
0.06887639313936234,
-0.1258983314037323,
-0.12239092588424683,
0.10967212915420532,
0.05702728405594826,
-0.06760378926992416,
0.059146273881196976,
-0.07484756410121918,
-0.07443231344223022,
-0.001427896087989211,
0.04404884949326515,
0.004636154510080814,
0.07033408433198929,
0.053981680423021317,
-0.09380005300045013,
-0.04534292593598366,
-0.0682666227221489,
0.027827063575387,
0.04630143567919731,
0.0013896781019866467,
-0.031696464866399765,
0.046477966010570526,
0.06150156259536743,
0.011529108509421349,
-0.031020840629935265,
-0.0007059158524498343,
0.0513775460422039,
0.023194391280412674,
-0.026886405423283577,
-0.015043215826153755,
0.10223372280597687,
-0.03389327600598335,
0.024174869060516357,
0.055131182074546814,
0.03296424821019173,
-0.05902194604277611,
-0.07618997246026993,
0.0564134418964386,
1.1616700092025145e-34,
0.06978970766067505,
-0.01802067458629608,
0.022713199257850647,
0.08254130184650421,
0.029110822826623917,
0.08939860016107559,
-0.002375858835875988,
-0.04104739427566528,
-0.028708146885037422,
-0.057220906019210815,
-0.02405296079814434,
-0.058685291558504105,
-0.043451812118291855,
-0.05065751075744629,
0.089378222823143,
0.014389230869710445,
-0.06997036933898926,
0.05919518321752548,
-0.0330674909055233,
0.029253307729959488,
0.0476570688188076,
0.014197231270372868,
-0.0206880085170269,
-0.09519944339990616,
-0.015213992446660995,
0.06332699209451675,
0.09280905872583389,
-0.05372940003871918,
-0.05923152714967728,
0.01571076177060604,
-0.13785317540168762,
-0.008859433233737946,
0.046453431248664856,
-0.029310468584299088,
0.03720146045088768,
-0.049840912222862244,
-0.025291807949543,
-0.03003765642642975,
-0.024486564099788666,
-0.017233116552233696,
0.006418165285140276,
0.01730857789516449,
0.057285308837890625,
0.00566150201484561,
-0.012544629164040089,
-0.039284493774175644,
-0.02512207441031933,
-0.07019131630659103,
-0.04937423765659332,
-0.06408470124006271,
0.049992501735687256,
0.048865046352148056,
-0.00267745996825397,
0.07510126382112503,
0.018526561558246613,
-0.056819915771484375,
0.02097475342452526,
0.00894837360829115,
0.03200310468673706,
0.017757689580321312,
0.013371857814490795,
0.06056782230734825,
0.006213781423866749,
0.04328792542219162,
0.06506622582674026,
0.0640740692615509,
-0.013428185135126114,
0.015819933265447617,
-0.022251956164836884,
-0.06174769252538681,
-0.027116229757666588,
-0.06443733721971512,
-0.01382069755345583,
-0.05368271842598915,
0.012767329812049866,
0.033162422478199005,
0.014513053931295872,
-0.009824393317103386,
0.020460253581404686,
-0.039083097130060196,
-0.0180270504206419,
-0.048323873430490494,
0.002308022463694215,
0.011322595179080963,
-0.019257210195064545,
0.0008142297156155109,
0.03797770291566849,
-0.03213462606072426,
-0.07684221863746643,
0.02997736819088459,
-0.025721602141857147,
-0.055129267275333405,
0.03175566345453262,
-0.03499816358089447,
0.01725962944328785,
-2.1601591324030772e-33,
-0.0752941146492958,
-0.010697943158447742,
0.03883586823940277,
0.004216072615236044,
-0.035734862089157104,
-0.12254678457975388,
0.0015260447980836034,
0.03996390849351883,
-0.0035530279856175184,
-0.09285341203212738,
-0.13772167265415192,
0.0034140583593398333,
0.062094949185848236,
-0.012762408703565598,
0.06587439030408859,
-0.02861800417304039,
0.023759875446558,
0.01047897431999445,
-0.020515350624918938,
0.03459678590297699,
0.03611772507429123,
0.04598882794380188,
-0.09786339849233627,
0.031233657151460648,
0.05150087550282478,
0.05853120610117912,
-0.016032060608267784,
0.07788503915071487,
0.025235923007130623,
-0.043710336089134216,
-0.08323454111814499,
0.0815022885799408,
0.03248853236436844,
0.01979190669953823,
-0.10876117646694183,
0.0011093359207734466,
0.035436663776636124,
-0.02584245800971985,
-0.029397979378700256,
0.1136433407664299,
0.018733685836195946,
-0.02237042412161827,
-0.071293406188488,
-0.05248964577913284,
-0.052809830754995346,
-0.031635090708732605,
-0.05273667722940445,
-0.036384157836437225,
0.05865173041820526,
-0.04449636861681938,
0.022302303463220596,
0.06969214975833893,
-0.06196959316730499,
0.05605742707848549,
-0.06054500490427017,
-0.11170806735754013,
-0.1265387237071991,
-0.03916371613740921,
0.07179347425699234,
0.08036760985851288,
-0.05649586394429207,
0.014036465436220169,
0.03861285746097565,
-0.0273121390491724,
-0.0045316992327570915,
-0.0315476655960083,
-0.04318462312221527,
-0.016545988619327545,
0.018361780792474747,
0.012051619589328766,
0.08270691335201263,
-0.06050441041588783,
-0.01966533623635769,
0.0531412810087204,
0.06324217468500137,
-0.022266030311584473,
0.04154961183667183,
0.02330995723605156,
-0.012954020872712135,
-0.023367295041680336,
0.04777536541223526,
0.05736906826496124,
0.023333624005317688,
0.060138557106256485,
0.09219148010015488,
0.0316823348402977,
0.013459229841828346,
-0.034616027027368546,
-0.023298213258385658,
-0.008984270505607128,
-0.03987935930490494,
0.05125585198402405,
0.009103137999773026,
0.04854808375239372,
-0.09221794456243515,
-3.95977899358968e-8,
-0.05001340061426163,
0.014717710204422474,
0.006498759146779776,
0.037552397698163986,
0.03810090944170952,
0.008438083343207836,
-0.07442621141672134,
-0.03196035698056221,
0.013867530971765518,
-0.07437177747488022,
0.08198302239179611,
-0.019977929070591927,
-0.03115314617753029,
-0.06678948551416397,
-0.020245524123311043,
0.11575358361005783,
-0.0279685165733099,
0.027106579393148422,
-0.004529056139290333,
0.012379292398691177,
0.0697367936372757,
0.04759515821933746,
-0.0012821671552956104,
-0.02481963112950325,
-0.00777990510687232,
-0.04114542156457901,
0.005562571808695793,
0.09828677028417587,
0.06769440323114395,
0.034349825233221054,
0.02258887328207493,
-0.009092618711292744,
-0.02865973673760891,
0.0456911101937294,
-0.008279519155621529,
0.11415329575538635,
0.04350494593381882,
-0.09144066274166107,
0.021214954555034637,
-0.03488712012767792,
0.10225842148065567,
-0.012926463969051838,
-0.08667055517435074,
0.0636795237660408,
0.016394339501857758,
-0.013547501526772976,
0.020955562591552734,
-0.057250045239925385,
0.030960189178586006,
-0.04813466966152191,
-0.012251130305230618,
-0.05101761966943741,
-0.01559642143547535,
0.04048973321914673,
0.04037351533770561,
0.035979531705379486,
0.016899190843105316,
-0.024090619757771492,
-0.0062745604664087296,
0.04986724257469177,
0.0414968878030777,
0.055916182696819305,
-0.021333741024136543,
0.013215096667408943
] |
Improving Equivariant Networks with Probabilistic Symmetry Breaking | https://openreview.net/forum?id=ZE6lrLvATd | [
"Hannah Lawrence",
"Vasco Portilheiro",
"Yan Zhang",
"Sékou-Oumar Kaba"
] | Poster | Equivariance encodes known symmetries into neural networks, often enhancing generalization. However, equivariant networks cannot *break* symmetries: the output of an equivariant network must, by definition, have at least the same self-symmetries as its input. This poses an important problem, both (1) for prediction tasks on domains where self-symmetries are common, and (2) for generative models, which must break symmetries in order to reconstruct from highly symmetric latent spaces. This fundamental limitation can in fact be addressed by considering *equivariant conditional distributions*, instead of equivariant functions. We therefore present novel theoretical results that establish necessary and sufficient conditions for representing such distributions. Concretely, this representation provides a practical framework for breaking symmetries in any equivariant network via randomized canonicalization. Our method, SymPE (Symmetry-breaking Positional Encodings), admits a simple interpretation in terms of positional encodings. This approach expands the representational power of equivariant networks while retaining the inductive bias of symmetry, which we justify through generalization bounds. Experimental results demonstrate that SymPE significantly improves performance of group-equivariant and graph neural networks across diffusion models for graphs, graph autoencoders, and lattice spin system modeling. | equivariance, symmetry, symmetry-breaking, canonicalization, graphs, GNNs | We propose a probabilistic framework for breaking symmetries, e.g. in generative models' latent spaces, by combining equivariant networks with canonicalization. | 1,165 | null | [
-0.09765928238630295,
-0.12213297933340073,
0.13125823438167572,
-0.05558568984270096,
0.054771263152360916,
0.004522357601672411,
-0.010151120834052563,
-0.07331866025924683,
0.0458790622651577,
-0.03655317798256874,
0.019047731533646584,
-0.04605943337082863,
-0.008573375642299652,
0.053859878331422806,
0.027713634073734283,
-0.04200247302651405,
0.019738363102078438,
0.10648473352193832,
-0.07124186307191849,
-0.012528959661722183,
0.03476537764072418,
0.014838415198028088,
-0.018708495423197746,
0.018773334100842476,
0.08216385543346405,
-0.024848761036992073,
0.07885482162237167,
0.01299461629241705,
-0.04628225415945053,
-0.0879150852560997,
0.063470259308815,
-0.004190725274384022,
-0.07550683617591858,
-0.04024279862642288,
-0.06569807231426239,
0.0391010157763958,
-0.053643856197595596,
-0.00018167542293667793,
0.04814647510647774,
-0.05487743765115738,
0.003506183624267578,
0.007399768568575382,
0.02623715251684189,
0.02513849176466465,
0.06061962619423866,
0.08749313652515411,
0.07558849453926086,
-0.04518232122063637,
-0.012035112828016281,
0.013078166171908379,
-0.006793174892663956,
0.0719408392906189,
-0.09533826261758804,
0.01698809303343296,
0.04195219278335571,
-0.03895242512226105,
0.010672777891159058,
-0.023195641115307808,
-0.04612388089299202,
-0.02681511454284191,
0.0015728017315268517,
-0.040958505123853683,
-0.04414509981870651,
0.011195734143257141,
0.01256309449672699,
0.002978420350700617,
0.03181285038590431,
0.09829167276620865,
0.021976210176944733,
-0.001684498623944819,
0.001681400346569717,
0.04617488756775856,
-0.14878173172473907,
0.09169293195009232,
0.04879332706332207,
0.12709249556064606,
0.07278912514448166,
-0.029941407963633537,
-0.0031787469051778316,
-0.0320785716176033,
-0.037694867700338364,
0.05897924676537514,
0.10671374201774597,
0.014442967250943184,
0.0864337906241417,
0.009378891438245773,
-0.06529390066862106,
0.02706720679998398,
0.05511292815208435,
-0.03750960901379585,
-0.05303233116865158,
-0.004190871026366949,
0.019870776683092117,
-0.02152910642325878,
0.08683524280786514,
0.01499936729669571,
0.07030342519283295,
-0.0781358927488327,
-0.00019318917475175112,
0.09872240573167801,
-0.028262563049793243,
-0.011248795315623283,
0.07631894201040268,
-0.00828003603965044,
0.12084053456783295,
-0.05632740259170532,
0.00411247881129384,
-0.019232075661420822,
0.03815917670726776,
-0.164340540766716,
-0.04566194862127304,
-0.030110258609056473,
-0.027815518900752068,
-0.023774029687047005,
-0.06697626411914825,
-0.019091378897428513,
-0.05250013247132301,
0.025164734572172165,
0.02912980504333973,
-0.008308480493724346,
-0.06021108105778694,
0.020834052935242653,
-0.04806642606854439,
0.026713598519563675,
0.030667956918478012,
0.0034972799476236105,
-0.07599447667598724,
5.245333136567444e-33,
0.01072360947728157,
0.02108912542462349,
0.036990921944379807,
0.04871810972690582,
0.06763115525245667,
0.07432176172733307,
0.01562114991247654,
0.005981373135000467,
0.07971351593732834,
0.02473287470638752,
-0.09067363291978836,
-0.026550080627202988,
-0.02760375663638115,
0.00909327995032072,
-0.006592517253011465,
0.07297766208648682,
-0.034517765045166016,
-0.024656329303979874,
-0.0016218911623582244,
-0.0566277913749218,
0.019269458949565887,
0.01569056324660778,
-0.060496117919683456,
0.04138335958123207,
-0.08437243849039078,
0.01046669576317072,
0.031421959400177,
-0.01693931221961975,
-0.10565181821584702,
-0.008793083019554615,
-0.03206692636013031,
0.041478198021650314,
-0.014789508655667305,
0.05095263198018074,
0.021628130227327347,
-0.07199986279010773,
0.012794439680874348,
-0.0769088938832283,
0.03919393941760063,
-0.04146561771631241,
0.006310392636805773,
0.02915884740650654,
-0.05413553863763809,
-0.01988578587770462,
-0.029172243550419807,
-0.06708443909883499,
0.07015503942966461,
0.023608578369021416,
-0.011378496885299683,
-0.06618542969226837,
0.05799424648284912,
0.05950414389371872,
0.032507117837667465,
-0.05639324337244034,
-0.028700759634375572,
0.04697684198617935,
0.08880797773599625,
0.03254597634077072,
0.017050044611096382,
0.02826242335140705,
-0.04045448824763298,
0.06397607177495956,
0.02150084264576435,
-0.05506708845496178,
0.023655330762267113,
0.03520391881465912,
-0.009016374126076698,
-0.032809577882289886,
0.1511651575565338,
-0.023717939853668213,
-0.006266674026846886,
0.03251224011182785,
-0.1370609700679779,
-0.06404336541891098,
0.04691212624311447,
0.03707859665155411,
-0.05028174817562103,
-0.10023488104343414,
0.020578421652317047,
0.030537128448486328,
-0.044761814177036285,
0.058328475803136826,
-0.08726368844509125,
0.002343851374462247,
-0.0008082716958597302,
0.024629002436995506,
0.046647194772958755,
-0.13423795998096466,
0.0078662671148777,
-0.06121816113591194,
-0.02001374401152134,
-0.09308203309774399,
-0.02856810763478279,
-0.0035425785463303328,
0.018795669078826904,
-6.360124591771284e-33,
-0.1013241782784462,
-0.0007971872109919786,
-0.027080660685896873,
0.04969450458884239,
0.006472059991210699,
-0.07447382062673569,
-0.0237258430570364,
0.0685981884598732,
-0.02148357778787613,
0.01628013141453266,
0.004978170618414879,
-0.021736087277531624,
-0.03419969230890274,
-0.01969071850180626,
0.017346486449241638,
0.04586338251829147,
-0.02961195446550846,
0.00499750766903162,
-0.027655594050884247,
0.019793109968304634,
-0.0023768609389662743,
0.07084953784942627,
-0.09235191345214844,
0.04868439957499504,
0.0676625519990921,
0.03767489641904831,
-0.035105928778648376,
0.014778059907257557,
0.07760590314865112,
0.019299115985631943,
-0.06497947126626968,
-0.036398883908987045,
-0.040335915982723236,
0.035241466015577316,
-0.00819002091884613,
0.025463668629527092,
-0.00590800866484642,
0.036591172218322754,
-0.07467452436685562,
0.07311670482158661,
0.02017739601433277,
-0.040802840143442154,
-0.044039350003004074,
0.060841649770736694,
0.014205857180058956,
-0.05462871864438057,
0.03447955846786499,
0.02846679463982582,
-0.011946373619139194,
0.04303186759352684,
-0.0014694920973852277,
0.06605889648199081,
-0.06295114010572433,
0.008411851711571217,
-0.02710852399468422,
-0.009134652093052864,
-0.02790742740035057,
-0.07382973283529282,
0.08130238205194473,
0.01632947474718094,
-0.07087422162294388,
-0.1075802743434906,
0.07055346667766571,
0.0123497499153018,
-0.043958455324172974,
-0.04136288911104202,
-0.031049193814396858,
-0.001212587347254157,
0.08505821973085403,
0.03768870234489441,
0.04276515915989876,
-0.01867668330669403,
-0.015446988865733147,
-0.028425633907318115,
0.023808086290955544,
-0.01606515608727932,
0.05942029133439064,
-0.01070362888276577,
0.03926496580243111,
-0.08163284510374069,
-0.02210550382733345,
-0.07032909244298935,
0.05401257798075676,
0.003937748726457357,
0.05740603059530258,
-0.010093288496136665,
0.05581819638609886,
0.01043158583343029,
0.10659386217594147,
-0.027750572189688683,
-0.0023670729715377092,
0.06573674827814102,
0.02628341317176819,
0.11366204172372818,
-0.015322186052799225,
-4.5521755254185337e-8,
-0.044333357363939285,
0.03124847635626793,
0.004349245689809322,
0.047471653670072556,
0.03101937845349312,
-0.00914017204195261,
0.01604151539504528,
0.038944877684116364,
-0.03172975406050682,
0.03345667943358421,
0.06011077016592026,
-0.0020932487677782774,
-0.060023244470357895,
-0.06550297886133194,
-0.027408812195062637,
0.035248711705207825,
0.03448433056473732,
0.1029462218284607,
-0.023473799228668213,
0.031317971646785736,
0.007939670234918594,
-0.004771044012159109,
-0.04557222127914429,
-0.03608148917555809,
0.07120922207832336,
-0.0815410315990448,
-0.1005067229270935,
-0.02726072072982788,
0.055783629417419434,
0.05245940759778023,
-0.027764612808823586,
-0.028641218319535255,
0.041186682879924774,
-0.04199103266000748,
-0.03527229279279709,
0.06317933648824692,
-0.007085380144417286,
-0.038727182894945145,
-0.00712954206392169,
-0.08593206107616425,
0.003643433563411236,
0.02034437656402588,
-0.04024878516793251,
0.05413435399532318,
0.051452383399009705,
0.027282875031232834,
0.09100685268640518,
-0.043549682945013046,
0.011026625521481037,
-0.020411740988492966,
-0.05661049857735634,
-0.02858871966600418,
-0.027612878009676933,
-0.009126530960202217,
-0.038800787180662155,
0.047456320375204086,
0.020017534494400024,
-0.03962040692567825,
0.04331204295158386,
0.008955234661698341,
-0.04238187521696091,
0.03651631996035576,
-0.07168794423341751,
-0.06284331530332565
] |
CraftRTL: High-quality Synthetic Data Generation for Verilog Code Models with Correct-by-Construction Non-Textual Representations and Targeted Code Repair | https://openreview.net/forum?id=8KQzoD5XAr | [
"Mingjie Liu",
"Yun-Da Tsai",
"Wenfei Zhou",
"Haoxing Ren"
] | Poster | Despite the significant progress made in code generation with large language models, challenges persist, especially with hardware description languages such as Verilog. This paper first presents an analysis of fine-tuned LLMs on Verilog coding, with synthetic data from prior methods. We identify two main issues: difficulties in handling non-textual representations (Karnaugh maps, state-transition diagrams and waveforms) and significant variability during training with models randomly making ''minor'' mistakes. To address these limitations, we enhance data curation by creating correct-by-construction data targeting non-textual representations. Additionally, we introduce an automated framework that generates error reports from various model checkpoints and injects these errors into open-source code to create targeted code repair data. Our fine-tuned Starcoder2-15B outperforms prior state-of-the-art results by 3.8\%, 10.9\%, 6.6\% for pass@1 on VerilogEval-Machine, VerilogEval-Human, and RTLLM. | Verilog Code Generation, Synthetic Data Generation, Large Language Models | We produce high-quality Verilog finetuning data that is correct-by-construction for non-textual representations and develop targeted code repair data by injecting errors into open-source code. | 1,163 | 2409.12993 | [
-0.039734311401844025,
-0.02610277384519577,
0.03504755347967148,
0.010208540596067905,
0.015162136405706406,
-0.06754611432552338,
-0.06546822190284729,
-0.029567668214440346,
-0.06350983679294586,
-0.03424399718642235,
-0.05216998606920242,
-0.04765085503458977,
0.032928112894296646,
-0.04563668370246887,
0.0037306114099919796,
0.012958663515746593,
-0.0031500360928475857,
0.08772046864032745,
-0.010821448639035225,
-0.0874987468123436,
-0.0007072772132232785,
0.07321011275053024,
0.0016286439495161176,
0.04515508934855461,
0.061008621007204056,
0.016515186056494713,
0.016739938408136368,
0.005148841999471188,
0.03485921025276184,
-0.05113963037729263,
-0.02120574750006199,
0.08119452744722366,
-0.04959065839648247,
0.04818546399474144,
0.020099394023418427,
0.08259093016386032,
-0.08661536127328873,
0.002358665456995368,
-0.04031975939869881,
-0.05349424108862877,
-0.05488535761833191,
0.0100408885627985,
0.014338006265461445,
0.07888633757829666,
0.04545818641781807,
-0.03329020366072655,
-0.011692643165588379,
-0.06728842109441757,
-0.03422894328832626,
0.0007373003172688186,
-0.12164723873138428,
0.004725185688585043,
0.016834678128361702,
-0.0004806767392437905,
0.032183196395635605,
-0.02278921939432621,
0.10715243220329285,
-0.006917341612279415,
0.027446290478110313,
-0.04260111227631569,
-0.017430484294891357,
-0.04466165602207184,
-0.015610836446285248,
-0.01880110800266266,
-0.053172413259744644,
0.0036911580245941877,
0.027262674644589424,
0.09720862656831741,
0.037054676562547684,
0.006613434292376041,
-0.04713459312915802,
0.04513494670391083,
-0.007167087402194738,
0.09571930021047592,
-0.011275422759354115,
0.05234607681632042,
0.04734620824456215,
0.01439130399376154,
0.07933144271373749,
-0.09270107001066208,
-0.04548962786793709,
-0.030857061967253685,
-0.019112881273031235,
0.03913459926843643,
0.08083945512771606,
0.03567267954349518,
0.05044589191675186,
0.08309878408908844,
0.08087319135665894,
0.0741889476776123,
-0.03866202011704445,
-0.050097763538360596,
0.06814887374639511,
0.006915548350661993,
0.0419393815100193,
0.043662089854478836,
0.015061993151903152,
-0.09377622604370117,
0.01931651495397091,
0.050049033015966415,
-0.027113281190395355,
0.11496321111917496,
0.06303318589925766,
-0.06487032771110535,
-0.006285531911998987,
0.008172726258635521,
0.0704549178481102,
0.019665196537971497,
-0.032307639718055725,
-0.07976581901311874,
0.03652101010084152,
0.043396975845098495,
0.011933865025639534,
-0.026156499981880188,
0.03509029373526573,
-0.012392183765769005,
-0.01392122358083725,
-0.015958262607455254,
-0.03958006948232651,
0.06562110036611557,
-0.01588025502860546,
0.0016933418810367584,
-0.020085284486413002,
0.07358045876026154,
0.01089561264961958,
-0.045345570892095566,
-0.03022117353975773,
6.179666096946404e-34,
0.04876803606748581,
0.027858899906277657,
-0.02790803276002407,
0.08221398293972015,
0.07132420688867569,
-0.0555986724793911,
0.01285607647150755,
0.05750332400202751,
-0.046901166439056396,
0.018583018332719803,
0.022224200889468193,
-0.0508047416806221,
-0.052138350903987885,
0.08169133961200714,
0.045803479850292206,
0.0139921298250556,
-0.060116082429885864,
-0.030020952224731445,
-0.11999144405126572,
0.04959112033247948,
-0.0022367602214217186,
-0.023987019434571266,
-0.022678455337882042,
-0.060708388686180115,
0.04781733825802803,
0.093756765127182,
0.040189892053604126,
-0.025451160967350006,
-0.011180154979228973,
0.022988352924585342,
-0.08277378231287003,
-0.04302658140659332,
0.1027546152472496,
0.05427665263414383,
0.028941845521330833,
-0.034738440066576004,
-0.04618149623274803,
-0.023181429132819176,
0.0073270672000944614,
0.03588174283504486,
0.035204410552978516,
0.022516250610351562,
-0.0021609507966786623,
-0.011009041219949722,
0.020891821011900902,
-0.11468219012022018,
-0.028872624039649963,
0.026061024516820908,
-0.004046241287142038,
-0.0595548078417778,
0.06584368646144867,
0.09701882302761078,
-0.03073912486433983,
-0.02173008769750595,
-0.029743677005171776,
0.002452115761116147,
0.019707508385181427,
0.02091730199754238,
-0.02077547460794449,
0.10203246027231216,
0.0011355876922607422,
0.04886738955974579,
0.006056233309209347,
0.029964976012706757,
0.03295950964093208,
0.016363194212317467,
-0.01092025451362133,
0.025665707886219025,
0.013138961046934128,
-0.025095943361520767,
-0.05640019103884697,
0.02061454765498638,
-0.0441640168428421,
0.05521942302584648,
0.008804267272353172,
-0.014786852523684502,
0.04782911390066147,
-0.06076396256685257,
-0.036091335117816925,
-0.04308583214879036,
0.030056912451982498,
0.003949769772589207,
-0.031107455492019653,
-0.0021821812260895967,
-0.03829718753695488,
-0.12159761041402817,
-0.0005598980933427811,
-0.045396625995635986,
-0.03164398670196533,
0.03546859323978424,
-0.025731662288308144,
-0.12082052230834961,
-0.012780340388417244,
-0.01925642229616642,
-0.012266624718904495,
-1.5608425211859497e-33,
-0.08799939602613449,
0.005979140754789114,
0.019442452117800713,
0.09545262157917023,
-0.09476184844970703,
-0.11118482053279877,
0.0039054169319570065,
0.008495016023516655,
0.041245464235544205,
-0.00015768903540447354,
-0.07127707451581955,
-0.05014364793896675,
0.015119180083274841,
-0.026791440322995186,
0.07178964465856552,
-0.02909206412732601,
-0.08632780611515045,
0.0096131581813097,
-0.012201623059809208,
0.061824701726436615,
0.04332702234387398,
0.13631467521190643,
-0.04778821021318436,
0.028660282492637634,
-0.025297537446022034,
0.08264704048633575,
-0.025903521105647087,
0.04380768910050392,
0.02863570675253868,
-0.022771093994379044,
0.052991047501564026,
-0.015022422187030315,
0.004401346202939749,
-0.009904596954584122,
-0.017524167895317078,
-0.0032856306061148643,
0.1083853542804718,
-0.02563616819679737,
0.0032282297033816576,
0.018442632630467415,
0.11230069398880005,
0.03621933236718178,
-0.10116135329008102,
0.016777198761701584,
-0.05104098469018936,
0.0021082081366330385,
-0.05939198657870293,
-0.02364722453057766,
0.025070561096072197,
-0.07027987390756607,
-0.01364881731569767,
-0.0038241432048380375,
-0.007987807504832745,
-0.045788656920194626,
-0.003154531354084611,
-0.0693407729268074,
-0.03585319593548775,
-0.03910725191235542,
0.0520741231739521,
0.01932601071894169,
-0.1041148453950882,
-0.07117273658514023,
0.11504699289798737,
-0.08423101156949997,
0.011541086249053478,
-0.08189968019723892,
-0.0247876588255167,
0.04103383421897888,
-0.022722333669662476,
-0.036664895713329315,
0.07213771343231201,
-0.01592002995312214,
-0.11810700595378876,
0.03357218950986862,
0.015508752316236496,
-0.02269790507853031,
-0.022940633818507195,
-0.012689761817455292,
0.009851684793829918,
-0.006848021876066923,
0.06554931402206421,
-0.028111563995480537,
0.03619430959224701,
0.10396687686443329,
0.02173059992492199,
0.05777058005332947,
0.08185078948736191,
0.09041725844144821,
-0.013626743108034134,
-0.01694668084383011,
-0.09816339612007141,
0.08116022497415543,
-0.013519142754375935,
0.059931956231594086,
-0.01816144585609436,
-4.2684767720402306e-8,
-0.024755313992500305,
0.0016496015014126897,
-0.04686833545565605,
0.0977722629904747,
0.021930180490016937,
-0.03973335772752762,
-0.08476965129375458,
0.10682430863380432,
-0.02527855522930622,
-0.007267723325639963,
0.014680826105177402,
0.01553700678050518,
-0.023977098986506462,
-0.13283000886440277,
0.0012988477246835828,
0.0908181294798851,
-0.05141497030854225,
0.005183648318052292,
-0.02133924327790737,
0.004936766345053911,
0.030922824516892433,
0.13060002028942108,
-0.009464528411626816,
0.005176004022359848,
-0.018349185585975647,
-0.06132461875677109,
0.002828967524692416,
0.07886587083339691,
0.07020611315965652,
-0.044097650796175,
0.03626151382923126,
-0.015547388233244419,
0.06096455827355385,
-0.0122941629961133,
-0.01017390750348568,
-0.0025740868877619505,
0.08909375965595245,
-0.014396515674889088,
0.07168998569250107,
-0.009456646628677845,
0.052017729729413986,
0.0853717029094696,
-0.09471873939037323,
0.00931281317025423,
-0.06514674425125122,
-0.04633506387472153,
-0.04785951226949692,
-0.10615484416484833,
-0.034687645733356476,
-0.06814854592084885,
-0.012618405744433403,
0.019432561472058296,
-0.06425710767507553,
0.011167935095727444,
-0.024510914459824562,
0.05193189159035683,
-0.058931928128004074,
-0.0015261408407241106,
0.06427601724863052,
0.04486655071377754,
-0.04657260701060295,
0.024774689227342606,
0.005390464793890715,
-0.015197222121059895
] |
FreSh: Frequency Shifting for Accelerated Neural Representation Learning | https://openreview.net/forum?id=zMjjzXxS64 | [
"Adam Kania",
"Marko Mihajlovic",
"Sergey Prokudin",
"Jacek Tabor",
"Przemysław Spurek"
] | Poster | Implicit Neural Representations (INRs) have recently gained attention as a powerful approach for continuously representing signals such as images, videos, and 3D shapes using multilayer perceptrons (MLPs). However, MLPs are known to exhibit a low-frequency bias, limiting their ability to capture high-frequency details accurately. This limitation is typically addressed by incorporating high-frequency input embeddings or specialized activation layers. In this work, we demonstrate that these embeddings and activations are often configured with hyperparameters that perform well on average but are suboptimal for specific input signals under consideration, necessitating a costly grid search to identify optimal settings. Our key observation is that the initial frequency spectrum of an untrained model's output correlates strongly with the model's eventual performance on a given target signal. Leveraging this insight, we propose frequency shifting (or FreSh), a method that selects embedding hyperparameters to align the frequency spectrum of the model’s initial output with that of the target signal. We show that this simple initialization technique improves performance across various neural representation methods and tasks, achieving results comparable to extensive hyperparameter sweeps but with only marginal computational overhead compared to training a single model with default hyperparameters. | spectral bias, automatic hyperparameter selection, implicit neural representation, discrete fourier transform | null | 1,157 | 2410.05050 | [
-0.04054093733429909,
-0.1813431978225708,
0.030652526766061783,
0.062289126217365265,
0.024425067007541656,
0.07708471268415451,
-0.045708365738391876,
-0.09331560134887695,
0.004842109512537718,
-0.12046198546886444,
-0.06901390850543976,
-0.05153976380825043,
-0.046225596219301224,
0.05325739085674286,
-0.037092216312885284,
0.0005135981482453644,
0.1577937752008438,
0.08335913717746735,
-0.03139190375804901,
-0.00800644513219595,
0.0151539146900177,
-0.08046187460422516,
0.016771160066127777,
-0.03501518443226814,
0.07607921957969666,
-0.007944628596305847,
-0.008926969952881336,
-0.0009326412691734731,
0.054735004901885986,
-0.011241331696510315,
0.03948858752846718,
-0.05151822045445442,
0.0010610900353640318,
0.07235828787088394,
-0.06125734746456146,
0.03257210925221443,
-0.014613190665841103,
0.044643521308898926,
-0.023824915289878845,
-0.06814329326152802,
0.055932383984327316,
-0.011227338574826717,
-0.01058895979076624,
0.05414971336722374,
0.03150317072868347,
0.030780786648392677,
0.043685223907232285,
-0.06695344299077988,
-0.025526558980345726,
0.00013894724543206394,
-0.021083101630210876,
-0.0028911172412335873,
-0.012145321816205978,
0.033410411328077316,
-0.045427411794662476,
0.025542961433529854,
0.051230281591415405,
0.018421374261379242,
-0.00024344702251255512,
0.0022319937124848366,
-0.046429507434368134,
-0.053346917033195496,
0.02847895212471485,
-0.013910137116909027,
-0.015676898881793022,
-0.021594202145934105,
0.06833936274051666,
0.007354489527642727,
-0.00017361454956699163,
0.025883641093969345,
-0.03437333181500435,
0.11830094456672668,
-0.06558260321617126,
-0.01383870467543602,
0.024598076939582825,
0.0014054026687517762,
0.012463562190532684,
-0.01761075109243393,
0.08356377482414246,
-0.07947935909032822,
0.04993514344096184,
-0.007220545317977667,
-0.010554908774793148,
-0.029308931902050972,
0.12981033325195312,
0.005308300722390413,
-0.0231193657964468,
0.023406658321619034,
-0.00008632004028186202,
-0.036726560443639755,
-0.08804532885551453,
-0.02691039815545082,
-0.007575931493192911,
-0.044398535043001175,
0.021300900727510452,
0.060906555503606796,
0.029229482635855675,
-0.08130069077014923,
-0.011219736188650131,
0.08707518875598907,
-0.01529746875166893,
0.014648069627583027,
0.02559572644531727,
0.02678576670587063,
0.022786540910601616,
-0.041973017156124115,
0.02343539148569107,
0.08730617165565491,
-0.0018910204526036978,
-0.06685587763786316,
0.06376569718122482,
0.013232676312327385,
-0.014082422479987144,
0.03768719732761383,
0.06878077983856201,
0.012821744196116924,
0.016152003780007362,
-0.03236031159758568,
0.02853565290570259,
0.05244677886366844,
-0.07862376421689987,
-0.006674155592918396,
-0.06303788721561432,
0.06287915259599686,
0.020430240780115128,
-0.00909800361841917,
-0.08215317130088806,
2.8029029130433955e-33,
-0.03191455453634262,
0.07894997298717499,
0.038177549839019775,
-0.008937343955039978,
0.010168449021875858,
0.026924097910523415,
0.056498777121305466,
0.08376675844192505,
0.09532430022954941,
0.07195651531219482,
-0.09734418988227844,
-0.03429313004016876,
-0.01983761601150036,
0.14874601364135742,
0.0005739727639593184,
-0.033980801701545715,
-0.0762130543589592,
0.05477320775389671,
-0.04286015033721924,
-0.034219540655612946,
0.016413070261478424,
-0.0013278544647619128,
-0.0028514538425952196,
-0.05278504267334938,
-0.07259590178728104,
0.030889155343174934,
0.037599317729473114,
-0.055035945028066635,
-0.0738753005862236,
-0.009155225940048695,
-0.053305406123399734,
-0.006330754142254591,
0.005745361093431711,
-0.016057666391134262,
0.03554413095116615,
-0.03762078285217285,
0.04209505394101143,
0.02657354436814785,
0.05648898333311081,
0.02327387034893036,
0.0036366344429552555,
0.024257315322756767,
0.011247151531279087,
-0.024502255022525787,
-0.06888370215892792,
-0.052487608045339584,
-0.03772357106208801,
0.0864325538277626,
-0.06353732943534851,
-0.031275343149900436,
0.051962822675704956,
0.06338513642549515,
-0.022671349346637726,
-0.12495286017656326,
-0.01573874242603779,
-0.0008185486076399684,
0.030558474361896515,
0.04854467883706093,
0.03357419744133949,
0.018103893846273422,
-0.0007680224953219295,
0.01429993100464344,
-0.027815144509077072,
0.09640643000602722,
0.020638488233089447,
-0.0070653799921274185,
0.02979179285466671,
0.1051040068268776,
0.01331543643027544,
0.08055902272462845,
0.008937695994973183,
0.10253260284662247,
-0.04777584969997406,
-0.09895729273557663,
0.02317117713391781,
0.012793796136975288,
0.010518014430999756,
-0.0570407509803772,
-0.037343770265579224,
0.03512338921427727,
0.0689144879579544,
0.04044107347726822,
-0.018030337989330292,
-0.035613544285297394,
-0.004456894937902689,
-0.017788201570510864,
0.01922403648495674,
-0.07716100662946701,
0.0635298639535904,
-0.07362323999404907,
-0.016495367512106895,
-0.003329142229631543,
-0.026206359267234802,
-0.059079982340335846,
-0.0007713439408689737,
-1.9467788060670076e-33,
0.053155314177274704,
0.06753155589103699,
-0.04484046623110771,
0.020349424332380295,
0.019680893048644066,
0.012555654160678387,
-0.018649041652679443,
0.05153343826532364,
-0.030791692435741425,
-0.021483389660716057,
-0.023998498916625977,
-0.07425349205732346,
-0.024420224130153656,
-0.03110099583864212,
-0.03602322190999985,
-0.02387331984937191,
-0.09772218763828278,
0.03825477883219719,
0.02581939287483692,
-0.03088337928056717,
-0.013862687163054943,
0.0008490277105011046,
-0.07576331496238708,
0.12290287017822266,
-0.041823286563158035,
0.037881039083004,
0.01674872823059559,
0.0651630312204361,
-0.018436096608638763,
0.012686644680798054,
-0.08230292052030563,
0.01559020858258009,
0.0420004166662693,
-0.04428383335471153,
0.008750546723604202,
0.06267297267913818,
0.04635120928287506,
-0.03597709909081459,
-0.0354611873626709,
0.07208558171987534,
0.06245236098766327,
0.005786466877907515,
-0.0223054438829422,
0.05377212166786194,
0.0046085091307759285,
0.007992556318640709,
-0.05860016494989395,
0.007590023335069418,
-0.03988615795969963,
0.01902209408581257,
-0.04012858495116234,
-0.0066057899966835976,
-0.06105652451515198,
-0.014862116426229477,
-0.028497446328401566,
-0.03587048128247261,
-0.07435452193021774,
-0.024398481473326683,
0.14377689361572266,
0.0021258238703012466,
-0.07082845270633698,
-0.02580074779689312,
-0.05694463849067688,
-0.09637582302093506,
0.004038603976368904,
-0.008727009408175945,
-0.010474988259375095,
-0.043426502496004105,
0.024852994829416275,
0.016856031492352486,
-0.03146643191576004,
-0.0654042437672615,
0.018432240933179855,
-0.03996450826525688,
-0.03918243199586868,
-0.0535842590034008,
0.052878487855196,
-0.05722380429506302,
-0.03571826219558716,
-0.037770189344882965,
-0.011042128317058086,
-0.06191054359078407,
-0.016539176926016808,
0.04542762041091919,
0.05906218662858009,
0.11605414003133774,
0.004761190619319677,
0.05603357031941414,
0.02699778787791729,
-0.026591520756483078,
-0.02402246929705143,
0.07705505192279816,
-0.056502971798181534,
0.0401705764234066,
0.014016841538250446,
-4.8718231226985154e-8,
-0.0794091448187828,
0.11787490546703339,
0.023681027814745903,
0.008939548395574093,
0.04541565850377083,
-0.024422362446784973,
0.06053410843014717,
0.06070147827267647,
-0.02406683750450611,
-0.09904389083385468,
0.09707919508218765,
-0.045395657420158386,
0.014397312887012959,
-0.03718584403395653,
0.04062819853425026,
0.05937960371375084,
0.04424180090427399,
0.01461031287908554,
0.000039880338590592146,
-0.00591752864420414,
-0.048407360911369324,
0.12064484506845474,
0.04963119700551033,
-0.07716143131256104,
0.01450948417186737,
-0.06592609733343124,
-0.062243856489658356,
0.13799265027046204,
0.10440022498369217,
0.017834559082984924,
-0.028587326407432556,
0.04453444108366966,
0.0659843385219574,
-0.0515868216753006,
-0.029843028634786606,
0.05998030677437782,
0.022274447605013847,
0.006476530339568853,
-0.019030334427952766,
-0.05847562476992607,
-0.04378381371498108,
0.10002420097589493,
-0.04178217053413391,
-0.0377434641122818,
-0.04865284264087677,
-0.009764834307134151,
0.023192746564745903,
-0.0681210458278656,
0.019933752715587616,
-0.008824308402836323,
-0.001907020341604948,
-0.027013378217816353,
0.00895992573350668,
0.09828196465969086,
0.0021895417012274265,
0.01735408417880535,
-0.05106087401509285,
-0.1004948616027832,
-0.033416248857975006,
0.052482765167951584,
0.023997198790311813,
0.02637760154902935,
-0.04562665894627571,
0.033318933099508286
] |
PointOBB-v2: Towards Simpler, Faster, and Stronger Single Point Supervised Oriented Object Detection | https://openreview.net/forum?id=R22JPTQYWV | [
"Botao Ren",
"Xue Yang",
"Yi Yu",
"Junwei Luo",
"Zhidong Deng"
] | Poster | Single point supervised oriented object detection has gained attention and made initial progress within the community. Diverse from those approaches relying on one-shot samples or powerful pretrained models (e.g. SAM), PointOBB has shown promise due to its prior-free feature. In this paper, we propose PointOBB-v2, a simpler, faster, and stronger method to generate pseudo rotated boxes from points without relying on any other prior. Specifically, we first generate a Class Probability Map (CPM) by training the network with non-uniform positive and negative sampling. We show that the CPM is able to learn the approximate object regions and their contours. Then, Principal Component Analysis (PCA) is applied to accurately estimate the orientation and the boundary of objects. By further incorporating a separation mechanism, we resolve the confusion caused by the overlapping on the CPM, enabling its operation in high-density scenarios. Extensive comparisons demonstrate that our method achieves a training speed 15.58$\times$ faster and an accuracy improvement of 11.60\%/25.15\%/21.19\% on the DOTA-v1.0/v1.5/v2.0 datasets compared to the previous state-of-the-art, PointOBB. This significantly advances the cutting edge of single point supervised oriented detection in the modular track. Code and models will be released. | Oriented Object Detection, Point Supervised Object Detection | A simpler, faster, and stronger single point supervised oriented object detection method. | 1,151 | null | [
0.020624585449695587,
-0.040313299745321274,
0.007303610444068909,
-0.02619810588657856,
0.06989613175392151,
0.0036316183395683765,
0.00365283596329391,
-0.0892966017127037,
0.01791595108807087,
-0.0046949791721999645,
-0.005700947251170874,
-0.0019143319223076105,
-0.010211966000497341,
0.0577709823846817,
0.040862418711185455,
-0.036341577768325806,
0.1240171268582344,
0.03083798475563526,
-0.014753631316125393,
0.07992604374885559,
-0.06995794922113419,
0.0011094147339463234,
0.010915339924395084,
0.02205228991806507,
0.0054076118394732475,
0.05585889890789986,
0.07248080521821976,
-0.04819401353597641,
0.014207073487341404,
-0.0896265059709549,
0.07926588505506516,
-0.027547897771000862,
-0.015386403538286686,
0.04713432118296623,
-0.035169947892427444,
-0.037281304597854614,
-0.014631876721978188,
0.04007851704955101,
0.019059233367443085,
-0.02267071232199669,
0.06926600635051727,
0.10316293686628342,
-0.01157174538820982,
0.0420212484896183,
0.11198998242616653,
0.0730462446808815,
0.01596835069358349,
-0.07865230739116669,
0.0864345133304596,
-0.013692399486899376,
-0.017365410923957825,
-0.042941026389598846,
-0.022990994155406952,
0.025798866525292397,
-0.05840341001749039,
0.056563954800367355,
0.04879029095172882,
-0.056850068271160126,
0.04772965610027313,
-0.021259550005197525,
-0.01463573332875967,
-0.06155848130583763,
0.06302201747894287,
-0.03140254318714142,
0.0011774192098528147,
-0.09225006401538849,
0.014986446127295494,
-0.02244751527905464,
0.031231023371219635,
-0.01542598381638527,
0.007367903832346201,
0.06925396621227264,
0.06927834451198578,
0.025349251925945282,
0.014787685126066208,
0.022061333060264587,
-0.06360585987567902,
-0.021915193647146225,
0.02357598766684532,
-0.07674628496170044,
-0.04125590994954109,
-0.02274613454937935,
0.08439663797616959,
0.03848065435886383,
0.05572328343987465,
0.02963455766439438,
-0.0926850363612175,
0.054309483617544174,
-0.06753315776586533,
0.022268028929829597,
-0.043018389493227005,
0.006795679684728384,
0.0015661465004086494,
-0.09940602630376816,
-0.03552558273077011,
-0.051668934524059296,
-0.009164047427475452,
-0.06773079186677933,
-0.03395678102970123,
0.0702064260840416,
-0.003062654286623001,
0.011800398118793964,
0.01521893497556448,
-0.06291744112968445,
0.07016229629516602,
0.01287791971117258,
-0.05528486892580986,
-0.01107066310942173,
0.12952940165996552,
0.004139898810535669,
-0.0008753807051107287,
-0.05861600860953331,
-0.031488317996263504,
-0.01813359744846821,
-0.02552046626806259,
-0.019255338236689568,
-0.034292273223400116,
0.0014975983649492264,
0.020890897139906883,
-0.004675987176597118,
-0.03850652649998665,
-0.0660003274679184,
-0.00691505242139101,
0.028157055377960205,
0.04145007207989693,
-0.04642977938055992,
-0.07312426716089249,
3.583750060749803e-34,
0.03348325565457344,
0.009517351165413857,
0.05459976941347122,
0.03511969745159149,
0.07027220726013184,
0.05356273427605629,
0.02378104440867901,
0.00008810847793938592,
0.0412355437874794,
0.050119370222091675,
-0.04460855573415756,
-0.070958711206913,
-0.020109692588448524,
0.01393404696136713,
0.06173853948712349,
-0.02913294918835163,
-0.04360446333885193,
-0.003786772955209017,
-0.06760011613368988,
0.050410836935043335,
0.0009991535916924477,
-0.055996377021074295,
0.014021609909832478,
-0.04092053323984146,
0.020576845854520798,
0.03934542089700699,
0.02220456302165985,
0.01789449155330658,
-0.03105795383453369,
0.003434406826272607,
-0.039421193301677704,
0.026957925409078598,
-0.04235750809311867,
0.027988266199827194,
0.0043396120890975,
-0.07002778351306915,
-0.00009479380969423801,
-0.043403636664152145,
-0.024536630138754845,
-0.04909460246562958,
0.014699151739478111,
0.04038345813751221,
0.0015810680342838168,
-0.13075707852840424,
-0.04177863150835037,
-0.007896114140748978,
-0.024120986461639404,
0.0903187245130539,
-0.018703125417232513,
0.058926090598106384,
0.07459034025669098,
-0.002743893302977085,
-0.056825149804353714,
-0.002250224817544222,
-0.05354011431336403,
0.02589312195777893,
0.0056337472051382065,
0.04168466106057167,
0.061491549015045166,
-0.008129275403916836,
0.039317015558481216,
0.07293596863746643,
-0.08674503862857819,
0.01527806930243969,
-0.056397225707769394,
-0.07278549671173096,
0.000489757105242461,
0.050705596804618835,
0.034912630915641785,
0.009987779892981052,
0.05540335923433304,
0.022553319111466408,
-0.05546622723340988,
-0.07344470173120499,
0.02277202159166336,
0.006547962315380573,
-0.013095458038151264,
0.004668013192713261,
0.07738068699836731,
0.05332280695438385,
-0.049966782331466675,
0.002739870687946677,
0.04666229709982872,
-0.12228045612573624,
-0.030868787318468094,
0.01923270709812641,
0.10863368958234787,
-0.03671331703662872,
-0.023660995066165924,
-0.03158498927950859,
-0.002327489200979471,
-0.006895129568874836,
-0.040674347430467606,
0.04424993321299553,
-0.05100218951702118,
1.0690733115757437e-34,
0.004943492356687784,
0.008579321205615997,
-0.007815634831786156,
0.018789958208799362,
-0.028946692124009132,
-0.03812875598669052,
-0.057830438017845154,
-0.027437806129455566,
-0.050044428557157516,
-0.06948930770158768,
-0.01518281176686287,
0.019703829661011696,
0.047929733991622925,
0.008654044941067696,
0.09301161020994186,
0.057861436158418655,
-0.06729251891374588,
0.01875673606991768,
-0.03622714802622795,
-0.04540913552045822,
-0.03518807142972946,
0.046884968876838684,
-0.06968233734369278,
0.028790710493922234,
-0.062041621655225754,
0.04722527042031288,
0.01583181694149971,
0.02774553932249546,
0.0032283172477036715,
0.01526991929858923,
-0.021772855892777443,
0.0196542926132679,
0.03005076013505459,
-0.02322576940059662,
-0.09066366404294968,
-0.011257624253630638,
0.011389928869903088,
-0.008971384726464748,
0.03637700527906418,
0.05822625383734703,
0.05195486918091774,
-0.04608159512281418,
-0.07754366844892502,
0.014820986427366734,
-0.07643868774175644,
-0.03954777121543884,
-0.0159076489508152,
0.11276842653751373,
-0.01529867947101593,
0.13027900457382202,
-0.1421479731798172,
0.03657737746834755,
-0.04836022108793259,
-0.006024875678122044,
-0.02879330888390541,
-0.011083918623626232,
-0.027890348806977272,
0.053283922374248505,
0.0774180218577385,
-0.014062317088246346,
-0.02987806685268879,
0.014860945753753185,
-0.04054080694913864,
0.0765993669629097,
0.029781002551317215,
0.016230473294854164,
0.003910078667104244,
0.004590282682329416,
0.03481534123420715,
0.011039619334042072,
0.028235556557774544,
0.14077994227409363,
-0.03461053967475891,
-0.018237151205539703,
-0.03357448801398277,
0.03424772992730141,
0.08489532023668289,
0.01680286042392254,
0.1273409128189087,
-0.0703757107257843,
-0.027375172823667526,
-0.05430466681718826,
-0.030317025259137154,
0.10261455178260803,
0.12152206897735596,
0.04650724679231644,
-0.0006556599983014166,
-0.08807270228862762,
0.0588422566652298,
0.036347415298223495,
-0.04071326553821564,
0.0652679055929184,
0.016483141109347343,
0.02013261429965496,
-0.016807591542601585,
-4.8237644989512773e-8,
-0.16281704604625702,
-0.002433473477140069,
0.021209023892879486,
0.02651548571884632,
0.0473545640707016,
-0.005104296375066042,
-0.040640201419591904,
0.12614917755126953,
0.01512683555483818,
-0.04130816087126732,
0.016721142455935478,
-0.03044651448726654,
-0.11556852608919144,
-0.05356244370341301,
0.05736682936549187,
0.07158804684877396,
-0.0031076727900654078,
0.033202651888132095,
-0.008146081119775772,
0.06936434656381607,
0.004374178592115641,
-0.028934577479958534,
0.03925537317991257,
0.015916937962174416,
-0.0049359556287527084,
-0.0570795014500618,
-0.0329008512198925,
0.09888125956058502,
0.06969603151082993,
-0.05948631465435028,
-0.08055883646011353,
-0.009802237153053284,
0.027876805514097214,
0.0029168741311877966,
0.11353258788585663,
0.07822634279727936,
0.04368341341614723,
0.01858183741569519,
-0.005765353795140982,
-0.039261117577552795,
-0.031133588403463364,
0.0265259500592947,
0.009447240270674229,
-0.047789040952920914,
-0.02051573432981968,
0.1103832796216011,
0.04439123719930649,
-0.10192535072565079,
-0.030461642891168594,
-0.08645527809858322,
-0.044188156723976135,
0.05140078067779541,
-0.024720706045627594,
0.07498697936534882,
-0.07440096884965897,
0.06541238725185394,
0.03196481987833977,
-0.06629399210214615,
0.03210647776722908,
0.11053043603897095,
-0.03041253238916397,
0.0244695246219635,
-0.08601883053779602,
-0.014608034864068031
] |
A Unified Theory of Quantum Neural Network Loss Landscapes | https://openreview.net/forum?id=fv8TTt9srF | [
"Eric Ricardo Anschuetz"
] | Poster | Classical neural networks with random initialization famously behave as Gaussian processes in the limit of many neurons, which allows one to completely characterize their training and generalization behavior. No such general understanding exists for quantum neural networks (QNNs), which—outside of certain special cases—are known to not behave as Gaussian processes when randomly initialized. We here prove that QNNs and their first two derivatives instead generally form what we call "Wishart processes," where certain algebraic properties of the network determine the hyperparameters of the process. This Wishart process description allows us to, for the first time: give necessary and sufficient conditions for a QNN architecture to have a Gaussian process limit; calculate the full gradient distribution, generalizing previously known barren plateau results; and calculate the local minima distribution of algebraically constrained QNNs. Our unified framework suggests a certain simple operational definition for the "trainability" of a given QNN model using a newly introduced, experimentally accessible quantity we call the "degrees of freedom" of the network architecture. | quantum machine learning, neural tangent kernel, loss landscape, spin glass, Kac–Rice formula | We unify previous results on the training behavior of quantum neural networks by proving they asymptotically behave as a novel class of random processes. | 1,140 | 2408.11901 | [
-0.07791978865861893,
-0.057093046605587006,
0.03789399191737175,
0.029978889971971512,
-0.062267519533634186,
0.08012522011995316,
-0.017044445499777794,
-0.04421082139015198,
0.039449337869882584,
-0.05179726332426071,
-0.017330603674054146,
0.038177020847797394,
0.006749587133526802,
0.049707818776369095,
-0.11012192815542221,
-0.03371373564004898,
0.05167064815759659,
0.01591303199529648,
-0.07058189809322357,
0.045201778411865234,
-0.00041639007395133376,
0.04651825875043869,
0.027905263006687164,
-0.0010415578726679087,
0.056583601981401443,
-0.04680435732007027,
0.003049674676731229,
0.03340080752968788,
0.04401222988963127,
0.020654235035181046,
0.08108121901750565,
-0.028667166829109192,
-0.02953232452273369,
0.011660391464829445,
-0.004494303837418556,
0.031243639066815376,
-0.054070230573415756,
-0.02801523357629776,
0.04618603363633156,
0.01488402672111988,
0.028741154819726944,
0.013872921466827393,
-0.08432140946388245,
0.10753058642148972,
0.11227809637784958,
0.1059948280453682,
0.03414519131183624,
-0.06720874458551407,
-0.09372816979885101,
-0.1236708015203476,
-0.0037676598876714706,
-0.003409158205613494,
-0.01353614591062069,
0.06203019246459007,
0.04762495681643486,
0.011236297897994518,
-0.015231752768158913,
0.009041963145136833,
-0.08631175011396408,
0.07360133528709412,
-0.04543302208185196,
-0.0920797735452652,
-0.023425456136465073,
-0.02140321210026741,
0.0718507245182991,
0.005531213711947203,
-0.052818749099969864,
0.042240265756845474,
0.036871928721666336,
0.010506222024559975,
-0.04989045858383179,
0.06555351614952087,
-0.08648574352264404,
0.03925945609807968,
0.035337675362825394,
0.006409695837646723,
0.06140998378396034,
0.10616979748010635,
-0.09357329457998276,
-0.02967800199985504,
0.02880118228495121,
-0.04619293287396431,
0.01431069616228342,
-0.03923456370830536,
0.05892122536897659,
0.015081987716257572,
-0.007545487955212593,
0.07091221213340759,
0.05231514573097229,
-0.08974812179803848,
-0.11130789667367935,
-0.005486980080604553,
-0.06281207501888275,
-0.030397245660424232,
0.08104699850082397,
-0.0020139310508966446,
0.03081059642136097,
-0.11016781628131866,
0.00622459314763546,
0.03203083574771881,
0.06197149306535721,
-0.04318387433886528,
0.079546257853508,
-0.035409003496170044,
0.10845255106687546,
-0.008599202148616314,
0.06485278159379959,
0.040217578411102295,
0.04479705169796944,
-0.10773806273937225,
-0.0014109349576756358,
-0.01926925592124462,
0.023147786036133766,
-0.03212172910571098,
-0.018426762893795967,
0.045569173991680145,
-0.001020503230392933,
0.09597166627645493,
-0.03812652453780174,
0.06843948364257812,
-0.11503014713525772,
0.000364653707947582,
-0.06172385439276695,
-0.003140419954434037,
-0.002371021080762148,
-0.009345455095171928,
-0.09874334931373596,
3.820470883529655e-33,
0.041473980993032455,
-0.008825485594570637,
-0.01920783706009388,
-0.027201024815440178,
0.0837051123380661,
-0.0014962394488975406,
0.1205427274107933,
-0.04515441507101059,
0.04060390591621399,
-0.02196156419813633,
-0.013201707042753696,
-0.00548755656927824,
-0.07233680039644241,
0.0032595982775092125,
0.05249829590320587,
0.047716643661260605,
-0.0827699825167656,
-0.02242128737270832,
0.08428886532783508,
-0.05093064904212952,
0.09101703017950058,
-0.06388669461011887,
-0.06006256118416786,
0.0029473421163856983,
-0.0330452062189579,
-0.05393274500966072,
0.06598243117332458,
0.030171861872076988,
-0.08629267662763596,
-0.00755875650793314,
-0.04049721360206604,
0.0465511716902256,
-0.07471337914466858,
-0.05026024952530861,
0.016307126730680466,
0.009746958501636982,
0.010885795578360558,
0.056813228875398636,
-0.006923553068190813,
-0.07651907950639725,
-0.023781998082995415,
-0.003364586504176259,
0.06417030841112137,
-0.00726437708362937,
-0.03295459225773811,
-0.0451129674911499,
0.05929051339626312,
0.02285321243107319,
-0.00633983314037323,
-0.10038860887289047,
0.006238413508981466,
-0.05351686477661133,
-0.016549907624721527,
-0.03977217525243759,
0.028705667704343796,
0.025576572865247726,
0.04953145235776901,
0.056504011154174805,
-0.056439924985170364,
0.04842387139797211,
0.014771453104913235,
0.009014179930090904,
-0.060624368488788605,
0.02894020266830921,
-0.042051590979099274,
0.027072647586464882,
-0.02942419797182083,
0.012844315730035305,
-0.0026324319187551737,
0.07612588256597519,
-0.019311122596263885,
0.0020688371732831,
-0.03245893865823746,
-0.06387300789356232,
0.07635530829429626,
0.002671990543603897,
-0.005311861634254456,
-0.10963098704814911,
0.016162434592843056,
0.012448050081729889,
-0.010179135948419571,
0.01330340001732111,
-0.08692982792854309,
0.02256711944937706,
-0.04064406082034111,
-0.06719717383384705,
0.024791115894913673,
-0.08270954340696335,
-0.025638902559876442,
-0.01853206939995289,
-0.042089711874723434,
-0.03515303134918213,
0.02805347368121147,
0.022207269445061684,
-0.0639762207865715,
-3.1691452594443646e-33,
-0.025906957685947418,
0.052420999854803085,
-0.057955265045166016,
0.07523433864116669,
-0.0018339873058721423,
-0.0004568006843328476,
-0.038885705173015594,
0.020217357203364372,
-0.031056400388479233,
-0.003076741937547922,
-0.025017375126481056,
0.03611702844500542,
0.011687858030200005,
0.06369128078222275,
-0.05955732986330986,
-0.005305367987602949,
-0.019274046644568443,
0.010469107888638973,
0.0949658676981926,
-0.04861762747168541,
0.052715495228767395,
0.05360567942261696,
-0.08289829641580582,
-0.03049657493829727,
0.042456064373254776,
0.05529385805130005,
0.017162885516881943,
0.10847482830286026,
0.005658895708620548,
-0.013094311580061913,
-0.0886363536119461,
-0.14329826831817627,
-0.09869826585054398,
0.0027317730709910393,
-0.017554696649312973,
0.05248278006911278,
0.0804530531167984,
0.0023584833834320307,
-0.04055254906415939,
0.0317356176674366,
0.018542954698204994,
-0.054197899997234344,
0.025599777698516846,
0.062389906495809555,
0.0758073627948761,
-0.005718903616070747,
0.01130988635122776,
-0.030548978596925735,
-0.05642422288656235,
0.06525380164384842,
0.0036191788967698812,
0.08504560589790344,
-0.057039860635995865,
0.06892944872379303,
-0.06878301501274109,
-0.0002625947236083448,
-0.010899879038333893,
0.028516855090856552,
0.06508736312389374,
0.009083576500415802,
-0.020549485459923744,
-0.06997707486152649,
-0.03352705016732216,
0.05836569890379906,
-0.055114924907684326,
-0.00942954234778881,
-0.039418186992406845,
0.10759396851062775,
0.069573774933815,
0.007768227253109217,
-0.002148569328710437,
-0.015807563439011574,
0.029604973271489143,
0.044365689158439636,
-0.006852756720036268,
0.013831247575581074,
0.02401787042617798,
-0.05762464925646782,
-0.018953435122966766,
-0.053503524512052536,
0.06192130222916603,
0.026002274826169014,
0.0015391819179058075,
0.020380383357405663,
0.1007302924990654,
0.08527086675167084,
0.053108103573322296,
0.014501267112791538,
0.08119272440671921,
-0.0659172534942627,
0.06619222462177277,
0.05075816065073013,
0.03036978840827942,
0.024820365011692047,
-0.06796661019325256,
-4.220663640808198e-8,
-0.04037896543741226,
0.011617840267717838,
-0.01081819087266922,
0.02264021895825863,
0.12302330136299133,
-0.02760753035545349,
0.07284006476402283,
0.0584915466606617,
-0.055741894990205765,
-0.013703989796340466,
0.003617119276896119,
0.01068010926246643,
-0.07335001975297928,
0.015319363214075565,
-0.0013094572350382805,
0.06240588054060936,
-0.033346496522426605,
-0.003895294386893511,
0.03617970272898674,
0.009089088067412376,
0.009880335070192814,
0.025849319994449615,
0.011109670624136925,
-0.00007687448669457808,
-0.010864142328500748,
-0.1055898442864418,
-0.015874972566962242,
0.015538734383881092,
-0.026952138170599937,
0.055346112698316574,
0.022278254851698875,
0.004172563552856445,
0.05333567038178444,
0.028574347496032715,
-0.07016851007938385,
0.08673841506242752,
-0.09797332435846329,
-0.04341033101081848,
0.01563650742173195,
-0.005086143966764212,
-0.09340162575244904,
0.032579585909843445,
0.016986148431897163,
0.020992916077375412,
0.010406644083559513,
-0.01608603075146675,
-0.004193327389657497,
-0.05920198932290077,
0.06640534102916718,
0.021600935608148575,
0.03389228507876396,
-0.0023523445706814528,
0.003989949356764555,
-0.02467118762433529,
0.01028570719063282,
0.0148857356980443,
0.009595655836164951,
-0.10409599542617798,
-0.04428158327937126,
0.0372367724776268,
-0.009606846608221531,
0.0762300044298172,
-0.031531378626823425,
0.02767311781644821
] |
Enhancing Robust Fairness via Confusional Spectral Regularization | https://openreview.net/forum?id=lW0ZndAimF | [
"Gaojie Jin",
"Sihao Wu",
"Jiaxu Liu",
"Tianjin Huang",
"Ronghui Mu"
] | Poster | Recent research has highlighted a critical issue known as ``robust fairness", where robust accuracy varies significantly across different classes, undermining the reliability of deep neural networks (DNNs).
A common approach to address this has been to dynamically reweight classes during training, giving more weight to those with lower empirical robust performance.
However, we find there is a divergence of class-wise robust performance between training set and testing set, which limits the effectiveness of these explicit reweighting methods, indicating the need for a principled alternative.
In this work, we derive a robust generalization bound for the worst-class robust error within the PAC-Bayesian framework, accounting for unknown data distributions.
Our analysis shows that the worst-class robust error is influenced by two main factors: the spectral norm of the empirical robust confusion matrix and the information embedded in the model and training set.
While the latter has been extensively studied, we propose a novel regularization technique targeting the spectral norm of the robust confusion matrix to improve worst-class robust accuracy and enhance robust fairness.
We validate our approach through comprehensive experiments on various datasets and models, demonstrating its effectiveness in enhancing robust fairness. | Adversarial robustness, Worst-class robust accuracy, Robust fairness, PAC-Bayes | null | 1,137 | 2501.13273 | [
-0.10590913891792297,
-0.08862632513046265,
0.008407223969697952,
0.04071145877242088,
0.00187548971734941,
0.08643587678670883,
0.024977227672934532,
-0.026179902255535126,
-0.07138420641422272,
-0.017824461683630943,
-0.06070471182465553,
-0.01344449631869793,
0.062253691256046295,
-0.028329910710453987,
-0.12003997713327408,
-0.03913918882608414,
0.1324784904718399,
0.10989345610141754,
-0.16876660287380219,
0.04766828194260597,
0.04193361476063728,
0.0666160136461258,
0.007963920012116432,
0.025927841663360596,
0.02941405214369297,
-0.06932314485311508,
-0.011942578479647636,
-0.028455795720219612,
-0.0017725145444273949,
-0.01041216030716896,
0.02960822917521,
-0.104890838265419,
-0.007417352870106697,
-0.0013544613029807806,
-0.017275789752602577,
-0.03513624891638756,
0.02295967936515808,
0.0043550594709813595,
0.03674273565411568,
0.058710940182209015,
0.007058747112751007,
0.02837158553302288,
0.01716701127588749,
0.05062425509095192,
0.0011891305912286043,
0.006824397947639227,
0.04274683818221092,
-0.04136616364121437,
0.043303802609443665,
-0.02825254201889038,
0.04680641368031502,
0.036645062267780304,
-0.019186735153198242,
0.047117456793785095,
-0.05397004261612892,
0.036013081669807434,
0.02592085301876068,
0.031634699553251266,
-0.023057352751493454,
0.02362380363047123,
-0.04716545343399048,
-0.14167997241020203,
0.031522803008556366,
0.00624213507398963,
-0.017316171899437904,
-0.0020690576639026403,
-0.034353092312812805,
0.0033250872511416674,
-0.0011026867432519794,
0.039426982402801514,
-0.027764679864048958,
0.1124560683965683,
-0.04526989907026291,
0.06427934020757675,
0.02658923529088497,
0.09194391965866089,
0.04321152716875076,
-0.013185090385377407,
0.007783181499689817,
-0.027806060388684273,
0.08633334189653397,
0.03571959212422371,
0.014526624232530594,
-0.010994330048561096,
0.18260133266448975,
0.0016512988368049264,
-0.010066757909953594,
0.07680293917655945,
0.02296380326151848,
-0.02166637033224106,
-0.06213238090276718,
0.0017242229077965021,
0.01499718613922596,
-0.005204231012612581,
-0.0038657865952700377,
0.03406045585870743,
0.05457601323723793,
-0.08031052350997925,
-0.0031574643217027187,
0.11383847892284393,
-0.0090646892786026,
0.021173758432269096,
-0.008663037791848183,
0.006325496360659599,
0.036924757063388824,
0.02370307967066765,
0.031791795045137405,
0.018131380900740623,
0.008802389726042747,
-0.04502561315894127,
0.03581971302628517,
-0.004468436818569899,
-0.009309087879955769,
0.012033641338348389,
-0.02881486527621746,
0.041751179844141006,
0.011644306592643261,
-0.015950271859765053,
-0.048472750931978226,
0.03030787967145443,
-0.08643880486488342,
-0.02606123685836792,
0.06473492830991745,
-0.016665982082486153,
0.040298040956258774,
0.004030976910144091,
-0.06836830824613571,
3.139434272385347e-33,
0.025603096932172775,
0.039601389318704605,
-0.006752542685717344,
-0.05728808790445328,
0.0051115876995027065,
-0.02849840745329857,
-0.07727200537919998,
0.007630001753568649,
-0.05925583839416504,
0.05110124498605728,
0.0031335155945271254,
0.003494556760415435,
-0.09678902477025986,
0.010737042874097824,
0.022217435762286186,
0.10193938761949539,
-0.06257890909910202,
0.0009693556348793209,
-0.042333509773015976,
-0.09748824685811996,
0.08248674869537354,
-0.04066922888159752,
-0.009043069556355476,
-0.04433654248714447,
0.0007231303607113659,
0.02006373554468155,
0.06889785081148148,
0.021174203604459763,
0.026985296979546547,
0.0016393164405599236,
-0.11042235046625137,
-0.021367831155657768,
0.10249045491218567,
0.0012511772802099586,
0.04173315688967705,
-0.027297893539071083,
0.002036972204223275,
0.04071863740682602,
0.0167836993932724,
-0.00424140365794301,
0.0033139747101813555,
0.028430119156837463,
0.046938858926296234,
-0.0240932647138834,
-0.013675503432750702,
-0.055369917303323746,
0.061197277158498764,
-0.09411058574914932,
-0.051818158477544785,
-0.06371169537305832,
0.01854855753481388,
0.008977607823908329,
0.0021412663627415895,
-0.02990568056702614,
-0.07107844203710556,
-0.08783641457557678,
0.046497516334056854,
0.045919474214315414,
0.005230077542364597,
0.04635974392294884,
-0.007136721163988113,
-0.0028974160086363554,
-0.02063891477882862,
0.025676069781184196,
-0.06988447159528732,
0.020102718845009804,
-0.0024039370473474264,
0.044272229075431824,
0.025204554200172424,
-0.03656066209077835,
0.052682310342788696,
0.001512251328676939,
-0.042184121906757355,
-0.06299331039190292,
0.03660469129681587,
0.03740634024143219,
0.07274255901575089,
0.01812126301229,
-0.05519390478730202,
-0.005585282109677792,
0.007962898351252079,
-0.004385954234749079,
-0.03925328701734543,
0.033309273421764374,
-0.14361780881881714,
-0.0019013498676940799,
0.03853752464056015,
-0.0029481472447514534,
-0.006832522805780172,
0.022966697812080383,
-0.01689695194363594,
-0.011949311010539532,
-0.01733575575053692,
-0.00943523645401001,
-0.02203337661921978,
-1.8781257115461628e-33,
-0.07014267146587372,
0.044960662722587585,
-0.09779786318540573,
0.0924006775021553,
-0.005390573758631945,
-0.05166448652744293,
0.019351286813616753,
0.09458393603563309,
-0.033667393028736115,
-0.075581394135952,
-0.008685426786541939,
-0.052155882120132446,
0.019507458433508873,
-0.029580580070614815,
0.06042155995965004,
-0.057234350591897964,
-0.08898723125457764,
0.08562161773443222,
0.004721696022897959,
-0.012631348334252834,
0.10913068801164627,
0.030939266085624695,
-0.06864366680383682,
0.06519076973199844,
-0.008921136148273945,
0.03757534548640251,
-0.11477339267730713,
0.08738794177770615,
-0.002624803688377142,
-0.018760012462735176,
-0.07449080795049667,
-0.02228492684662342,
-0.031635843217372894,
0.012775500304996967,
0.03236312419176102,
0.029855523258447647,
0.008036294020712376,
-0.007579973433166742,
-0.08146942406892776,
0.08022761344909668,
-0.0005463757552206516,
0.041910648345947266,
-0.11476750671863556,
-0.012635585851967335,
0.0028066213708370924,
0.007690878119319677,
0.04098352789878845,
-0.07219923287630081,
0.06932669878005981,
-0.0404292494058609,
-0.01870161108672619,
-0.04619593545794487,
-0.040854427963495255,
0.1648789346218109,
-0.04451007768511772,
0.008464338257908821,
0.048970241099596024,
0.010677049867808819,
0.03527142107486725,
0.11873380094766617,
-0.06876814365386963,
0.004271161742508411,
-0.08478926122188568,
0.020087670534849167,
-0.007662294432520866,
-0.013850804418325424,
-0.06011870130896568,
-0.005572193767875433,
0.05812123417854309,
0.06390689313411713,
0.010940318927168846,
-0.009899950586259365,
0.0515827015042305,
0.03807200491428375,
-0.021396875381469727,
0.028734609484672546,
0.03621499985456467,
0.00493203941732645,
-0.03273453190922737,
0.1291392594575882,
0.03391969949007034,
-0.018273772671818733,
0.059792447835206985,
0.07403156161308289,
0.07003165781497955,
0.08621276915073395,
0.03757326304912567,
-0.000751605024561286,
-0.038604021072387695,
-0.0213300958275795,
-0.02609354443848133,
0.027961192652583122,
0.019200770184397697,
0.025500399991869926,
-0.05733635649085045,
-4.771933959091257e-8,
-0.06236260384321213,
0.025498544797301292,
-0.024942481890320778,
0.02955625206232071,
-0.0001684020389802754,
-0.06154793128371239,
-0.016948522999882698,
0.01731960102915764,
-0.033302728086709976,
-0.009767059236764908,
0.046189598739147186,
-0.03146098181605339,
-0.02369694598019123,
-0.06679096817970276,
-0.03597601503133774,
0.11476690322160721,
-0.03977172449231148,
0.11087173223495483,
0.030941126868128777,
0.027331504970788956,
0.018080728128552437,
0.012144241482019424,
0.03274908661842346,
-0.08133076131343842,
-0.031229430809617043,
-0.05166875943541527,
0.006865885574370623,
0.1039753407239914,
0.05635881423950195,
0.06744087487459183,
-0.08135151118040085,
-0.05034228786826134,
-0.023405252024531364,
-0.03731968626379967,
-0.023218441754579544,
0.13176418840885162,
-0.025501515716314316,
-0.0504213348031044,
0.006382989697158337,
0.0008353045559488237,
0.014053171500563622,
0.04556674137711525,
-0.019182113930583,
-0.016393011435866356,
0.05266811326146126,
0.027323471382260323,
-0.002380393911153078,
0.0028582692611962557,
0.03772696107625961,
-0.05082773044705391,
0.03972822055220604,
0.017324168235063553,
-0.004134646616876125,
0.03180382028222084,
0.029402179643511772,
-0.04214493930339813,
-0.03129241615533829,
-0.08242180198431015,
-0.1183295026421547,
0.044960733503103256,
0.028137385845184326,
0.04982021450996399,
-0.04008130729198456,
-0.0002498312678653747
] |
Perplexity Trap: PLM-Based Retrievers Overrate Low Perplexity Documents | https://openreview.net/forum?id=U1T6sq12uj | [
"Haoyu Wang",
"Sunhao Dai",
"Haiyuan Zhao",
"Liang Pang",
"Xiao Zhang",
"Gang Wang",
"Zhenhua Dong",
"Jun Xu",
"Ji-Rong Wen"
] | Poster | Previous studies have found that PLM-based retrieval models exhibit a preference for LLM-generated content, assigning higher relevance scores to these documents even when their semantic quality is comparable to human-written ones. This phenomenon, known as source bias, threatens the sustainable development of the information access ecosystem. However, the underlying causes of source bias remain unexplored. In this paper, we explain the process of information retrieval with a causal graph and discover that PLM-based retrievers learn perplexity features for relevance estimation, causing source bias by ranking the documents with low perplexity higher. Theoretical analysis further reveals that the phenomenon stems from the positive correlation between the gradients of the loss functions in language modeling task and retrieval task. Based on the analysis, a causal-inspired inference-time debiasing method is proposed, called **C**ausal **D**iagnosis and **C**orrection (CDC). CDC first diagnoses the bias effect of the perplexity and then separates the bias effect from the overall estimated relevance score. Experimental results across three domains demonstrate the superior debiasing effectiveness of CDC, emphasizing the validity of our proposed explanatory framework. Source codes are available at https://github.com/WhyDwelledOnAi/Perplexity-Trap. | Source Bias, LLM-Generated Content, Infomation Retrieval, Large Language Models | We verify the causal effect of document perplexity on estimated relevance scores in PLM-based retrievers to explain the cause of source bias. | 1,134 | 2503.08684 | [
-0.028808023780584335,
-0.02908366173505783,
-0.002007941482588649,
0.059005364775657654,
0.08176206797361374,
0.019705284386873245,
0.017754467204213142,
0.04286660626530647,
0.07708937674760818,
-0.0013881344348192215,
-0.06263983994722366,
0.06629391014575958,
0.11703258752822876,
0.014592299237847328,
-0.03538236767053604,
0.0486394464969635,
0.12434782087802887,
0.022254742681980133,
-0.06785200536251068,
-0.04472159594297409,
0.028410155326128006,
0.06156354025006294,
0.033199090510606766,
-0.04391592741012573,
-0.044931504875421524,
-0.04637874290347099,
-0.0556398406624794,
0.051490623503923416,
0.051085300743579865,
-0.029020771384239197,
0.0022955797612667084,
0.0928817167878151,
-0.07220803201198578,
-0.012327232398092747,
-0.017437802627682686,
0.00026634082314558327,
-0.0449075922369957,
-0.006706864573061466,
0.024754585698246956,
0.011279276572167873,
0.047619935125112534,
0.06431155651807785,
0.003908604849129915,
0.05908503755927086,
0.06412658095359802,
0.0008159549324773252,
-0.00906556099653244,
0.012718947604298592,
-0.03564628213644028,
0.04578257352113724,
-0.06730034947395325,
0.008128484711050987,
0.021180301904678345,
-0.006999047007411718,
0.040697526186704636,
-0.021079324185848236,
-0.031686846166849136,
0.06661692261695862,
-0.12376149743795395,
-0.10328956693410873,
-0.11012526601552963,
-0.15678544342517853,
-0.07843974232673645,
-0.03322490304708481,
0.017346493899822235,
-0.03012515790760517,
0.01860445737838745,
0.012286399491131306,
-0.07073540985584259,
0.01293943915516138,
0.026313822716474533,
-0.011111319996416569,
-0.008195905014872551,
0.07448232918977737,
0.009495277889072895,
0.017861608415842056,
0.005387639161199331,
0.007109812926501036,
0.04515939578413963,
-0.003783205524086952,
-0.001963587012141943,
0.052769482135772705,
0.048230547457933426,
-0.03545799106359482,
0.060264069586992264,
-0.040770549327135086,
0.11003842204809189,
-0.014179080724716187,
0.016219494864344597,
0.04326478764414787,
0.008914516307413578,
-0.09583593159914017,
0.0713171437382698,
-0.004556427709758282,
0.04219720512628555,
0.029015114530920982,
0.003107307944446802,
0.009805737063288689,
0.04839172959327698,
0.03622064366936684,
0.013911825604736805,
0.054305173456668854,
-0.012300506234169006,
-0.04015738517045975,
0.00928870402276516,
0.023286456242203712,
-0.03320451080799103,
0.14339064061641693,
0.01638900302350521,
-0.08785240352153778,
-0.06091396138072014,
0.030138378962874413,
0.031155791133642197,
-0.004849984310567379,
0.004610838368535042,
-0.0017738284077495337,
0.05861172452569008,
-0.03587726503610611,
0.02939351089298725,
0.013853163458406925,
-0.04752129316329956,
-0.015034392476081848,
0.03127238526940346,
-0.036084625869989395,
-0.03008306212723255,
-0.04817323386669159,
-0.05503040924668312,
1.1204516441740416e-33,
0.009052393026649952,
0.0675707459449768,
-0.02668113075196743,
0.00866621546447277,
0.029399942606687546,
-0.025803973898291588,
-0.0018832647474482656,
0.06189453974366188,
0.005143024958670139,
-0.006058734375983477,
-0.030375216156244278,
0.06235668808221817,
-0.016353286802768707,
-0.03087017312645912,
-0.023084161803126335,
0.008399724029004574,
-0.0523507185280323,
0.04160922393202782,
-0.05923616141080856,
-0.03163888305425644,
0.07439478486776352,
-0.031921617686748505,
0.05748860910534859,
-0.04464210569858551,
0.036337774246931076,
0.04130232334136963,
0.04161076992750168,
-0.06485390663146973,
-0.015409338288009167,
-0.029615528881549835,
-0.03306062892079353,
-0.0473637729883194,
0.07783208042383194,
-0.08245744556188583,
-0.0065425182692706585,
-0.03549716994166374,
-0.0739673599600792,
0.03268173709511757,
0.01915321871638298,
0.023992590606212616,
-0.06314360350370407,
-0.06259824335575104,
0.024513941258192062,
-0.037941109389066696,
-0.06505659967660904,
-0.09192153811454773,
0.006880090571939945,
-0.009508362039923668,
0.002833140315487981,
-0.025282064452767372,
0.08309046924114227,
0.003579601878300309,
0.005920009687542915,
0.0524180643260479,
0.0433674082159996,
0.0788756012916565,
-0.03199390321969986,
-0.006514326203614473,
-0.026985304430127144,
-0.022204501554369926,
0.07972434908151627,
0.022190876305103302,
0.032196562737226486,
0.02535281330347061,
0.05556439980864525,
0.05352014675736427,
-0.07708241790533066,
0.033185023814439774,
0.05705921724438667,
0.05228413641452789,
0.04769182950258255,
-0.013616110198199749,
-0.012147242203354836,
-0.08088679611682892,
0.07586817443370819,
-0.012248235754668713,
-0.06769724935293198,
-0.02938436158001423,
0.0670178011059761,
-0.0037812343798577785,
-0.00812654010951519,
-0.08324553072452545,
0.02045721374452114,
-0.08830731362104416,
-0.12325004488229752,
-0.002485639415681362,
0.0018266397528350353,
-0.09429959207773209,
0.060123514384031296,
0.0052560484036803246,
0.05085572972893715,
0.025066355243325233,
-0.027426570653915405,
-0.026299593970179558,
-0.020670626312494278,
-1.964241417496946e-34,
-0.09129504859447479,
-0.08097568154335022,
0.05773507058620453,
0.009471075609326363,
-0.05236341059207916,
-0.02057506889104843,
-0.0328165628015995,
0.04436487704515457,
-0.021001700311899185,
-0.0013474118895828724,
-0.033160679042339325,
-0.06903410702943802,
0.01730756275355816,
0.023085718974471092,
0.03074575774371624,
0.09855156391859055,
0.01610822044312954,
0.031238509342074394,
-0.08504202216863632,
0.06365374475717545,
0.008692295290529728,
0.030625028535723686,
-0.08545597642660141,
0.008840754628181458,
0.039223767817020416,
0.04855819046497345,
0.026930760592222214,
0.004580898210406303,
0.06333789974451065,
-0.0049359481781721115,
-0.07297277450561523,
0.0032876257319003344,
-0.08735402673482895,
-0.02522602304816246,
-0.09540393948554993,
0.07086627930402756,
0.01355831790715456,
0.024734754115343094,
-0.1083611473441124,
0.07370591163635254,
0.017645524814724922,
0.031113000586628914,
-0.0786726325750351,
0.001318995957262814,
0.002421214245259762,
-0.005326753482222557,
-0.13367481529712677,
0.06659689545631409,
0.07117694616317749,
0.009484256617724895,
-0.008282558061182499,
-0.006805578246712685,
0.02651558816432953,
0.052374016493558884,
-0.09556226432323456,
-0.08412022143602371,
-0.012652014382183552,
-0.0626644715666771,
0.03684833645820618,
0.023227030411362648,
-0.007672547362744808,
-0.0323062501847744,
-0.0718904361128807,
0.0012799487449228764,
0.020107196643948555,
-0.050463587045669556,
0.03830000013113022,
-0.034472785890102386,
-0.04431263357400894,
0.011392226442694664,
0.06017047539353371,
-0.008489511907100677,
-0.05884665623307228,
-0.12514829635620117,
0.05723565071821213,
0.1315278857946396,
-0.012825672514736652,
0.005073541775345802,
-0.06455092877149582,
-0.024858001619577408,
0.021057914942502975,
0.050063423812389374,
-0.002655884250998497,
0.047911737114191055,
0.03314957022666931,
0.11314090341329575,
0.04693794623017311,
0.055555470287799835,
-0.018980322405695915,
-0.02509242855012417,
-0.024856988340616226,
-0.06004972755908966,
-0.0493948757648468,
0.09070756286382675,
-0.02592521719634533,
-4.497528749425328e-8,
-0.103045254945755,
0.003629878396168351,
-0.09717819094657898,
0.06511332094669342,
-0.008953875862061977,
-0.06970106065273285,
-0.035197868943214417,
0.10784773528575897,
-0.03543844819068909,
-0.009510890580713749,
0.028314391151070595,
-0.026773283258080482,
-0.0886370837688446,
0.020131990313529968,
0.07967842370271683,
0.07236138731241226,
0.06679276376962662,
-0.0009558394667692482,
0.024402527138590813,
-0.08403492718935013,
0.020667506381869316,
-0.012338217347860336,
0.05176934972405434,
0.045260488986968994,
0.0350269190967083,
-0.024997759610414505,
0.012478327378630638,
0.052358921617269516,
0.04914403334259987,
-0.03681127354502678,
-0.009075853042304516,
0.054743971675634384,
-0.03883035108447075,
-0.02489444985985756,
0.05430106446146965,
0.10801483690738678,
0.0892060324549675,
-0.04825737327337265,
-0.055826686322689056,
0.05342187359929085,
0.04626529663801193,
0.08260726183652878,
0.005041253753006458,
0.06982078403234482,
0.03911424055695534,
-0.0006833908846601844,
-0.01755218394100666,
-0.06132283806800842,
0.05991198867559433,
-0.05340312048792839,
0.039738696068525314,
-0.0913662388920784,
0.006073989439755678,
-0.026230664923787117,
0.029355578124523163,
0.012611479498445988,
0.014313735999166965,
-0.0029445074032992125,
0.004807054530829191,
0.03901189565658569,
0.07445471733808517,
0.02997892163693905,
0.02783016487956047,
-0.009113963693380356
] |
HART: Efficient Visual Generation with Hybrid Autoregressive Transformer | https://openreview.net/forum?id=q5sOv4xQe4 | [
"Haotian Tang",
"Yecheng Wu",
"Shang Yang",
"Enze Xie",
"Junsong Chen",
"Junyu Chen",
"Zhuoyang Zhang",
"Han Cai",
"Yao Lu",
"Song Han"
] | Poster | We introduce Hybrid Autoregressive Transformer (HART), the first autoregressive (AR) visual generation model capable of directly generating 1024x1024 images, rivaling diffusion models in image generation quality. Existing AR models face limitations due to the poor image reconstruction quality of their discrete tokenizers and the prohibitive training costs associated with generating 1024px images. To address these challenges, we present the hybrid tokenizer, which decomposes the continuous latents from the autoencoder into two components: discrete tokens representing the big picture and continuous tokens representing the residual components that cannot be represented by the discrete tokens. The discrete component is modeled by a scalable-resolution discrete AR model, while the continuous component is learned with a lightweight residual diffusion module with only 37M parameters. Compared with the discrete-only VAR tokenizer, our hybrid approach improves reconstruction FID from 2.11 to 0.30 on MJHQ-30K, leading to a 31% generation FID improvement from 7.85 to 5.38. HART also outperforms state-of-the-art diffusion models in both FID and CLIP score, with 4.5-7.7$\times$ higher throughput and 6.9-13.4$\times$ lower MACs. Our code is open sourced at https://github.com/mit-han-lab/hart. | autoregressive models, image generation, text-to-image | An autoregressive model that can directly generate 1024x1024 images. It takes advantage of hybrid tokenization and residual diffusion to model both continuous and discrete tokens. | 1,128 | 2410.10812 | [
-0.042448412626981735,
-0.05958633869886398,
-0.009502656757831573,
0.04779956862330437,
0.028621859848499298,
-0.0203789621591568,
-0.0522252582013607,
-0.03394043818116188,
0.09323804825544357,
-0.03163008764386177,
-0.01023262832313776,
-0.0012933771358802915,
-0.031126534566283226,
0.0475732646882534,
-0.034295227378606796,
0.00047836825251579285,
-0.00024310313165187836,
0.08158841729164124,
-0.09000159054994583,
-0.06599799543619156,
-0.015287667512893677,
-0.044549308717250824,
-0.0639563575387001,
-0.07610440254211426,
0.051173362880945206,
0.08634451031684875,
0.0742993950843811,
-0.09097244590520859,
0.04942501708865166,
-0.05474388971924782,
0.01628776453435421,
-0.024948401376605034,
-0.05322914198040962,
0.0018259428907185793,
-0.0035080586094409227,
0.0391589030623436,
-0.0357554666697979,
0.05902346968650818,
0.05701950937509537,
-0.07070506364107132,
0.035688504576683044,
0.01613963581621647,
0.002329327864572406,
-0.06361906230449677,
0.06521843373775482,
0.0028489509131759405,
-0.002859782427549362,
-0.03669361770153046,
0.03148748353123665,
-0.013630026951432228,
0.0002196219575125724,
0.06215719133615494,
-0.037645500153303146,
0.034136541187763214,
-0.03378256410360336,
0.012153779156506062,
0.03864200785756111,
-0.07866480946540833,
0.026789309456944466,
-0.01371940691024065,
-0.08906634896993637,
0.02941286191344261,
0.049629732966423035,
-0.041103266179561615,
-0.0676192045211792,
-0.015613132156431675,
0.05536015331745148,
-0.0016667599556967616,
0.039651867002248764,
-0.0018310515442863107,
-0.07731053978204727,
0.010878046043217182,
-0.06867093592882156,
-0.019528962671756744,
0.01862191967666149,
-0.008836854249238968,
0.07130259275436401,
-0.08207407593727112,
0.10217928886413574,
-0.0302041694521904,
0.06584446877241135,
0.02374880015850067,
0.09695138037204742,
0.006767186801880598,
0.04315529018640518,
0.023088177666068077,
-0.016721686348319054,
0.07333239912986755,
-0.00975632295012474,
0.003300333395600319,
-0.0896601602435112,
0.0401843823492527,
-0.044716235250234604,
0.0437181331217289,
0.01738417148590088,
-0.021770324558019638,
0.0006680826190859079,
-0.041627220809459686,
0.008405597880482674,
0.09343162924051285,
0.06670264154672623,
-0.005298331379890442,
0.08719141036272049,
-0.004957642871886492,
0.030704034492373466,
-0.061008427292108536,
0.08842609077692032,
0.051329322159290314,
0.029797332361340523,
0.0022600372321903706,
-0.011165102012455463,
0.04712367430329323,
-0.010256261564791203,
-0.019709359854459763,
0.07042825222015381,
-0.04865274950861931,
-0.06355075538158417,
-0.02643592841923237,
0.06968045979738235,
0.019225269556045532,
0.023824235424399376,
-0.00267314282245934,
-0.07370705902576447,
0.10290335863828659,
0.019206566736102104,
-0.033987727016210556,
0.010142002254724503,
3.569356637519213e-33,
-0.09188541024923325,
0.01128749642521143,
0.022357314825057983,
-0.015386125072836876,
0.0680304542183876,
0.00791392382234335,
0.03533479943871498,
-0.04095114395022392,
0.026002923026680946,
-0.05276671424508095,
-0.04696797952055931,
-0.03986244276165962,
-0.05009497329592705,
0.13530047237873077,
0.026583356782794,
-0.06858654320240021,
-0.07206866145133972,
0.06457055360078812,
0.01871212013065815,
-0.03289385885000229,
0.012833028100430965,
0.029715096578001976,
0.021723296493291855,
0.0316290557384491,
-0.04517592489719391,
-0.03959260508418083,
0.07643023133277893,
-0.006394308526068926,
-0.004488440230488777,
0.013784498907625675,
-0.15462501347064972,
-0.04067937657237053,
0.06858488917350769,
-0.015440690331161022,
0.019897334277629852,
0.01579585298895836,
-0.008944361470639706,
-0.03708544373512268,
-0.08101695030927658,
-0.024607202038168907,
-0.006800519302487373,
0.10648699849843979,
-0.02312987670302391,
-0.030202187597751617,
-0.10519210994243622,
0.027231762185692787,
0.01041994709521532,
0.07923439145088196,
-0.09149310737848282,
-0.00937573704868555,
0.030708730220794678,
0.0016835315618664026,
-0.03822413459420204,
-0.07103614509105682,
0.016512257978320122,
-0.03744751587510109,
0.03562679886817932,
-0.039477571845054626,
0.048078540712594986,
-0.008959564380347729,
0.010092107579112053,
0.03019094653427601,
0.0065426304936409,
0.09769708663225174,
0.023765793070197105,
-0.05318885296583176,
0.039456821978092194,
0.0410991832613945,
-0.032689157873392105,
0.03667454421520233,
-0.011204508133232594,
0.04324070364236832,
-0.019146418198943138,
-0.14119985699653625,
0.07970703393220901,
-0.04251400753855705,
0.0822785347700119,
0.05077364668250084,
-0.048654213547706604,
0.03566398099064827,
-0.11934860795736313,
0.0463787205517292,
0.028920438140630722,
-0.106552354991436,
0.024707235395908356,
-0.00883396714925766,
-0.005987491458654404,
-0.045836448669433594,
-0.03034970909357071,
-0.12105432897806168,
0.03242705762386322,
-0.011200672946870327,
0.018727583810687065,
0.023013658821582794,
-0.0036891803611069918,
-2.6954638314041924e-33,
0.032339032739400864,
0.15350249409675598,
-0.00030711459112353623,
0.047730669379234314,
-0.09333198517560959,
-0.05450822785496712,
0.055097855627536774,
0.06087623909115791,
0.01625056564807892,
-0.11471660435199738,
-0.032752394676208496,
0.011720153503119946,
-0.0285582784563303,
0.00944476667791605,
0.011671110056340694,
-0.03352315351366997,
0.007541315630078316,
-0.012082193978130817,
-0.02458513341844082,
0.03736238554120064,
-0.006788571365177631,
0.051352012902498245,
-0.012045339681208134,
0.06024844944477081,
-0.11752264946699142,
0.0761447623372078,
-0.05590394511818886,
0.05288758501410484,
0.008978371508419514,
-0.027087802067399025,
0.09142448753118515,
0.010807247832417488,
0.0246463343501091,
-0.030873946845531464,
-0.09412418305873871,
-0.005676123779267073,
0.031934890896081924,
-0.002290488453581929,
-0.008693594485521317,
0.057240117341279984,
0.04384437948465347,
-0.0679699257016182,
-0.019519204273819923,
0.06057066470384598,
-0.039124298840761185,
0.008025683462619781,
0.06213851645588875,
-0.04633864760398865,
0.03403565660119057,
-0.000814804167021066,
0.018571563065052032,
0.041225239634513855,
-0.09263765811920166,
0.046875961124897,
-0.03941376134753227,
-0.08680678904056549,
-0.0337333008646965,
0.012996526435017586,
0.05793998762965202,
-0.0054106819443404675,
-0.011324366554617882,
-0.05416932702064514,
0.06322700530290604,
-0.10704892128705978,
-0.012729369103908539,
-0.051309093832969666,
-0.0368688590824604,
0.02115248143672943,
-0.039203159511089325,
0.08667756617069244,
0.006666169501841068,
-0.0011383637320250273,
0.024889733642339706,
0.08502549678087234,
0.05317579209804535,
-0.0931171253323555,
0.06784113496541977,
-0.007962502539157867,
0.07003780454397202,
-0.010061931796371937,
-0.03413412719964981,
-0.0030163400806486607,
-0.0029225118923932314,
0.04927663505077362,
-0.022245043888688087,
-0.02444474771618843,
-0.0025499926414340734,
-0.02063746750354767,
0.060459040105342865,
0.06418836861848831,
-0.009375326335430145,
0.030384240671992302,
0.043747011572122574,
0.027282128110527992,
-0.027895506471395493,
-4.648353169045549e-8,
-0.04401980713009834,
0.07048996537923813,
-0.04798921197652817,
-0.02600114978849888,
-0.014145583845674992,
-0.0544695220887661,
0.020238887518644333,
0.018582655116915703,
0.027840230613946915,
-0.010172649286687374,
0.06153137981891632,
-0.08276509493589401,
-0.07283356040716171,
-0.03359771519899368,
-0.017193568870425224,
0.03080650232732296,
0.06438694149255753,
-0.0170882735401392,
0.051229991018772125,
-0.041444048285484314,
-0.12775596976280212,
0.06213124096393585,
-0.025699084624648094,
-0.04829855263233185,
0.06192054599523544,
-0.0028190051671117544,
0.004189212806522846,
-0.004885320086032152,
0.08232226967811584,
0.008300023153424263,
-0.008464055135846138,
0.09724901616573334,
0.02177317440509796,
-0.05119613930583,
-0.04668503254652023,
0.07632792741060257,
0.05870494246482849,
0.007269814610481262,
-0.04329114779829979,
-0.0832889974117279,
-0.031005386263132095,
0.06306831538677216,
-0.040881071239709854,
-0.03172063082456589,
-0.05218249931931496,
0.014087053015828133,
0.04326200485229492,
-0.07144881784915924,
-0.0149905476719141,
-0.002526544500142336,
0.03192214295268059,
-0.0647868886590004,
0.026950335130095482,
0.053638435900211334,
0.055735327303409576,
-0.03542414680123329,
0.038429927080869675,
-0.054289303719997406,
0.10646297037601471,
0.06154191121459007,
0.02815290167927742,
-0.011024626903235912,
-0.06970443576574326,
0.07321053743362427
] |
ANaGRAM: A Natural Gradient Relative to Adapted Model for efficient PINNs learning | https://openreview.net/forum?id=o1IiiNIoaA | [
"Nilo Schwencke",
"Cyril Furtlehner"
] | Poster | In the recent years, Physics Informed Neural Networks (PINNs) have received strong interest as a method to solve PDE driven systems, in particular for data assimilation purpose. This method is still in its infancy, with many shortcomings and failures that remain not properly understood.
In this paper we propose a natural gradient approach to PINNs which contributes to speed-up and improve the accuracy of the training.
Based on an in depth analysis of the differential geometric structures of the problem, we come up with two distinct contributions:
(i) a new natural gradient algorithm that scales as $\min(P^2S, S^2P)$, where $P$ is the number of parameters, and $S$ the batch size;
(ii) a mathematically principled reformulation of the PINNs problem that allows the extension of natural gradient to it, with proved connections to Green's function theory. | PINNs, SciML, PDEs, Natural Gradient, Neural Tangent Kernel | We propose a new natural gradient that scales as $\min(P^2S, S^2P)$, $P$ the number of parameters and $S$ the batch size, focusing on its application to PINNs traning. | 1,125 | 2412.10782 | [
-0.08914659917354584,
-0.11425312608480453,
0.0714353695511818,
0.02301786094903946,
0.0031509497202932835,
-0.023969464004039764,
-0.006603100802749395,
-0.0065138209611177444,
0.02921871840953827,
-0.0016500463243573904,
0.009697535075247288,
0.02657443657517433,
-0.032332174479961395,
0.011315827257931232,
-0.08578965812921524,
0.02809346467256546,
-0.06216380372643471,
0.05827467516064644,
-0.07793596386909485,
-0.05426177382469177,
-0.05954263359308243,
0.025673087686300278,
-0.011379851959645748,
0.07658467441797256,
0.029909882694482803,
-0.03176823630928993,
0.05967813357710838,
0.016785738989710808,
-0.010593746788799763,
-0.019092757254838943,
0.08587256819009781,
0.008952249772846699,
0.00788131169974804,
0.026062337681651115,
-0.10362691432237625,
0.04149199649691582,
-0.06968428939580917,
0.03524046018719673,
-0.04345734044909477,
0.005971670150756836,
0.033954378217458725,
0.067106232047081,
0.04848162829875946,
0.011585087515413761,
0.1728089302778244,
-0.012361393310129642,
0.0026664058677852154,
-0.01028874609619379,
0.0392393134534359,
-0.005014423280954361,
0.017642349004745483,
-0.003320063930004835,
-0.06603677570819855,
-0.027551934123039246,
0.06473834067583084,
-0.007883117534220219,
0.06117960438132286,
-0.03241885453462601,
0.009252320975065231,
-0.014936023391783237,
0.06418101489543915,
-0.06965110450983047,
-0.02380094863474369,
0.01979690231382847,
0.06394421309232712,
-0.006580080837011337,
0.009207532741129398,
0.07521732151508331,
0.0618942528963089,
0.02171708084642887,
-0.05540977790951729,
0.08883829414844513,
-0.060413651168346405,
-0.027448199689388275,
0.029966523870825768,
-0.0028822701424360275,
0.10442915558815002,
0.081257663667202,
0.026705419644713402,
-0.025940213352441788,
-0.0054184007458388805,
0.007912972941994667,
0.02020164765417576,
-0.09811757504940033,
0.04240913689136505,
0.029890989884734154,
-0.10343725234270096,
0.028926907107234,
0.01718778908252716,
-0.02392946369946003,
-0.007252719718962908,
-0.019619954749941826,
-0.05493656173348427,
-0.03826048597693443,
0.022287100553512573,
-0.007960470393300056,
-0.02263464592397213,
-0.03850642219185829,
-0.006981901824474335,
0.09065297245979309,
-0.014693979173898697,
0.06976041197776794,
-0.013296256773173809,
-0.030564045533537865,
0.03043445758521557,
-0.013579295948147774,
0.051542267203330994,
0.08525576442480087,
0.011518223211169243,
-0.055759888142347336,
0.03135926276445389,
0.0003471264790277928,
-0.011752118356525898,
0.014891504310071468,
-0.02490508370101452,
-0.021232569590210915,
0.054803796112537384,
-0.061860620975494385,
-0.009054870344698429,
0.05003533884882927,
-0.045009028166532516,
-0.020582199096679688,
-0.08526932448148727,
0.04055741801857948,
-0.04393232241272926,
0.024833127856254578,
-0.07367241382598877,
5.622398204355136e-34,
-0.00665662158280611,
0.021243300288915634,
0.037983473390340805,
-0.05631057545542717,
0.033470019698143005,
-0.003074533771723509,
0.040474556386470795,
-0.10040486603975296,
-0.031129959970712662,
0.034108828753232956,
-0.09839076548814774,
-0.020968208089470863,
-0.041320137679576874,
0.04098227992653847,
-0.00931147113442421,
-0.06189180165529251,
-0.000477785273687914,
0.03326618671417236,
-0.03076297603547573,
-0.024297913536429405,
0.05138838291168213,
-0.0683351680636406,
-0.0363478884100914,
-0.0765269473195076,
0.023431677371263504,
-0.0035348457749933004,
0.05320940166711807,
0.054131198674440384,
-0.019515423104166985,
-0.006316778715699911,
0.005159750580787659,
-0.05722733959555626,
-0.07233896851539612,
0.04702974483370781,
0.08127005398273468,
-0.07584352046251297,
0.07301289588212967,
0.004449819214642048,
-0.021651970222592354,
0.004132966510951519,
0.08045338839292526,
0.041070882230997086,
0.053434669971466064,
0.01595025137066841,
-0.06614817678928375,
0.036366138607263565,
-0.012308652512729168,
0.03828207403421402,
-0.0094527592882514,
-0.06969809532165527,
-0.05537905916571617,
-0.02876298315823078,
-0.012328186072409153,
-0.06482207775115967,
0.08625366538763046,
-0.03722785785794258,
0.03552878275513649,
0.0059355865232646465,
0.01837148144841194,
0.019001184031367302,
0.022054502740502357,
0.0067602284252643585,
-0.002118233125656843,
-0.022829802706837654,
0.042668748646974564,
0.07975342869758606,
-0.012210377492010593,
0.06214975193142891,
0.021871311590075493,
-0.059380222111940384,
0.008028947748243809,
0.022960683330893517,
-0.02940177731215954,
-0.03595563396811485,
0.13119906187057495,
-0.025098342448472977,
0.06387849152088165,
-0.043212294578552246,
-0.008842221461236477,
-0.002300830325111747,
-0.13219429552555084,
0.05801018327474594,
-0.08383769541978836,
0.06033959612250328,
-0.054088469594717026,
-0.10325987637042999,
-0.015901891514658928,
-0.03123074769973755,
-0.0009599714539945126,
-0.06908640265464783,
-0.15958857536315918,
0.0021583375055342913,
-0.017122428864240646,
-0.03615447133779526,
0.00474552670493722,
-7.909023539339902e-34,
-0.04446975886821747,
0.0819944217801094,
-0.037140168249607086,
0.045632168650627136,
-0.047563180327415466,
-0.04581185057759285,
-0.06102703511714935,
0.07917210459709167,
0.039523739367723465,
-0.060503795742988586,
-0.035094212740659714,
0.014184858649969101,
-0.0025707806926220655,
-0.011161086149513721,
0.05163975805044174,
0.05424908176064491,
-0.07169775664806366,
0.053478918969631195,
0.061709266155958176,
-0.03626969829201698,
-0.06923177093267441,
0.07838121056556702,
-0.013632961548864841,
0.012585503049194813,
-0.051033683121204376,
0.03448637202382088,
-0.01599917560815811,
0.07412360608577728,
-0.03795776888728142,
0.024646082893013954,
-0.06840978562831879,
-0.001516780466772616,
-0.10177136957645416,
0.04478729888796806,
-0.13493914902210236,
0.11125825345516205,
0.031125383451581,
0.01808982715010643,
0.018543120473623276,
0.11363530158996582,
-0.018240205943584442,
-0.014927097596228123,
0.014375346712768078,
-0.04574510455131531,
0.0054439641535282135,
-0.015426610596477985,
0.03357567638158798,
-0.004016286693513393,
-0.03926264867186546,
0.08270519226789474,
0.016107438132166862,
-0.008296745829284191,
0.012004614807665348,
0.040822938084602356,
-0.05576682090759277,
0.008457670919597149,
0.0018235777970403433,
-0.015396314673125744,
0.0991450846195221,
-0.10336682200431824,
-0.06295863538980484,
-0.1099088117480278,
0.014812403358519077,
0.03165693208575249,
-0.05071491375565529,
-0.004361086059361696,
-0.03255496174097061,
0.08834316581487656,
0.09300673007965088,
-0.04487689211964607,
0.011860948987305164,
0.013433278538286686,
0.058291319757699966,
0.08133258670568466,
0.0033421034459024668,
-0.07158011943101883,
0.01496822852641344,
-0.07976846396923065,
0.01509160827845335,
-0.04659273475408554,
-0.020793631672859192,
0.016807755455374718,
0.08813980221748352,
0.025319550186395645,
0.11643379926681519,
-0.0007234172080643475,
0.08284164220094681,
0.004233459942042828,
0.05776868388056755,
-0.03695249184966087,
0.047098610550165176,
0.04656210169196129,
0.07459012418985367,
0.020990699529647827,
-0.09356187283992767,
-5.013142612142474e-8,
-0.0857037752866745,
-0.002842402085661888,
0.01267955917865038,
0.0004423451900947839,
0.0873081237077713,
0.03641755133867264,
-0.024836240336298943,
0.044748228043317795,
-0.04134076461195946,
-0.03245196118950844,
0.023099850863218307,
0.010696719400584698,
-0.04634826257824898,
-0.0644906684756279,
0.007865091785788536,
0.07461892068386078,
-0.0011949519393965602,
0.04269763082265854,
-0.05383032560348511,
-0.002018602332100272,
0.006882588844746351,
0.10456682741641998,
0.0033263261429965496,
0.007058660965412855,
0.0735025629401207,
-0.05314626172184944,
-0.003334415378049016,
0.017881348729133606,
0.06002138927578926,
0.0114839943125844,
-0.030018024146556854,
0.04357951879501343,
0.002603787463158369,
0.07753773033618927,
-0.0032620250713080168,
0.04457519203424454,
0.013711215928196907,
0.012522927485406399,
-0.09316213428974152,
-0.019556956365704536,
-0.10860669612884521,
0.007525650784373283,
0.00021533014660235494,
-0.03195302560925484,
0.015494371764361858,
-0.020756516605615616,
0.014614811167120934,
-0.007595245726406574,
0.03860945627093315,
0.03704708442091942,
0.06839446723461151,
0.03712077438831329,
0.004166972823441029,
0.021283186972141266,
0.09473256766796112,
-0.0456061027944088,
-0.04426119849085808,
-0.09456979483366013,
0.05881744995713234,
0.04199707880616188,
-0.06806440651416779,
0.08018959313631058,
-0.009927771054208279,
-0.038646191358566284
] |
EdgeRunner: Auto-regressive Auto-encoder for Artistic Mesh Generation | https://openreview.net/forum?id=81cta3WQVI | [
"Jiaxiang Tang",
"Zhaoshuo Li",
"Zekun Hao",
"Xian Liu",
"Gang Zeng",
"Ming-Yu Liu",
"Qinsheng Zhang"
] | Poster | Current auto-regressive mesh generation methods suffer from issues such as incompleteness, insufficient detail, and poor generalization.
In this paper, we propose an Auto-regressive Auto-encoder (ArAE) model capable of generating high-quality 3D meshes with up to 4,000 faces at a spatial resolution of $512^3$.
We introduce a novel mesh tokenization algorithm that efficiently compresses triangular meshes into 1D token sequences, significantly enhancing training efficiency.
Furthermore, our model compresses variable-length triangular meshes into a fixed-length latent space, enabling training latent diffusion models for better generalization.
Extensive experiments demonstrate the superior quality, diversity, and generalization capabilities of our model in both point cloud and image-conditioned mesh generation tasks. | 3D Generation, Auto-regressive Mesh Generation | An auto-regressive auto-encoder that compresses variable-length meshes into fixed-length latent codes. | 1,123 | 2409.18114 | [
-0.0231939684599638,
-0.033936724066734314,
0.05141556262969971,
0.014112774282693863,
0.06579094380140305,
-0.001840193523094058,
-0.08045736700296402,
-0.12171201407909393,
0.028511008247733116,
0.01291196420788765,
-0.05647142231464386,
-0.04954202100634575,
0.007770612370222807,
0.09197510033845901,
0.008231289684772491,
0.014996222220361233,
0.020620297640562057,
0.06258136034011841,
-0.07637390494346619,
-0.008460721001029015,
-0.0041739256121218204,
0.012412747368216515,
-0.07818923890590668,
-0.07354506105184555,
0.12342999130487442,
0.007750510238111019,
0.1052742451429367,
-0.05415237694978714,
0.06211654096841812,
0.006493250373750925,
0.052583567798137665,
-0.0640096440911293,
-0.017849460244178772,
0.02330215647816658,
0.0014537905808538198,
0.025752710178494453,
-0.09853336960077286,
0.06306181102991104,
-0.03171931579709053,
0.0073588574305176735,
0.022083811461925507,
0.0707252025604248,
-0.02454450912773609,
0.00924686249345541,
0.09187248349189758,
0.00930428970605135,
-0.08434620499610901,
-0.03776028752326965,
0.006701871287077665,
-0.03277656063437462,
-0.0057984162122011185,
-0.07252141833305359,
-0.073942631483078,
-0.052202172577381134,
-0.02674386091530323,
-0.0011607257183641195,
0.05820419639348984,
-0.03448953106999397,
0.04570234566926956,
-0.012584573589265347,
-0.0012220254866406322,
-0.06768936663866043,
-0.006987167987972498,
-0.06132205203175545,
0.011183260008692741,
-0.07152280956506729,
0.02664947509765625,
0.025019284337759018,
-0.05666763707995415,
0.06468690186738968,
-0.027842387557029724,
-0.007128037977963686,
-0.0495154969394207,
0.03934916853904724,
0.04831521958112717,
0.031104033812880516,
0.027765169739723206,
-0.048481088131666183,
-0.034179627895355225,
0.0017066624714061618,
0.013842562213540077,
0.06189759075641632,
0.03280038386583328,
0.02519807405769825,
-0.0665331482887268,
0.031799305230379105,
-0.009161507710814476,
0.09299685060977936,
-0.036301389336586,
-0.00787391234189272,
-0.005637417081743479,
-0.016260825097560883,
-0.050350774079561234,
-0.034884948283433914,
0.0038214572705328465,
0.020045576617121696,
0.030411597341299057,
-0.09838539361953735,
0.0036163239274173975,
0.07718829065561295,
0.011685626581311226,
-0.0371401272714138,
0.01908576488494873,
0.004130842629820108,
0.06596112996339798,
0.0033077318221330643,
0.08388835191726685,
0.09440957009792328,
0.02611454576253891,
-0.012559176422655582,
0.03461163491010666,
0.004837943706661463,
-0.018701087683439255,
-0.0005920880939811468,
0.01255076564848423,
-0.04851755499839783,
-0.047344498336315155,
-0.04281041398644447,
0.10210218280553818,
0.06731971353292465,
0.002530662575736642,
0.006770963314920664,
-0.05966540426015854,
0.012923093512654305,
0.012745524756610394,
-0.0511804036796093,
-0.05160544812679291,
1.0391377408108714e-33,
0.0028623624239116907,
0.06576564162969589,
0.057290054857730865,
0.06094084307551384,
0.13916470110416412,
0.04430896043777466,
-0.006368399132043123,
-0.07829981297254562,
0.05532138794660568,
0.026253052055835724,
-0.14944109320640564,
-0.07753977179527283,
-0.06952790915966034,
0.14409175515174866,
0.010527897626161575,
-0.03304382041096687,
-0.058731332421302795,
0.008114428259432316,
0.00213775341399014,
0.08865848183631897,
0.045120708644390106,
-0.047136008739471436,
0.0005342014483176172,
-0.018142135813832283,
-0.10490058362483978,
0.023828301578760147,
0.06563394516706467,
-0.016248203814029694,
0.010210244916379452,
-0.0133266756311059,
-0.06389980763196945,
-0.004718214273452759,
0.03129154443740845,
0.0021460570860654116,
-0.0055940039455890656,
0.042934268712997437,
0.003929978236556053,
-0.017360063269734383,
-0.016158821061253548,
-0.04220617562532425,
0.08017680794000626,
0.1100175529718399,
-0.020827243104577065,
0.0044516753405332565,
-0.07155368477106094,
-0.020091470330953598,
0.030170103535056114,
0.05348183587193489,
-0.06839411705732346,
0.02347777783870697,
0.1065797433257103,
0.05802737921476364,
-0.028404196724295616,
-0.012448626570403576,
-0.04839363694190979,
-0.047867611050605774,
0.021168552339076996,
-0.08762934058904648,
0.08468855172395706,
0.010409594513475895,
0.04864450916647911,
-0.03268427774310112,
-0.02342930994927883,
0.010232323780655861,
-0.012830711901187897,
-0.05450666323304176,
0.09398414194583893,
0.05676042661070824,
-0.010790588334202766,
-0.012026860378682613,
-0.0500689260661602,
0.024859599769115448,
-0.017989719286561012,
-0.08674950152635574,
0.0016246051527559757,
-0.04926401749253273,
0.0815635696053505,
0.0073736985214054585,
0.00485294871032238,
0.010576830245554447,
-0.09029307216405869,
0.062330059707164764,
0.013132438994944096,
-0.0994839072227478,
-0.02936255745589733,
-0.04894661530852318,
0.014908898621797562,
-0.06811825931072235,
-0.013825295493006706,
-0.04877340793609619,
-0.010805349797010422,
-0.08177031576633453,
-0.001852407236583531,
0.0356055423617363,
-0.04235568642616272,
-6.939798600550675e-34,
0.0011738160392269492,
0.07250478863716125,
-0.05815376341342926,
0.033449459820985794,
0.03792895749211311,
-0.017167778685688972,
0.013058748096227646,
0.04191776365041733,
-0.031576428562402725,
-0.05543064326047897,
-0.010193455033004284,
0.02192971482872963,
0.027339467778801918,
-0.025647543370723724,
0.0589507594704628,
-0.02717783860862255,
0.01963138021528721,
-0.059207167476415634,
-0.054031576961278915,
-0.005461588967591524,
-0.011185802519321442,
0.016747495159506798,
0.01906980760395527,
0.008616777136921883,
-0.04308605194091797,
0.061165325343608856,
-0.04390386864542961,
0.07151768356561661,
-0.01922103948891163,
0.057175230234861374,
0.03984401375055313,
-0.02969515137374401,
0.0344696044921875,
0.03132745251059532,
-0.06267619878053665,
0.01972559094429016,
0.06298567354679108,
0.06274556368589401,
-0.011312033981084824,
0.06406150758266449,
0.04749780520796776,
-0.07189071178436279,
-0.037320930510759354,
0.1458868384361267,
-0.019443508237600327,
0.014379489235579967,
-0.02820221334695816,
-0.019769402220845222,
0.0031406402122229338,
0.047298677265644073,
-0.03776491433382034,
0.015168596059083939,
-0.06411672383546829,
0.01544866431504488,
-0.008924810215830803,
-0.10889942198991776,
-0.06487686932086945,
0.010333065874874592,
0.039093274623155594,
0.017651328817009926,
-0.06489154696464539,
-0.05827340856194496,
-0.0040460871532559395,
-0.008679335005581379,
-0.016606135293841362,
-0.058467887341976166,
-0.05590379610657692,
-0.025687528774142265,
0.0013006050139665604,
0.10935423523187637,
0.0135793536901474,
-0.00292384997010231,
-0.042194101959466934,
0.08242524415254593,
-0.00045030590263195336,
-0.09920988976955414,
0.050015274435281754,
0.024456432089209557,
0.059147655963897705,
-0.049485743045806885,
-0.022356392815709114,
0.02210334502160549,
0.028589457273483276,
0.0692124292254448,
0.04126707464456558,
0.02132601849734783,
0.055245909839868546,
0.04989602789282799,
-0.0018141487380489707,
0.07451970130205154,
-0.0017812824808061123,
0.02760005183517933,
-0.005169407930225134,
0.035134490579366684,
-0.038645412772893906,
-3.4577528396084745e-8,
-0.06294743716716766,
0.04681006446480751,
0.05905221775174141,
-0.03824963793158531,
-0.05898973345756531,
-0.0511227622628212,
0.0702039822936058,
0.04259826987981796,
0.019437845796346664,
-0.03678947314620018,
0.07898382097482681,
-0.1082552969455719,
-0.06818731874227524,
-0.021381786093115807,
0.008104954846203327,
0.09498944878578186,
-0.017557084560394287,
0.0576595701277256,
0.01826026849448681,
-0.013889623805880547,
-0.12803803384304047,
0.08820568770170212,
-0.02262536808848381,
-0.07321460545063019,
0.08351520448923111,
-0.058525171130895615,
-0.04570283740758896,
0.0817633643746376,
0.04164442420005798,
-0.02993619069457054,
-0.025420043617486954,
0.05758390203118324,
0.018810685724020004,
-0.042331840842962265,
-0.04741589352488518,
0.06010217219591141,
0.017095515504479408,
-0.06545775383710861,
-0.024361036717891693,
-0.11576608568429947,
-0.023027153685688972,
0.1055845245718956,
0.011830301955342293,
-0.06531252712011337,
-0.000730481930077076,
0.0840269923210144,
0.09455800801515579,
-0.04975242540240288,
0.018410738557577133,
0.0035373650025576353,
0.022777356207370758,
-0.024281907826662064,
0.0020303158089518547,
0.024080974981188774,
0.0472167544066906,
0.03384408354759216,
-0.03771997615695,
-0.019251050427556038,
0.11431165039539337,
-0.006052615121006966,
-0.04577726125717163,
0.012192018330097198,
-0.017876649275422096,
0.012652180157601833
] |
Beyond Random Augmentations: Pretraining with Hard Views | https://openreview.net/forum?id=AK1C55o4r7 | [
"Fabio Ferreira",
"Ivo Rapant",
"Jörg K.H. Franke",
"Frank Hutter"
] | Poster | Self-Supervised Learning (SSL) methods typically rely on random image augmentations, or views, to make models invariant to different transformations. We hypothesize that the efficacy of pretraining pipelines based on conventional random view sampling can be enhanced by explicitly selecting views that benefit the learning progress. A simple yet effective approach is to select hard views that yield a higher loss. In this paper, we propose Hard View Pretraining (HVP), a learning-free strategy that extends random view generation by exposing models to more challenging samples during SSL pretraining. HVP encompasses the following iterative steps: 1) randomly sample multiple views and forward each view through the pretrained model, 2) create pairs of two views and compute their loss, 3) adversarially select the pair yielding the highest loss according to the current model state, and 4) perform a backward pass with the selected pair. In contrast to existing hard view literature, we are the first to demonstrate hard view pretraining's effectiveness at scale, particularly training on the full ImageNet-1k dataset, and evaluating across multiple SSL methods, Convolutional Networks, and Vision Transformers. As a result, HVP sets a new state-of-the-art on DINO ViT-B/16, reaching 78.8% linear evaluation accuracy (a 0.6% improvement) and consistent gains of 1% for both 100 and 300 epoch pretraining, with similar improvements across transfer tasks in DINO, SimSiam, iBOT, and SimCLR. | Self-Supervised Learning, Data Augmentation, Pretraining | We propose Hard View Pretraining, a learning-free self-supervised method generating challenging samples based on the model’s state, improving multiple baselines by 1% linear eval. accuracy, including a 0.6% boost for DINO ViT-B/16, reaching 78.80%. | 1,120 | 2310.03940 | [
-0.07354116439819336,
-0.11709168553352356,
0.030857505276799202,
0.051658228039741516,
0.059170566499233246,
0.04221322014927864,
-0.04312701150774956,
-0.08854838460683823,
-0.006309741176664829,
-0.06680714339017868,
-0.07664841413497925,
-0.051901474595069885,
0.024009423330426216,
0.011314592324197292,
-0.07212738692760468,
-0.04551064223051071,
0.07775703817605972,
0.09351957589387894,
-0.1062258929014206,
0.007531596347689629,
-0.005699377506971359,
-0.04199714586138725,
0.026508547365665436,
-0.012841840274631977,
0.0761583223938942,
-0.06038736552000046,
0.048283591866493225,
-0.03690085560083389,
0.031900666654109955,
-0.0928729772567749,
0.0360199399292469,
-0.013084937818348408,
-0.04349286109209061,
0.018663497641682625,
-0.05488954111933708,
-0.0009306945721618831,
-0.029645081609487534,
0.06868863105773926,
0.012699334882199764,
-0.036010853946208954,
0.0014592546503990889,
-0.02857377380132675,
-0.023822102695703506,
-0.03299736976623535,
0.012544387951493263,
0.0029019492212682962,
0.05732160434126854,
-0.07335203886032104,
0.052352678030729294,
0.023569142445921898,
-0.09152542054653168,
-0.03465288132429123,
-0.04230758920311928,
0.024024417623877525,
-0.041088804602622986,
0.032411620020866394,
0.01624996028840542,
0.01641249470412731,
-0.019504854455590248,
0.06668712198734283,
0.04550237953662872,
-0.08142444491386414,
-0.024254780262708664,
0.04441126808524132,
-0.052856411784887314,
-0.011698458343744278,
0.076156847178936,
0.029175741598010063,
0.029677841812372208,
-0.021688096225261688,
-0.02845660224556923,
0.0806235820055008,
-0.027533255517482758,
-0.009628931991755962,
0.04814005270600319,
-0.0039010413456708193,
0.05176052451133728,
-0.08183799684047699,
0.049614813178777695,
-0.12611567974090576,
0.04641232267022133,
0.013875013217329979,
0.0051318877376616,
0.02996494434773922,
0.08393823355436325,
0.015079403296113014,
-0.04543018341064453,
0.02283802628517151,
0.019473982974886894,
0.013765143230557442,
-0.045138902962207794,
0.02358371950685978,
-0.028009267523884773,
-0.014720664359629154,
0.04138828068971634,
-0.006681348662823439,
0.03545849770307541,
-0.10324496775865555,
-0.04416798800230026,
0.07250184565782547,
0.026106949895620346,
0.0010846678633242846,
0.05848466232419014,
-0.002355229342356324,
0.05651630461215973,
-0.04368272423744202,
0.09404192864894867,
0.0421891026198864,
0.0408213771879673,
-0.029217546805739403,
0.07196758687496185,
0.051983792334795,
-0.005788914859294891,
-0.00960750225931406,
-0.005336552392691374,
0.06543479114770889,
-0.05815429985523224,
-0.03924668952822685,
-0.012633596546947956,
0.02402263693511486,
-0.05617433041334152,
0.004845532588660717,
0.03223069757223129,
0.02848670445382595,
0.030922146514058113,
-0.060898344963788986,
-0.05048852786421776,
2.3009172340743535e-33,
0.05139633268117905,
0.020700037479400635,
0.019043833017349243,
-0.037384480237960815,
0.05249869450926781,
-0.026452720165252686,
0.039585236459970474,
0.04275120794773102,
-0.0064911083318293095,
0.010205474682152271,
0.028193339705467224,
-0.04485246539115906,
-0.017459824681282043,
0.08852946758270264,
0.028003038838505745,
-0.03347013518214226,
-0.09965817630290985,
0.05119042098522186,
-0.03561834990978241,
0.023103173822164536,
-0.0011877442011609674,
-0.026252705603837967,
-0.0060492693446576595,
-0.0758831799030304,
-0.010133021511137486,
0.04295262694358826,
0.033551257103681564,
0.038460515439510345,
-0.05252012982964516,
0.030779333785176277,
-0.055973898619413376,
-0.018757594749331474,
0.00039553659735247493,
0.009554526768624783,
-0.00015025777975097299,
-0.018709909170866013,
0.04081558436155319,
0.002240241039544344,
0.0007514965254813433,
-0.01887025684118271,
-0.008062758482992649,
0.10434083640575409,
0.01323020365089178,
-0.027538977563381195,
-0.07917569577693939,
0.008858528919517994,
0.00553384143859148,
0.055276770144701004,
-0.132920041680336,
-0.024806637316942215,
0.07447018474340439,
-0.03720474615693092,
-0.061742812395095825,
-0.08631150424480438,
-0.09711829572916031,
0.018982110545039177,
0.045656077563762665,
0.040653932839632034,
0.05177968367934227,
0.08481492102146149,
-0.08069314062595367,
0.026466969400644302,
-0.06306656450033188,
0.14847812056541443,
-0.04456929862499237,
0.002094207564368844,
0.10618199408054352,
0.02856495976448059,
0.019829852506518364,
-0.033264681696891785,
-0.06939107179641724,
0.005118857137858868,
-0.011226124130189419,
-0.11849500238895416,
0.09820765256881714,
-0.004394476767629385,
0.04039355739951134,
-0.018952131271362305,
0.06744213402271271,
0.08402470499277115,
-0.018175404518842697,
0.05399954691529274,
0.03747950866818428,
-0.03398556634783745,
-0.04866041988134384,
0.019853275269269943,
0.02626434713602066,
-0.1155727431178093,
-0.016558146104216576,
0.0009735868661664426,
-0.024014020338654518,
-0.024677302688360214,
0.005456569138914347,
-0.04502318054437637,
0.01865607500076294,
-1.6182951749243736e-33,
-0.021438010036945343,
0.05741886422038078,
0.009988942183554173,
0.08051466196775436,
0.025953764095902443,
-0.06610175222158432,
0.021039778366684914,
0.07266144454479218,
0.03254742920398712,
-0.09419050067663193,
0.0002215223212260753,
-0.0355529710650444,
0.006497217807918787,
-0.005334826651960611,
0.03307453542947769,
-0.07110996544361115,
-0.03514609858393669,
0.0037091479171067476,
-0.07718513160943985,
0.04491749033331871,
0.038804344832897186,
0.11404341459274292,
-0.018025433644652367,
0.026836561039090157,
0.002360266400501132,
0.07363034039735794,
-0.018049782142043114,
0.09509290754795074,
0.06459428369998932,
-0.04658328741788864,
0.010662849992513657,
-0.008495204150676727,
0.009433990344405174,
-0.062211282551288605,
-0.08582010865211487,
0.11106611043214798,
0.02053668163716793,
0.04147404432296753,
-0.04240367189049721,
0.07681845873594284,
0.03984774649143219,
-0.03314516320824623,
-0.04244941100478172,
0.017127620056271553,
-0.05955878645181656,
-0.04946648329496384,
-0.012936300598084927,
0.025029858574271202,
-0.02607211284339428,
0.03307365998625755,
-0.07469982653856277,
0.05928463861346245,
-0.11336106806993484,
0.05558668076992035,
-0.01490450743585825,
-0.08532309532165527,
-0.008418955840170383,
0.01989295892417431,
0.05347975715994835,
0.03731263428926468,
-0.05856940522789955,
-0.05556762218475342,
-0.01589841954410076,
-0.03503906726837158,
0.014589400961995125,
-0.03374158963561058,
-0.10238169878721237,
-0.003501612925902009,
-0.004938675556331873,
0.07617174088954926,
-0.018611827865242958,
-0.030936039984226227,
-0.012769920751452446,
-0.014073414728045464,
0.009835679084062576,
-0.07670506089925766,
0.028335634618997574,
0.010186219587922096,
0.06730227917432785,
-0.013459921814501286,
0.027928780764341354,
-0.023077502846717834,
-0.024135911837220192,
0.05250522494316101,
0.10511885583400726,
0.048447560518980026,
0.03187333047389984,
0.005897463299334049,
0.0612221360206604,
-0.008873426355421543,
-0.08108002692461014,
0.038613755255937576,
0.014090478420257568,
0.0337204709649086,
-0.028938908129930496,
-3.692665373478121e-8,
-0.07970783114433289,
0.08549294620752335,
0.08380161970853806,
0.0634884238243103,
0.025308161973953247,
-0.005064297001808882,
-0.041156478226184845,
0.14160668849945068,
-0.024391861632466316,
-0.09293115139007568,
0.04544137045741081,
0.005118581932038069,
0.005831239745020866,
-0.01497693546116352,
-0.0423913300037384,
0.044738661497831345,
0.016974512487649918,
0.04108364135026932,
-0.010588555596768856,
-0.02044093981385231,
-0.012023513205349445,
-0.014645048417150974,
0.023713259026408195,
-0.06602911651134491,
0.011096331290900707,
-0.04613834619522095,
-0.024647096171975136,
0.026649579405784607,
0.08058904111385345,
-0.04017253220081329,
-0.034199051558971405,
0.0187413040548563,
0.024488816037774086,
0.031086144968867302,
0.030554521828889847,
0.06644594669342041,
-0.05274798721075058,
-0.01514842826873064,
-0.0051628705114126205,
0.020822999998927116,
-0.019905932247638702,
0.048863302916288376,
0.026115430518984795,
0.02137012965977192,
-0.07371093332767487,
-0.001468854257836938,
0.03466098755598068,
-0.09147734194993973,
0.05363873019814491,
-0.06379340589046478,
-0.03609814494848251,
-0.1039137989282608,
0.034061893820762634,
0.018974775448441505,
0.05700763687491417,
0.07320038974285126,
0.06257025897502899,
-0.03578563779592514,
0.0430283360183239,
0.15967011451721191,
0.024314315989613533,
0.011905187740921974,
-0.12406912446022034,
0.00814863946288824
] |
Accelerating Auto-regressive Text-to-Image Generation with Training-free Speculative Jacobi Decoding | https://openreview.net/forum?id=LZfjxvqw0N | [
"Yao Teng",
"Han Shi",
"Xian Liu",
"Xuefei Ning",
"Guohao Dai",
"Yu Wang",
"Zhenguo Li",
"Xihui Liu"
] | Poster | The current large auto-regressive models can generate high-quality, high-resolution images, but these models require hundreds or even thousands of steps of next-token prediction during inference, resulting in substantial time consumption. In existing studies, Jacobi decoding, an iterative parallel decoding algorithm, has been used to accelerate the auto-regressive generation and can be executed without training. However, the Jacobi decoding relies on a deterministic criterion to determine the convergence of iterations. Thus, it works for greedy decoding but is incompatible with sampling-based decoding which is crucial for visual quality and diversity in the current auto-regressive text-to-image generation. In this paper, we propose a training-free probabilistic parallel decoding algorithm, Speculative Jacobi Decoding (SJD), to accelerate auto-regressive text-to-image generation. By introducing a probabilistic convergence criterion, our SJD accelerates the inference of auto-regressive text-to-image generation while maintaining the randomness in sampling-based token decoding and allowing the model to generate diverse images. Specifically, SJD facilitates the model to predict multiple tokens at each step and accepts tokens based on the probabilistic criterion, enabling the model to generate images with fewer steps than the conventional next-token-prediction paradigm. We also investigate the token initialization strategies that leverage the spatial locality of visual data to further improve the acceleration ratio under specific scenarios. We conduct experiments for our proposed SJD on multiple auto-regressive text-to-image generation models, showing the effectiveness of model acceleration without sacrificing the visual quality. The code of our work is available here: https://github.com/tyshiwo1/Accelerating-T2I-AR-with-SJD/. | auto-regressive image generation, acceleration, training-free, image generation | null | 1,119 | 2410.01699 | [
-0.07897312194108963,
-0.07463112473487854,
0.019229840487241745,
-0.011240134947001934,
0.10646912455558777,
-0.021462813019752502,
-0.009485635906457901,
-0.025633489713072777,
0.042681723833084106,
0.005116773769259453,
-0.04469188302755356,
-0.001146335736848414,
0.044976815581321716,
0.027889668941497803,
-0.006356328260153532,
-0.016887594014406204,
0.01994679681956768,
0.10961546748876572,
-0.0681932270526886,
-0.11394060403108597,
0.03194554150104523,
-0.03907712921500206,
0.07938937097787857,
-0.03973067179322243,
0.1179409921169281,
0.03958389535546303,
0.0013708366313949227,
0.000677448813803494,
0.003820679849013686,
0.05031796917319298,
0.06300445646047592,
-0.04464806243777275,
-0.005699751898646355,
-0.054139502346515656,
-0.0003408067859709263,
0.10856684297323227,
-0.022804901003837585,
0.010925979353487492,
0.07014233618974686,
0.005315398797392845,
-0.008535753935575485,
0.04641425609588623,
0.046171072870492935,
0.014143061824142933,
0.08883735537528992,
0.00848422385752201,
-0.01035345159471035,
-0.06837490946054459,
0.06329952925443649,
0.010556953959167004,
-0.0599769651889801,
-0.0043632881715893745,
-0.034782957285642624,
-0.06180895119905472,
-0.018895551562309265,
0.03112865425646305,
0.01340752374380827,
-0.009206417016685009,
0.005421931855380535,
0.03588138148188591,
-0.05116352438926697,
-0.06977719813585281,
0.00445025647059083,
0.02195752039551735,
0.034491874277591705,
-0.041014548391103745,
0.06711634993553162,
0.045185625553131104,
-0.012371333315968513,
0.001976060913875699,
-0.08172200620174408,
0.03658754378557205,
-0.004118395037949085,
0.04914672300219536,
-0.00595643138512969,
0.045409902930259705,
0.007231485564261675,
-0.05995938926935196,
0.07858344912528992,
-0.07473933696746826,
0.03593399375677109,
-0.0021251055877655745,
0.09400638937950134,
0.024076608940958977,
0.019959211349487305,
-0.011967770755290985,
0.01027417927980423,
0.05614551901817322,
0.03520260006189346,
0.07329335808753967,
-0.13233034312725067,
-0.09750556945800781,
0.041277993470430374,
0.054846808314323425,
-0.04562072455883026,
0.04999645799398422,
-0.0133794154971838,
-0.05538726970553398,
-0.04100550338625908,
0.08284251391887665,
0.03422820195555687,
0.020738771185278893,
0.07148978114128113,
0.00700724683701992,
0.04876750335097313,
-0.06289447098970413,
0.10771644860506058,
0.08075027167797089,
0.03330717980861664,
-0.03187934309244156,
0.052649687975645065,
0.049581415951251984,
0.009125749580562115,
0.0006271814927458763,
-0.022409817203879356,
0.004867475014179945,
-0.03371536359190941,
-0.04924172908067703,
0.028978116810321808,
0.079585500061512,
-0.01678195223212242,
0.0009068408398889005,
-0.06500434130430222,
0.001078123808838427,
0.0027676692698150873,
-0.04684119299054146,
-0.0015128779923543334,
7.505761018740701e-35,
-0.020929008722305298,
0.09040655195713043,
0.10886022448539734,
0.041786421090364456,
0.003865305334329605,
-0.024351535364985466,
-0.07241500914096832,
-0.03484489396214485,
-0.05885658785700798,
0.0035541646648198366,
-0.05171527341008186,
-0.09005939215421677,
-0.070158451795578,
0.13693100214004517,
-0.030973447486758232,
-0.05409780517220497,
-0.05386830493807793,
-0.008787917904555798,
-0.006220645271241665,
0.05222014710307121,
0.03503299504518509,
-0.03497419133782387,
-0.023255134001374245,
-0.033995434641838074,
-0.026158899068832397,
-0.002364145824685693,
0.10967545956373215,
-0.008118395693600178,
-0.04357685148715973,
0.007530892267823219,
-0.10066327452659607,
-0.013912844471633434,
0.028222784399986267,
0.00011441124661359936,
0.06203969568014145,
-0.05989596247673035,
0.007921840995550156,
0.009307404048740864,
-0.09173890203237534,
-0.03383475914597511,
0.018859943374991417,
0.06054307147860527,
0.02656385861337185,
-0.015322570689022541,
-0.03856875002384186,
-0.05384538322687149,
0.0241505429148674,
-0.016103670001029968,
-0.0662757083773613,
0.015341460704803467,
0.06461170315742493,
0.016219612210989,
-0.08999878913164139,
-0.014021632261574268,
-0.021213265135884285,
-0.05014084652066231,
0.0916495993733406,
-0.026222163811326027,
0.07753938436508179,
0.048970501869916916,
-0.03721736744046211,
0.04217037558555603,
-0.02601230889558792,
0.07231546938419342,
-0.013781272806227207,
0.02883642539381981,
0.028725910931825638,
-0.04116106033325195,
-0.038364119827747345,
0.021745163947343826,
0.0001214171716128476,
0.02209535427391529,
-0.0007643201970495284,
-0.05263259634375572,
0.05890549346804619,
0.07393231987953186,
0.051577378064394,
-0.009226731956005096,
-0.03774992376565933,
0.04601375758647919,
-0.06248810887336731,
-0.04414478316903114,
0.00023730838438495994,
-0.053135234862565994,
-0.016565972939133644,
-0.03421841561794281,
0.029293637722730637,
-0.06752778589725494,
0.025082852691411972,
-0.05024219676852226,
-0.03716925531625748,
-0.05092569440603256,
-0.06658531725406647,
-0.03973979875445366,
0.011945007368922234,
-3.561014117377229e-34,
0.0035736113786697388,
0.04094797372817993,
-0.06262411177158356,
0.026662083342671394,
-0.03297905623912811,
-0.08803538978099823,
0.01798914186656475,
0.028724445030093193,
0.026410922408103943,
-0.11363623291254044,
-0.08238639682531357,
0.009900357574224472,
0.03943924233317375,
0.0346946120262146,
0.01147854421287775,
-0.00985585805028677,
0.024064943194389343,
-0.03585413843393326,
-0.06674394756555557,
0.00619228882715106,
-0.004750455729663372,
0.062601737678051,
0.0156027190387249,
0.013354825787246227,
0.04348596930503845,
0.08814603835344315,
-0.01836565136909485,
0.13074341416358948,
-0.012775485403835773,
-0.031206458806991577,
-0.006700312718749046,
-0.0338842049241066,
0.018340324983000755,
-0.0077161844819784164,
-0.08515802770853043,
0.0430833138525486,
0.030024979263544083,
0.01943540759384632,
-0.044450998306274414,
0.12953723967075348,
0.08915118128061295,
-0.06095189228653908,
-0.03595960512757301,
0.06702413409948349,
-0.03478367626667023,
-0.023510059341788292,
-0.05571452155709267,
-0.07493678480386734,
0.08767077326774597,
0.0680626705288887,
-0.028790157288312912,
0.06508966535329819,
-0.04268122836947441,
0.06087951362133026,
-0.003134176367893815,
-0.10679541528224945,
-0.13611064851284027,
0.02452899143099785,
0.04019579291343689,
-0.012603290379047394,
0.01789272390305996,
-0.054909560829401016,
-0.018318846821784973,
-0.0421716570854187,
0.013919807970523834,
-0.08513303101062775,
-0.05155566334724426,
-0.07321170717477798,
0.06891674548387527,
0.033417873084545135,
0.02043730579316616,
-0.06030239909887314,
-0.02817104198038578,
0.11083824187517166,
0.005836633965373039,
-0.04636332392692566,
0.010878042317926884,
0.03318478912115097,
0.006296806503087282,
-0.04529721289873123,
0.07320026308298111,
0.007642301730811596,
0.020815430209040642,
0.08837564289569855,
0.07787743955850601,
0.028285089880228043,
-0.028876762837171555,
-0.00478819664567709,
0.00045862363185733557,
-0.011306258849799633,
-0.014921760186553001,
0.07554066926240921,
0.010284102521836758,
0.012647438794374466,
-0.02131710946559906,
-3.552140981355478e-8,
-0.054491542279720306,
0.006356227211654186,
-0.0040528010576963425,
-0.006635392550379038,
0.021967321634292603,
-0.009356075897812843,
-0.04303513467311859,
-0.03735260292887688,
0.07993437349796295,
-0.09635673463344574,
0.10304275155067444,
-0.03612108156085014,
-0.046168480068445206,
-0.04543180763721466,
-0.022751834243535995,
0.05879893898963928,
0.029335778206586838,
-0.0158096756786108,
0.020796004682779312,
0.01777859404683113,
-0.04952026903629303,
0.025957291945815086,
-0.016268735751509666,
-0.08930116891860962,
0.00477713905274868,
-0.04594571888446808,
-0.00018661690410226583,
0.02753244899213314,
0.0771697387099266,
-0.05391271039843559,
-0.03692372515797615,
0.05067020654678345,
-0.025651363655924797,
-0.10968506336212158,
-0.0563012920320034,
0.020912431180477142,
0.02900085411965847,
-0.028041746467351913,
0.04141952469944954,
-0.12845833599567413,
0.04807404428720474,
0.033437807112932205,
-0.04867004230618477,
-0.007924068719148636,
0.015550442971289158,
-0.014246813021600246,
0.09287285059690475,
-0.040503498166799545,
0.0029233768582344055,
-0.04585588723421097,
0.006953075062483549,
-0.06356643885374069,
0.055324703454971313,
0.06006321683526039,
0.0032301987521350384,
-0.0630747452378273,
0.03613414242863655,
0.02256033569574356,
0.07211008667945862,
0.1434992253780365,
0.004844208713620901,
0.03843829408288002,
-0.06217130646109581,
-0.029120396822690964
] |
The Breakdown of Gaussian Universality in Classification of High-dimensional Linear Factor Mixtures | https://openreview.net/forum?id=UrKbn51HjA | [
"Xiaoyi MAI",
"Zhenyu Liao"
] | Poster | The assumption of Gaussian or Gaussian mixture data has been extensively exploited in a long series of precise performance analyses of machine learning (ML) methods, on large datasets having comparably numerous samples and features.
To relax this restrictive assumption, subsequent efforts have been devoted to establish "Gaussian equivalent principles" by studying scenarios of Gaussian universality where the asymptotic performance of ML methods on non-Gaussian data remains unchanged when replaced with Gaussian data having the *same mean and covariance*.
Beyond the realm of Gaussian universality, there are few exact results on how the data distribution affects the learning performance.
In this article, we provide a precise high-dimensional characterization of empirical risk minimization, for classification under a general mixture data setting of *linear factor models* that extends Gaussian mixtures.
The Gaussian universality is shown to break down under this setting, in the sense that the asymptotic learning performance depends on the data distribution *beyond* the class means and covariances.
To clarify the limitations of Gaussian universality in the classification of mixture data and to understand the impact of its breakdown, we specify conditions for Gaussian universality and discuss their implications for the choice of loss function. | High-dimensional statistics, random matrix theory, Gaussian universality, empirical risk minimization, mixture models, linear factor models | Establish asymptotic performance of ERM classifiers under linear factor mixture model, and derive conditions for Gaussian universality. | 1,113 | 2410.05609 | [
0.01201686356216669,
-0.14145734906196594,
0.009438985027372837,
-0.002803924959152937,
0.12111099064350128,
0.030792849138379097,
0.03501638025045395,
-0.06947283446788788,
-0.033232033252716064,
-0.05973665416240692,
0.09839478135108948,
0.030271651223301888,
0.04060067608952522,
-0.0035831041168421507,
-0.031019823625683784,
-0.10960233956575394,
0.12482420355081558,
0.07145079225301743,
-0.11813371628522873,
0.03176817297935486,
0.013288765214383602,
0.036736853420734406,
-0.022474929690361023,
-0.0051487889140844345,
0.030566761270165443,
-0.005823702551424503,
0.08221989870071411,
-0.049515124410390854,
0.004945379216223955,
-0.004646788351237774,
0.01687963306903839,
0.04199701547622681,
0.038491979241371155,
-0.00661763921380043,
-0.049975063651800156,
0.07123154401779175,
-0.0053776842541992664,
0.057967133820056915,
0.0006163276848383248,
0.0539524182677269,
0.0067144897766411304,
-0.00487142987549305,
-0.01281460840255022,
0.07420358061790466,
0.0551934540271759,
0.05524953827261925,
0.025339288637042046,
-0.14549772441387177,
0.017125442624092102,
0.018624627962708473,
0.0036457895766943693,
0.06508851796388626,
-0.07230430841445923,
0.044773124158382416,
-0.06380727887153625,
-0.07690556347370148,
0.036689214408397675,
0.02691824920475483,
-0.002773834392428398,
0.08012688159942627,
-0.04616594687104225,
-0.08339091390371323,
-0.0024378628004342318,
0.008980890735983849,
0.032287877053022385,
0.0013183874543756247,
-0.01646086573600769,
-0.017857838422060013,
-0.04034740477800369,
0.05814284086227417,
-0.04360943287611008,
0.09566060453653336,
-0.09570213407278061,
0.050786759704351425,
-0.00925815012305975,
0.03243052586913109,
-0.0020769541151821613,
0.0007234068471007049,
-0.024595018476247787,
0.0003948590892832726,
0.003151627955958247,
0.05961427465081215,
0.01384789403527975,
-0.07040049880743027,
0.1141824871301651,
0.01799304038286209,
-0.018569136038422585,
0.07713030278682709,
-0.017900967970490456,
-0.035932205617427826,
-0.0446605458855629,
0.037979304790496826,
0.040651921182870865,
-0.031036628410220146,
0.008689165115356445,
-0.009961552917957306,
0.013900794088840485,
-0.07483150064945221,
0.1106242835521698,
0.01942865550518036,
-0.04891297593712807,
-0.036026906222105026,
0.06166926026344299,
0.05149642378091812,
0.04281381517648697,
-0.07856612652540207,
0.019593074917793274,
-0.017443807795643806,
0.04824630916118622,
-0.049299612641334534,
-0.0038506658747792244,
0.04052024707198143,
0.005603463388979435,
-0.07197622209787369,
-0.060455072671175,
-0.016029248014092445,
0.046384699642658234,
-0.02154245786368847,
-0.033161796629428864,
0.06934688240289688,
0.03440083935856819,
0.031061535701155663,
0.07735303789377213,
0.012394006364047527,
0.032467979937791824,
0.061384011059999466,
-0.10678711533546448,
5.34278308761855e-33,
0.034144237637519836,
-0.0231538787484169,
0.008471169508993626,
0.0008729043183848262,
-0.05121466889977455,
0.029638923704624176,
-0.00870849285274744,
-0.000026560299374978058,
0.10978610813617706,
0.05934673920273781,
0.0016292056534439325,
0.098527692258358,
-0.045667972415685654,
0.08749649673700333,
0.004181832540780306,
0.1171434074640274,
-0.04392906650900841,
0.005380979739129543,
0.02634972333908081,
-0.07880046963691711,
0.08205261826515198,
-0.03878790885210037,
0.0316888727247715,
0.004129345063120127,
-0.09105295687913895,
-0.05381973460316658,
0.039201103150844574,
-0.056212443858385086,
-0.02895263396203518,
0.014513027854263783,
-0.08138769865036011,
0.04176275059580803,
-0.03559470921754837,
0.054936397820711136,
0.005551623646169901,
-0.02725905179977417,
-0.0024268091656267643,
-0.01573766954243183,
0.004502692259848118,
-0.07094690948724747,
-0.06599915027618408,
0.016093099489808083,
0.07187435775995255,
-0.05275418236851692,
-0.07945574820041656,
-0.017104683443903923,
0.0492149293422699,
0.0014015641063451767,
0.00632532499730587,
-0.07545343786478043,
-0.02347797155380249,
-0.08463548123836517,
-0.005944037809967995,
0.02584383450448513,
-0.05426469445228577,
0.052881814539432526,
-0.01085542980581522,
0.020313555374741554,
-0.025891108438372612,
0.009451311081647873,
-0.06987401843070984,
0.015056840144097805,
0.028613829985260963,
0.033112164586782455,
-0.03834458068013191,
-0.01463391538709402,
0.008739463984966278,
0.029376322403550148,
0.004696439951658249,
0.006200202275067568,
0.022053712978959084,
0.0139397531747818,
-0.06840918958187103,
-0.04523848742246628,
0.021385522559285164,
0.01564904674887657,
0.015471861697733402,
-0.03955765441060066,
-0.0334491990506649,
-0.007857831194996834,
-0.0638568252325058,
0.04090897738933563,
-0.013545935973525047,
-0.0953991711139679,
-0.07525802403688431,
0.042137324810028076,
-0.009266044944524765,
-0.04146084934473038,
-0.07730605453252792,
-0.03294183313846588,
-0.08500571548938751,
0.06596271693706512,
-0.015019841492176056,
-0.0035263269674032927,
-0.0040790908969938755,
-5.8467667396420055e-33,
-0.08954153954982758,
-0.014242378994822502,
0.018674947321414948,
0.07665091007947922,
0.05224107578396797,
-0.027781015262007713,
-0.013206816278398037,
-0.006692956201732159,
-0.03493449464440346,
0.040333401411771774,
0.01160117331892252,
-0.023587273433804512,
0.06338167935609818,
0.004826373420655727,
0.01167084090411663,
0.054096899926662445,
-0.05064161866903305,
0.1495196670293808,
0.0707358866930008,
-0.014013159088790417,
0.03767714649438858,
0.009692237712442875,
-0.046771347522735596,
-0.007877064868807793,
-0.004116576164960861,
-0.01446679886430502,
-0.08834321051836014,
0.06865783035755157,
-0.018979238346219063,
0.03966649994254112,
-0.07745467871427536,
0.04355699196457863,
-0.09935104101896286,
-0.02465505711734295,
-0.07312361896038055,
0.032968636602163315,
0.051168955862522125,
0.015665533021092415,
-0.043372515588998795,
0.0010825617937371135,
0.01618296280503273,
0.02591370977461338,
-0.045179806649684906,
-0.04133666679263115,
0.05672697722911835,
-0.07812903076410294,
0.03125280514359474,
-0.06890235096216202,
0.05393851548433304,
0.007755689788609743,
-0.030190737918019295,
0.04963626340031624,
-0.03815315663814545,
0.08677742630243301,
-0.030698541551828384,
0.05031281337141991,
0.011913931928575039,
-0.05641142278909683,
-0.012905391864478588,
0.004127889405936003,
0.03677554056048393,
0.011016900651156902,
-0.05975332111120224,
0.017816727980971336,
-0.02047448419034481,
-0.0642457976937294,
0.03873025253415108,
0.014622332528233528,
0.019121335819363594,
0.027149591594934464,
-0.04793205484747887,
-0.015901165083050728,
0.017859188839793205,
0.010390815325081348,
-0.07095746695995331,
0.04503569006919861,
0.07798773795366287,
-0.015105719678103924,
-0.028413284569978714,
0.005183993373066187,
0.11955583095550537,
-0.010898066684603691,
-0.0224900022149086,
0.028091995045542717,
0.034571193158626556,
0.10231132060289383,
0.006783220451325178,
-0.02294352278113365,
0.06518939882516861,
-0.09357137233018875,
-0.05298607796430588,
-0.0029660193249583244,
-0.02843446657061577,
0.10910230875015259,
-0.03532177954912186,
-4.434077638393319e-8,
-0.024413004517555237,
0.02143552526831627,
0.020338445901870728,
-0.13893842697143555,
0.08679112792015076,
0.014612725004553795,
-0.09634256362915039,
0.09632842242717743,
-0.024902435019612312,
-0.02296270616352558,
0.05436846241354942,
0.013782598078250885,
-0.03085319511592388,
-0.03800274804234505,
-0.007242489140480757,
0.035703144967556,
-0.046168435364961624,
-0.01605893112719059,
0.02796885184943676,
0.046489812433719635,
0.04669856280088425,
-0.048665158450603485,
0.009511462412774563,
-0.05232549086213112,
0.03104718215763569,
-0.10397451370954514,
0.04747190326452255,
0.0962415412068367,
0.025756381452083588,
0.02885206788778305,
-0.08573979884386063,
-0.026587635278701782,
-0.09996425360441208,
-0.06245462968945503,
-0.02198716625571251,
0.05994878709316254,
0.01994418539106846,
-0.03743835911154747,
-0.026332128793001175,
-0.04266977310180664,
0.0027306436095386744,
0.010796567425131798,
-0.011852447874844074,
0.0017312511336058378,
0.06053481996059418,
0.06702692806720734,
0.028082598000764847,
-0.08772145956754684,
0.06434989720582962,
0.04417577385902405,
-0.007810753770172596,
-0.009493482299149036,
0.02839815244078636,
0.10238039493560791,
0.07623909413814545,
-0.051258161664009094,
-0.002416748320683837,
-0.04914804548025131,
0.06111382320523262,
-0.09501450508832932,
0.05370934680104256,
0.030097443610429764,
-0.009157903492450714,
0.004905135370790958
] |
MoDGS: Dynamic Gaussian Splatting from Casually-captured Monocular Videos with Depth Priors | https://openreview.net/forum?id=2prShxdLkX | [
"Qingming LIU",
"Yuan Liu",
"Jiepeng Wang",
"Xianqiang Lyu",
"Peng Wang",
"Wenping Wang",
"Junhui Hou"
] | Poster | In this paper, we propose MoDGS, a new pipeline to render novel-view images in dynamic scenes using only casually captured monocular videos. Previous monocular dynamic NeRF or Gaussian Splatting methods strongly rely on the rapid movement of input cameras to construct multiview consistency but fail to reconstruct dynamic scenes on casually captured input videos whose cameras are static or move slowly. To address this challenging task, MoDGS adopts recent single-view depth estimation methods to guide the learning of the dynamic scene. Then, a novel 3D-aware initialization method is proposed to learn a reasonable deformation field and a new robust depth loss is proposed to guide the learning of dynamic scene geometry. Comprehensive experiments demonstrate that MoDGS is able to render high-quality novel view images of dynamic scenes from just a casually captured monocular video, which outperforms baseline methods by a significant margin. Project page: https://MoDGS.github.io | 3D Gaussian Splatting, Dynamic Novel-view Synthesis, Neural Rendering | null | 1,109 | null | [
-0.05203358456492424,
-0.1005406603217125,
0.06559162586927414,
-0.012996499426662922,
0.0690813660621643,
-0.04667414352297783,
-0.001344654243439436,
-0.0012621150817722082,
0.02257906273007393,
-0.05415690690279007,
-0.020929817110300064,
-0.008492312394082546,
-0.05879737809300423,
0.09239723533391953,
-0.03329599276185036,
-0.011813787743449211,
0.057315338402986526,
0.07992701977491379,
-0.07503479719161987,
0.036863408982753754,
0.028750289231538773,
-0.049562595784664154,
0.022025447338819504,
0.027527913451194763,
0.09406068176031113,
0.0130244679749012,
0.036636848002672195,
0.03343445435166359,
0.04330287501215935,
-0.012953593395650387,
0.010012373328208923,
-0.013103609904646873,
-0.05912276729941368,
-0.017305277287960052,
-0.03579607978463173,
0.0037003671750426292,
0.05790741369128227,
0.023519638925790787,
-0.028251787647604942,
-0.012797131203114986,
-0.04152457416057587,
0.09772451967000961,
0.0022314819507300854,
0.004068661946803331,
0.014765716157853603,
0.07216241955757141,
0.07174728065729141,
-0.08178263157606125,
0.037109408527612686,
-0.004973505157977343,
-0.04623529314994812,
0.007819874212145805,
0.02043835259974003,
0.030905408784747124,
0.042682986706495285,
0.0888623297214508,
-0.00732993520796299,
-0.09163972735404968,
0.008995776996016502,
0.005882741417735815,
0.022366637364029884,
-0.056053631007671356,
0.04064493998885155,
0.027046218514442444,
-0.045853789895772934,
-0.0015894201351329684,
0.0753975585103035,
-0.014098122715950012,
0.004828272853046656,
0.0204267930239439,
-0.11002156883478165,
0.03636664152145386,
-0.0013488158583641052,
-0.026349635794758797,
-0.05041998252272606,
0.006395642179995775,
0.0054557425901293755,
-0.09195704758167267,
-0.0018205324886366725,
-0.06430333107709885,
0.11926164478063583,
-0.009434911422431469,
0.013147903606295586,
0.03567548096179962,
-0.049546271562576294,
0.023085039108991623,
-0.00007742152956780046,
0.10826361179351807,
-0.005201996769756079,
0.037722114473581314,
-0.04719084873795509,
-0.022651366889476776,
-0.04053451865911484,
-0.062335580587387085,
-0.03127175197005272,
-0.07753738760948181,
-0.002935308963060379,
-0.13840296864509583,
0.04071389511227608,
0.058273669332265854,
-0.008112908340990543,
-0.0026233678217977285,
0.011824182234704494,
0.040384527295827866,
0.06561248004436493,
0.005749791394919157,
0.0264139324426651,
0.07723439484834671,
0.03154274448752403,
0.05199122428894043,
0.033539317548274994,
0.030794747173786163,
0.039759311825037,
0.015337872318923473,
-0.014681358821690083,
0.0747629925608635,
0.00916987843811512,
0.02768132835626602,
-0.05715188384056091,
0.0838884636759758,
0.010406791232526302,
-0.011009345762431622,
-0.08714541792869568,
0.0367276631295681,
0.03218250349164009,
-0.05743452161550522,
-0.006383726838976145,
3.224164272534602e-33,
0.0977320522069931,
0.031096836552023888,
0.06576625257730484,
0.08872489631175995,
-0.004874495789408684,
0.025626182556152344,
0.020522108301520348,
0.03645069897174835,
-0.044636573642492294,
0.013796285726130009,
-0.009408220648765564,
-0.10111834108829498,
0.011727963574230671,
0.0443117581307888,
-0.05377715080976486,
-0.024408044293522835,
-0.035578060895204544,
0.003263762453570962,
-0.01666359417140484,
0.12907572090625763,
-0.04499047249555588,
-0.002501640236005187,
-0.0021203448995947838,
-0.0325896255671978,
-0.04811187461018562,
0.0028949466068297625,
-0.015051339752972126,
0.022401021793484688,
0.030222473666071892,
-0.024118611589074135,
-0.048148397356271744,
-0.011741166934370995,
-0.06730365008115768,
-0.023166505619883537,
0.00865732878446579,
-0.045941516757011414,
-0.015446973033249378,
-0.06553085148334503,
-0.025960465893149376,
-0.035334959626197815,
-0.04394960030913353,
0.05920056253671646,
-0.012212594039738178,
0.0007031714776530862,
-0.08568552881479263,
0.03182562440633774,
0.05148780345916748,
0.0946941003203392,
-0.1186993196606636,
0.0693928599357605,
0.09408798068761826,
-0.009897072799503803,
-0.011114622466266155,
-0.07536067813634872,
-0.07739747315645218,
-0.010156490840017796,
0.015878252685070038,
-0.03534426540136337,
0.05238306522369385,
-0.001235215924680233,
0.004956621211022139,
0.009835933335125446,
-0.11592979729175568,
0.02064705453813076,
-0.07419958710670471,
0.003023371798917651,
0.030622029677033424,
0.04101981967687607,
0.02221503295004368,
0.020273784175515175,
-0.03729642927646637,
-0.02836204506456852,
0.00009073571709450334,
-0.005467107053846121,
0.03275338560342789,
0.05262153968214989,
0.010751704685389996,
-0.05638590827584267,
0.051754359155893326,
0.06417308002710342,
-0.06006987392902374,
0.05774342641234398,
0.11259226500988007,
-0.008270466700196266,
0.00823867879807949,
-0.039297446608543396,
0.02838265150785446,
-0.03664286807179451,
-0.013246551156044006,
-0.06697336584329605,
-0.0662495419383049,
-0.09701698273420334,
-0.03212753310799599,
-0.07771776616573334,
-0.005820252932608128,
-9.001550947868176e-34,
0.10620658099651337,
-0.0021881002467125654,
-0.11691241711378098,
0.05424438416957855,
-0.02957407757639885,
-0.0385977067053318,
-0.01817202940583229,
0.033062390983104706,
0.06213333085179329,
-0.11054805666208267,
-0.016351083293557167,
0.03361532837152481,
-0.022711750119924545,
0.017580734565854073,
0.029496463015675545,
-0.008123424835503101,
-0.04084188863635063,
-0.044778790324926376,
-0.035190168768167496,
-0.039077259600162506,
0.04762659966945648,
-0.03389694541692734,
-0.0019986345432698727,
-0.015259424224495888,
-0.05123114213347435,
0.13023166358470917,
-0.07779420912265778,
0.07358434796333313,
0.018398839980363846,
-0.022427115589380264,
0.005215758923441172,
-0.024071097373962402,
0.06609503924846649,
-0.03487589210271835,
-0.10744240134954453,
0.05014153569936752,
0.0021284909453243017,
0.01269641425460577,
-0.0032999692484736443,
-0.008937687613070011,
-0.038158610463142395,
0.029458587989211082,
-0.017550788819789886,
0.06592490524053574,
-0.007069757673889399,
0.011633363552391529,
0.04628666117787361,
-0.035737160593271255,
0.02490735985338688,
0.10689155012369156,
-0.14847922325134277,
0.0909038782119751,
-0.10500257462263107,
-0.013158498331904411,
0.007125644478946924,
-0.08177906274795532,
-0.1255943328142166,
-0.003513377858325839,
0.0723889097571373,
0.027451200410723686,
0.02632874995470047,
-0.008820821531116962,
-0.10440339893102646,
0.03869735822081566,
0.02214636839926243,
0.014210282824933529,
-0.0820593535900116,
-0.05219664424657822,
0.008407266810536385,
0.04354714974761009,
0.009158936329185963,
0.01867506466805935,
-0.017147231847047806,
-0.03339650109410286,
0.05079340934753418,
-0.011245468631386757,
-0.0018856849055737257,
-0.0060085728764534,
0.0414513498544693,
0.005545393563807011,
0.09459438174962997,
-0.06179433315992355,
0.06105630844831467,
0.14230002462863922,
0.06695827841758728,
0.09930986911058426,
-0.00870591215789318,
-0.011960839852690697,
0.06366854161024094,
0.015347071923315525,
-0.04062242433428764,
0.00982946902513504,
0.015095873735845089,
0.0017283853376284242,
0.03641253337264061,
-3.7251702167395706e-8,
-0.04328598454594612,
0.05288659781217575,
0.007506826426833868,
-0.013972784392535686,
-0.011549836955964565,
-0.034423355013132095,
0.04276837781071663,
0.0646078884601593,
0.024116678163409233,
-0.13512484729290009,
0.016382301226258278,
-0.0522843673825264,
0.0013649184256792068,
-0.014238902367651463,
-0.02004607766866684,
0.10335637629032135,
-0.018322652205824852,
-0.01753082126379013,
-0.06455834209918976,
0.1023162379860878,
-0.054820019751787186,
0.005233820527791977,
0.010411890223622322,
0.0063385427929461,
-0.01635795459151268,
-0.009510582312941551,
-0.015350287780165672,
0.032616548240184784,
0.12794078886508942,
-0.019887981936335564,
-0.017422590404748917,
-0.009382779709994793,
0.056385528296232224,
-0.040962040424346924,
-0.03317280486226082,
0.07222215831279755,
-0.009547711350023746,
0.01384789776057005,
0.061088304966688156,
-0.04509150609374046,
-0.07484466582536697,
-0.014841768890619278,
0.0367470346391201,
0.023286370560526848,
0.008246166631579399,
0.08160329610109329,
0.04754893481731415,
-0.0877743735909462,
-0.011098765768110752,
-0.045339975506067276,
0.016599662601947784,
0.031004512682557106,
-0.004471407271921635,
0.08581406623125076,
0.0957733690738678,
0.01529650017619133,
0.056988686323165894,
-0.022354284301400185,
0.0318651981651783,
0.0463566891849041,
-0.06742414832115173,
-0.0036865549627691507,
0.0032693345565348864,
-0.010785480029881
] |
HaDeMiF: Hallucination Detection and Mitigation in Large Language Models | https://openreview.net/forum?id=VwOYxPScxB | [
"Xiaoling Zhou",
"Mingjie Zhang",
"Zhemg Lee",
"Wei Ye",
"Shikun Zhang"
] | Poster | The phenomenon of knowledge hallucinations has raised substantial concerns about the security and reliability of deployed large language models (LLMs). Current methods for detecting hallucinations primarily depend on manually designed individual metrics, such as prediction uncertainty and consistency, and fall short in effectively calibrating model predictions, thus constraining their detection accuracy and applicability in practical applications. In response, we propose an advanced framework, termed HaDeMiF, for detecting and mitigating hallucinations in LLMs. Specifically, hallucinations within the output and semantic spaces of LLMs are comprehensively captured through two compact networks—a novel, interpretable tree model known as the Deep Dynamic Decision Tree (D3T) and a Multilayer Perceptron (MLP)—which take as input a set of prediction characteristics and the hidden states of tokens, respectively. The predictions of LLMs are subsequently calibrated using the outputs from the D3T and MLP networks, aiming to mitigate hallucinations and enhance model calibration. HaDeMiF can be applied during both the inference and fine-tuning phases of LLMs, introducing less than 2% of the parameters relative to the LLMs through the training of two small-scale networks. Extensive experiments conclusively demonstrate the effectiveness of our framework in hallucination detection and model calibration across text generation tasks with responses of varying lengths. | Large language model, knowledge hallucination, hallucination detection, model calibration, deep neural decision tree | This study presents a comprehensive framework, named HaDeMiF, for the detection and mitigation of hallucinations, which harnesses the extensive knowledge embedded in both the output space and the internal hidden states of LLMs. | 1,099 | null | [
0.006041019689291716,
-0.05831287428736687,
0.003349275328218937,
0.020176280289888382,
0.053261030465364456,
-0.006301646586507559,
-0.02942400611937046,
-0.04471995308995247,
0.08827183395624161,
-0.028100160881876945,
-0.1103193387389183,
-0.08075827360153198,
0.06963355094194412,
0.029031619429588318,
-0.03880493715405464,
0.022401759400963783,
0.027893131598830223,
0.10489323735237122,
-0.07982148230075836,
-0.07851430028676987,
0.09799890965223312,
0.039559271186590195,
-0.007223370485007763,
-0.0332866907119751,
0.06523209065198898,
0.02927989885210991,
-0.035739678889513016,
-0.030084846541285515,
-0.061112720519304276,
-0.031670693308115005,
0.024136479943990707,
0.05022914335131645,
0.015498831868171692,
-0.0008492924971506,
-0.034907322376966476,
0.022719483822584152,
-0.055284127593040466,
-0.004905045963823795,
-0.012146757915616035,
-0.05893605202436447,
-0.022684186697006226,
0.03988867998123169,
0.04055369645357132,
0.055937912315130234,
-0.009018605574965477,
0.033551037311553955,
-0.01857009530067444,
0.03792272508144379,
-0.05184486508369446,
-0.03481847792863846,
-0.02401447668671608,
-0.005099268164485693,
0.04412267729640007,
0.031217021867632866,
-0.057061318308115005,
0.02698863297700882,
0.016776874661445618,
0.038401663303375244,
-0.04590420052409172,
0.004322937224060297,
-0.06667756289243698,
-0.053254351019859314,
0.013541027903556824,
-0.038038406521081924,
0.026196062564849854,
0.07279670983552933,
-0.045874759554862976,
0.0350547693669796,
0.04028674215078354,
0.008708188310265541,
0.013133540749549866,
0.04557668790221214,
-0.096995048224926,
0.020567916333675385,
0.05316545441746712,
0.026130016893148422,
0.023508023470640182,
-0.06507068872451782,
0.11115557700395584,
-0.11606761068105698,
-0.0012046134797856212,
0.05158853530883789,
0.08255793899297714,
-0.003080495400354266,
0.0998612642288208,
0.004694709088653326,
-0.042668409645557404,
0.03771042078733444,
-0.03328800946474075,
0.0572064146399498,
-0.05578386411070824,
-0.160908505320549,
-0.06380850821733475,
0.02768752910196781,
0.018004337325692177,
0.015407444909214973,
-0.04329923540353775,
-0.05828490108251572,
-0.052021294832229614,
0.058280378580093384,
-0.02624511532485485,
0.027053527534008026,
0.016373982653021812,
-0.12232306599617004,
0.029164882376790047,
-0.028317390009760857,
0.062027398496866226,
0.005624343641102314,
0.04007327929139137,
-0.0508490689098835,
0.006905094720423222,
0.09837530553340912,
0.02476906217634678,
0.05306708440184593,
-0.010166119784116745,
-0.03461166098713875,
-0.01588469184935093,
-0.0671829804778099,
0.01210570428520441,
0.06764133274555206,
-0.02552119456231594,
0.030410347506403923,
0.05406632274389267,
0.015193759463727474,
0.03871599957346916,
-0.08584876358509064,
-0.09724108874797821,
2.6623753184674758e-33,
-0.006656178273260593,
0.01890682801604271,
-0.018381260335445404,
0.06537505239248276,
0.054779380559921265,
0.016373829916119576,
-0.04933035746216774,
0.01696067862212658,
0.04850999638438225,
-0.03727460652589798,
-0.09132102876901627,
-0.014663685113191605,
-0.024587735533714294,
0.023874839767813683,
0.04464709758758545,
0.026159700006246567,
-0.018824469298124313,
0.007769939489662647,
-0.052346885204315186,
-0.03186550363898277,
0.06427690386772156,
0.01389576867222786,
-0.012222619727253914,
-0.022000862285494804,
0.025118321180343628,
0.06668912619352341,
0.03216666728258133,
0.02970978245139122,
0.00026519550010561943,
0.03475578501820564,
-0.13019095361232758,
-0.01557565201073885,
0.003434443147853017,
-0.020190667361021042,
0.04885087534785271,
-0.019849495962262154,
-0.021322529762983322,
-0.0040698181837797165,
0.006195282097905874,
-0.0839378759264946,
-0.005264511331915855,
0.01572929322719574,
-0.050621140748262405,
-0.04388664662837982,
-0.07151170819997787,
0.01873362623155117,
0.00037360796704888344,
0.014871783554553986,
-0.05526910722255707,
-0.002955201780423522,
0.05304697901010513,
0.0006782942218706012,
0.001184888300485909,
-0.01625009998679161,
-0.07363148033618927,
0.025783885270357132,
0.012849475257098675,
0.0010515089379623532,
0.09528646618127823,
0.07167740166187286,
0.04017204791307449,
0.09551212936639786,
0.027767518535256386,
0.057634156197309494,
0.04548739641904831,
0.00776510639116168,
-0.03151186555624008,
0.004726933781057596,
0.014696603640913963,
-0.11439698934555054,
-0.004714593756943941,
-0.0388210266828537,
0.05683882161974907,
-0.022121842950582504,
-0.08999662846326828,
-0.033991336822509766,
0.03615523502230644,
0.02013310231268406,
-0.014903516508638859,
-0.014159416779875755,
0.022298237308859825,
-0.03783031180500984,
0.0004225066222716123,
-0.0497124157845974,
-0.010421414859592915,
-0.013529359363019466,
0.032137490808963776,
-0.10785641521215439,
-0.11864781379699707,
-0.009965946897864342,
-0.02182571217417717,
-0.055585965514183044,
0.05989062786102295,
0.006065863184630871,
-0.019094469025731087,
-4.449738356121765e-33,
-0.0704277828335762,
-0.008093069307506084,
0.052443377673625946,
0.0386693999171257,
-0.02954421192407608,
-0.04776742681860924,
-0.04267251864075661,
0.040181420743465424,
-0.02107541449368,
-0.04265081137418747,
-0.007443941198289394,
-0.017899632453918457,
-0.0017631265800446272,
-0.06426888704299927,
0.09086962044239044,
-0.03877504914999008,
0.00841333158314228,
0.044672079384326935,
0.00974031537771225,
0.15282003581523895,
-0.006929345428943634,
0.0819246768951416,
-0.1908782571554184,
0.004792693071067333,
-0.0337471105158329,
0.03162950649857521,
0.021026166155934334,
0.02301177754998207,
0.02551448345184326,
-0.01354547031223774,
-0.0076722996309399605,
0.04672963544726372,
-0.05858376994729042,
-0.0010795387206599116,
0.006053033750504255,
-0.03519769757986069,
0.03655795380473137,
-0.03828859701752663,
-0.06926516443490982,
-0.026467926800251007,
0.042957186698913574,
0.02097724750638008,
-0.12572485208511353,
-0.008702656254172325,
-0.013480059802532196,
0.02691180445253849,
-0.0690610334277153,
0.017290256917476654,
0.07488837093114853,
0.007457782980054617,
0.05822870880365372,
-0.04525307938456535,
-0.038143716752529144,
0.026103707030415535,
-0.09937430918216705,
-0.04128655046224594,
-0.020384041592478752,
0.01952970214188099,
0.08538944274187088,
0.04182558134198189,
-0.10737674683332443,
-0.05451836809515953,
0.018214860931038857,
-0.07375790178775787,
0.0011010910384356976,
-0.0033877904061228037,
-0.08856616169214249,
0.07757742702960968,
0.03645523265004158,
-0.001946018310263753,
0.04640498384833336,
-0.01061827503144741,
-0.018287841230630875,
0.10945399105548859,
-0.008678383193910122,
-0.04516567289829254,
-0.016303308308124542,
-0.09914921969175339,
0.013793748803436756,
-0.024045759811997414,
-0.030455077067017555,
-0.03554745391011238,
-0.015209744684398174,
0.07871399074792862,
0.10900283604860306,
-0.05473628267645836,
0.04765549302101135,
0.08259040117263794,
0.03039284609258175,
0.007029852829873562,
-0.03125068172812462,
0.024186989292502403,
-0.013053926639258862,
0.06392249464988708,
-0.040090519934892654,
-4.776037698661639e-8,
-0.07342696934938431,
-0.009966510348021984,
-0.0062628802843391895,
0.0039595989510416985,
0.025652796030044556,
-0.09467584639787674,
0.003768288530409336,
0.10294338315725327,
-0.001954521518200636,
-0.040812309831380844,
0.09608177095651627,
-0.05178140103816986,
-0.06045025587081909,
-0.11837198585271835,
0.021028997376561165,
0.08633434772491455,
-0.0077780974097549915,
0.06262797117233276,
-0.00836178194731474,
0.022882364690303802,
0.08737505227327347,
0.02853861264884472,
0.011940657161176205,
0.01206921599805355,
0.09040126204490662,
-0.02446589060127735,
-0.021478652954101562,
0.04353548586368561,
-0.029802296310663223,
0.046945542097091675,
0.04624698683619499,
0.042484015226364136,
0.026326332241296768,
0.03495953232049942,
-0.043161116540431976,
0.12864165008068085,
-0.025365233421325684,
-0.04989210516214371,
0.06878206878900528,
0.01802758499979973,
0.04448263347148895,
0.014422304928302765,
-0.03664392605423927,
0.04262584447860718,
-0.03126493841409683,
-0.05232862010598183,
0.03035147488117218,
-0.04770250990986824,
0.03640197217464447,
-0.0035623570438474417,
-0.02940886840224266,
0.09632370620965958,
-0.019468111917376518,
0.05524354800581932,
0.037479374557733536,
0.03296133503317833,
0.00904246885329485,
-0.024177640676498413,
0.02870706096291542,
0.057001277804374695,
0.08879711478948593,
-0.043803807348012924,
-0.089778833091259,
-0.02417636290192604
] |
LongPO: Long Context Self-Evolution of Large Language Models through Short-to-Long Preference Optimization | https://openreview.net/forum?id=qTrEq31Shm | [
"Guanzheng Chen",
"Xin Li",
"Michael Shieh",
"Lidong Bing"
] | Poster | Large Language Models (LLMs) have demonstrated remarkable capabilities through pretraining and alignment. However, superior short-context LLMs may underperform in long-context scenarios due to insufficient long-context alignment. This alignment process remains challenging due to the impracticality of human annotation for extended contexts and the difficulty in balancing short- and long-context performance. To address these challenges, we introduce LongPO, that enables short-context LLMs to self-evolve to excel on long-context tasks by internally transferring short-context capabilities. LongPO harnesses LLMs to learn from self-generated short-to-long preference data, comprising paired responses generated for identical instructions with long-context inputs and their compressed short-context counterparts, respectively. This preference reveals capabilities and potentials of LLMs cultivated during short-context alignment that may be diminished in under-aligned long-context scenarios. Additionally, LongPO incorporates a short-to-long KL constraint to mitigate short-context performance decline during long-context alignment. When applied to Mistral-7B-Instruct-v0.2 from 128K to 512K context lengths, LongPO fully retains short-context performance and largely outperforms naive SFT and DPO in both long- and short-context tasks. Specifically, LongPO-trained models can achieve results on long-context benchmarks comparable to, or even surpassing, those of superior LLMs (e.g., GPT-4-128K) that involve extensive long-context annotation and larger parameter scales. Our code is available at https://github.com/DAMO-NLP-SG/LongPO. | Long Context LLMs and Alignment, Large Language Models, Preference Optimization, Self-Evolution of LLMs | A new long-context alignment method that enables LLMs self-evolve to excel on long-context tasks without annotated data and short-performance drops. | 1,086 | 2502.13922 | [
-0.009139981120824814,
0.004156146198511124,
0.029962018132209778,
0.006239621434360743,
0.02090783230960369,
0.015643084421753883,
-0.04100753366947174,
-0.004276524297893047,
0.044941674917936325,
-0.058805759996175766,
-0.02351502515375614,
-0.08339688926935196,
0.0040442426688969135,
-0.048483707010746,
0.059385184198617935,
-0.021509308367967606,
0.0551108680665493,
0.04727881774306297,
-0.08379915356636047,
-0.015369446016848087,
0.030247673392295837,
0.0005027547595091164,
0.007378589827567339,
0.012154772877693176,
0.004375385120511055,
0.006931846961379051,
-0.031288690865039825,
-0.005431498400866985,
0.07844217866659164,
-0.06512083858251572,
0.054211124777793884,
0.09545641392469406,
0.038145557045936584,
0.04264557734131813,
-0.06234082952141762,
0.07759420573711395,
-0.12535496056079865,
-0.027781233191490173,
-0.005914449226111174,
-0.08777201920747757,
-0.0072150640189647675,
-0.004953842144459486,
0.04371418431401253,
0.059834275394678116,
-0.0024569707456976175,
-0.009660285897552967,
0.017783954739570618,
-0.014556010253727436,
0.01915351115167141,
0.051221784204244614,
-0.14435656368732452,
0.005913010332733393,
0.024311387911438942,
0.06606403738260269,
-0.03661840036511421,
0.04103867709636688,
0.0016138525679707527,
0.10147855430841446,
-0.009011797606945038,
0.03705867379903793,
-0.07694332301616669,
-0.09046869724988937,
-0.041194718331098557,
0.04760347679257393,
-0.037325575947761536,
-0.0139132309705019,
0.00721738813444972,
0.07501460611820221,
-0.05626120790839195,
0.02097255550324917,
-0.042814213782548904,
0.04852038994431496,
-0.05974716693162918,
0.13444015383720398,
0.03184417635202408,
0.10151120275259018,
-0.005290912929922342,
-0.013653747737407684,
0.01304851658642292,
-0.1029062494635582,
0.031658947467803955,
0.047170329838991165,
0.050034865736961365,
0.011837953701615334,
0.06934065371751785,
0.00012664686073549092,
0.007429169490933418,
0.03873132914304733,
0.014651913195848465,
0.010365140624344349,
-0.011929458007216454,
-0.1172100380063057,
0.04252505674958229,
-0.009875025600194931,
0.0321088470518589,
0.061679381877183914,
0.03131090849637985,
-0.010084243491292,
-0.10279785841703415,
0.04738079383969307,
0.059181030839681625,
0.04220504313707352,
-0.0585026741027832,
0.03715445101261139,
-0.048430319875478745,
-0.058878518640995026,
-0.02536415308713913,
0.014509362168610096,
0.0014435917837545276,
-0.08229687809944153,
0.025916652753949165,
0.03937365859746933,
0.015923917293548584,
-0.0004879895132035017,
0.019562941044569016,
-0.07907769083976746,
-0.0671369656920433,
-0.08995743840932846,
0.0032083396799862385,
0.031529754400253296,
-0.06338681280612946,
-0.034101102501153946,
-0.06476722657680511,
0.019579166546463966,
-0.04119044542312622,
0.008152578957378864,
-0.03875451162457466,
9.532521746134603e-34,
0.007576475851237774,
0.06952179968357086,
0.01926644891500473,
0.04141651466488838,
0.010763227008283138,
-0.0072656176052987576,
0.013381194323301315,
0.05376463010907173,
-0.09065460413694382,
0.011782517656683922,
0.02805989794433117,
0.022908980026841164,
-0.030472667887806892,
0.09001830965280533,
0.015535324811935425,
-0.01011695433408022,
-0.053476180881261826,
0.11512625217437744,
-0.08514668047428131,
0.009840675629675388,
0.04598463326692581,
-0.0032546077854931355,
0.011701044626533985,
-0.017177945002913475,
0.09246385097503662,
0.006450249347835779,
0.0760892778635025,
-0.043356772512197495,
-0.07946357131004333,
0.0015734781045466661,
-0.10643235594034195,
-0.0460362434387207,
0.02972872368991375,
0.007748445495963097,
0.047539226710796356,
-0.09765516966581345,
0.023366903886198997,
-0.044548965990543365,
-0.003775425022467971,
0.006479871924966574,
-0.0032650637440383434,
0.0577574297785759,
0.06169901043176651,
0.055065903812646866,
-0.09054739028215408,
-0.1051073670387268,
0.021784024313092232,
0.0900956317782402,
-0.01798774115741253,
-0.03360147029161453,
0.04188544675707817,
0.014363777823746204,
-0.017804116010665894,
-0.01261773332953453,
-0.03686663508415222,
0.011922480538487434,
0.04241708293557167,
0.029114648699760437,
0.016598695889115334,
0.05685381591320038,
-0.022932350635528564,
0.0686553567647934,
-0.03646734729409218,
0.1132315844297409,
0.06714737415313721,
0.00269096321426332,
0.005769531708210707,
0.04556852579116821,
0.07937190681695938,
-0.0716627836227417,
-0.01350653637200594,
-0.0636584460735321,
-0.00824467558413744,
0.0407196469604969,
0.012819310650229454,
-0.016171954572200775,
-0.0026723253540694714,
-0.047385163605213165,
-0.0022753283847123384,
0.02268645539879799,
0.02601529099047184,
0.013445599004626274,
-0.022735556587576866,
-0.025621091946959496,
-0.040631525218486786,
-0.0961523950099945,
0.03488245606422424,
-0.15529893338680267,
-0.023732591420412064,
0.010038191452622414,
0.007980612106621265,
-0.01389224361628294,
-0.015527259558439255,
-0.0042104702442884445,
0.004582734778523445,
-8.937875888252071e-34,
0.005021281074732542,
-0.04475013166666031,
-0.02587524615228176,
0.05094568803906441,
0.028139900416135788,
-0.019319260492920876,
0.016797108575701714,
0.0758834034204483,
-0.08773283660411835,
-0.024828476831316948,
-0.020481456071138382,
-0.008856788277626038,
0.02351655438542366,
-0.02129266783595085,
0.047510333359241486,
0.015159085392951965,
0.008139541372656822,
0.05669914558529854,
0.02436782233417034,
-0.017955999821424484,
0.04026859626173973,
0.0411599762737751,
-0.10722830146551132,
0.04550610110163689,
-0.013875648379325867,
0.01835743524134159,
-0.0220554918050766,
0.018232956528663635,
-0.030497398227453232,
0.021175306290388107,
-0.0698198601603508,
0.03765180706977844,
-0.07926124334335327,
-0.04231187328696251,
-0.01960630528628826,
0.036902595311403275,
-0.032649993896484375,
-0.008980798535048962,
0.02535264752805233,
0.04294548183679581,
0.08194586634635925,
-0.012701207771897316,
-0.003211762523278594,
-0.009365401230752468,
-0.011524537578225136,
-0.0487382709980011,
-0.11564291268587112,
-0.05950820818543434,
0.060120418667793274,
0.02210034243762493,
-0.030250851064920425,
0.08215942233800888,
-0.05273086950182915,
0.007976471446454525,
-0.05186992138624191,
-0.16716188192367554,
-0.040206462144851685,
-0.06256409734487534,
0.0385957695543766,
-0.0750211626291275,
-0.0243318323045969,
-0.05618423596024513,
0.10374965518712997,
-0.09685982763767242,
0.07648161053657532,
-0.017434095963835716,
0.03409089520573616,
-0.012996279634535313,
-0.01882179267704487,
-0.041587959975004196,
0.03414540737867355,
-0.04032677412033081,
-0.041684266179800034,
0.0005426772404462099,
-0.05696619674563408,
0.0031718246173113585,
-0.03870934993028641,
-0.043628666549921036,
-0.011975101195275784,
-0.031952302902936935,
0.004645019769668579,
0.0188408512622118,
-0.04885254055261612,
0.034348003566265106,
0.022229516878724098,
0.07977207005023956,
0.03287069872021675,
0.17945612967014313,
-0.0007388045196421444,
0.09351378679275513,
0.029110388830304146,
0.04623749107122421,
0.027853062376379967,
0.05741705745458603,
-0.0818679928779602,
-3.7799168239871506e-8,
-0.06730559468269348,
0.02581785060465336,
0.016309022903442383,
0.13636168837547302,
0.013380128890275955,
0.007569759152829647,
-0.034037746489048004,
0.10397037118673325,
0.023714685812592506,
0.028392046689987183,
-0.005159089341759682,
0.040415629744529724,
0.002689003711566329,
-0.02698022872209549,
0.005182374268770218,
0.05344364792108536,
-0.03577297553420067,
0.008525950834155083,
-0.027140246704220772,
0.015377979725599289,
0.050529904663562775,
0.07605203241109848,
-0.03434781730175018,
0.023133300244808197,
0.0274244025349617,
-0.07347699999809265,
-0.04400187358260155,
0.10185655951499939,
0.09828384220600128,
-0.055880431085824966,
-0.03702347353100777,
0.01921951211988926,
-0.08765792101621628,
-0.046685490757226944,
-0.04048142582178116,
0.017906824126839638,
0.026190584525465965,
0.023337898775935173,
0.03715125843882561,
-0.019182376563549042,
0.0003746003785636276,
0.06871230155229568,
-0.07760533690452576,
0.033148519694805145,
-0.000020666691852966323,
-0.012935549952089787,
-0.004851729609072208,
-0.0712694376707077,
-0.023407064378261566,
-0.06194441765546799,
0.02576710656285286,
-0.07030505686998367,
0.010012858547270298,
0.013023913837969303,
0.06247938424348831,
0.042584024369716644,
-0.04560188576579094,
0.033688727766275406,
0.12163715809583664,
0.06302883476018906,
-0.0021165625657886267,
0.022426649928092957,
-0.05122055485844612,
0.029352597892284393
] |
See What You Are Told: Visual Attention Sink in Large Multimodal Models | https://openreview.net/forum?id=7uDI7w5RQA | [
"Seil Kang",
"Jinyeong Kim",
"Junhyeok Kim",
"Seong Jae Hwang"
] | Poster | Large multimodal models (LMMs) "see" images by leveraging the attention mechanism between text and visual tokens in the transformer decoder. Ideally, these models should focus on key visual information relevant to the text token. However, recent findings indicate that LMMs have an extraordinary tendency to consistently allocate high attention weights to specific visual tokens, even when these tokens are irrelevant to the corresponding text. In this study, we investigate the property behind the appearance of these irrelevant visual tokens and examine their characteristics. Our findings show that this behavior arises due to the massive activation of certain hidden state dimensions, which resembles the attention sink found in language models. Hence, we refer to this phenomenon as the visual attention sink. In particular, our analysis reveals that removing the irrelevant visual sink tokens does not impact model performance, despite receiving high attention weights. Consequently, we recycle the attention to these tokens as surplus resources, redistributing the attention budget to enhance focus on the image. To achieve this, we introduce Visual Attention Redistribution (VAR), a method that redistributes attention in image-centric heads, which we identify as innately focusing on visual information. VAR can be seamlessly applied across different LMMs to improve performance on a wide range of tasks, including general vision-language tasks, visual hallucination tasks, and vision-centric tasks, all without the need for additional training, models, or inference steps. Experimental results demonstrate that VAR enables LMMs to process visual information more effectively by adjusting their internal attention mechanisms, offering a new direction to enhancing the multimodal capabilities of LMMs. | Large multimodal models, Visual attention sink, Visual attention redistribution | We find "visual attention sink" in LMMs, where irrelevant visual tokens with high attention weights arise from massive activation in specific dimensions. Also, we introduce Visual Attention Redistribution (VAR) to enhance attention on the image. | 1,085 | 2503.03321 | [
0.03006947971880436,
-0.08330737054347992,
0.060797419399023056,
0.008215771988034248,
0.1389719694852829,
-0.006956034805625677,
-0.012886254116892815,
0.0399496853351593,
0.10911925882101059,
-0.024445442482829094,
-0.011600201018154621,
-0.04058312624692917,
0.037014733999967575,
0.008600440807640553,
0.04024006426334381,
-0.013846242800354958,
0.05902399122714996,
0.07786554098129272,
-0.13651081919670105,
-0.04489006847143173,
0.08336098492145538,
-0.057967476546764374,
0.04866752400994301,
-0.029381917789578438,
0.0059089758433401585,
0.07321492582559586,
-0.046635061502456665,
-0.08932295441627502,
0.00958988070487976,
-0.046464864164590836,
-0.005030599422752857,
0.01955210603773594,
0.014356396161019802,
0.0647132471203804,
-0.09691247344017029,
0.13088621199131012,
-0.06954795122146606,
0.019690435379743576,
0.0005942180869169533,
-0.08396304398775101,
-0.02656395360827446,
0.049314554780721664,
0.0015063920291140676,
0.00886024534702301,
0.08816967904567719,
0.00449061393737793,
0.0347520187497139,
-0.021445946767926216,
-0.010532576590776443,
-0.11192193627357483,
-0.07074841111898422,
-0.0516437329351902,
-0.02891332469880581,
0.03774435818195343,
-0.051655903458595276,
0.011188658885657787,
0.06570042669773102,
-0.001714714802801609,
0.022447476163506508,
-0.008247570134699345,
-0.061062052845954895,
0.006354568526148796,
-0.0043050749227404594,
-0.043963704258203506,
0.05418741703033447,
-0.02456061914563179,
0.08418012410402298,
-0.07142866402864456,
-0.051529813557863235,
0.05284816026687622,
0.02288481593132019,
0.04502277076244354,
0.008765839971601963,
-0.03173605725169182,
0.06515197455883026,
0.040407221764326096,
0.048236947506666183,
-0.058857619762420654,
0.06599213182926178,
-0.07649584114551544,
0.07334944605827332,
-0.008515681140124798,
0.04708723723888397,
-0.042908307164907455,
0.12433074414730072,
0.022328674793243408,
-0.033154476433992386,
0.007841572165489197,
0.013233560137450695,
0.03665190935134888,
-0.0975521132349968,
-0.116504967212677,
0.04260077327489853,
-0.00004348475704318844,
0.04517296329140663,
-0.00583316944539547,
0.0013010052498430014,
-0.023686086758971214,
-0.0003990036784671247,
0.04993370920419693,
0.047029610723257065,
-0.026305556297302246,
0.039564669132232666,
-0.011230352334678173,
0.053613122552633286,
-0.020643701776862144,
0.07129520177841187,
0.04539523273706436,
-0.011649277992546558,
-0.055221084505319595,
0.051935307681560516,
0.010206053033471107,
-0.026368355378508568,
0.007402765564620495,
-0.005345126148313284,
-0.0061113969422876835,
0.03409164026379585,
-0.018004586920142174,
-0.0058246878907084465,
0.03223298862576485,
0.054294437170028687,
-0.0036413175985217094,
-0.028525255620479584,
-0.04461133852601051,
0.035047758370637894,
-0.023048095405101776,
-0.06383135914802551,
2.290817900890867e-33,
-0.025072505697607994,
0.011271907947957516,
0.012189655564725399,
0.02398899383842945,
0.0502922385931015,
-0.028461307287216187,
-0.01030777022242546,
-0.016932783648371696,
-0.016872115433216095,
-0.020636869594454765,
0.03353404253721237,
0.03709534928202629,
-0.04008256271481514,
0.17394083738327026,
-0.04711038991808891,
-0.07297838479280472,
-0.02414434216916561,
0.045906927436590195,
-0.02441173605620861,
0.0015679323114454746,
-0.004172642249614,
-0.003976692445576191,
-0.009101924486458302,
-0.02078082412481308,
-0.040376413613557816,
-0.018308566883206367,
0.02642674557864666,
-0.12060945481061935,
-0.07888440787792206,
0.020162975415587425,
-0.15291255712509155,
0.03251790627837181,
0.05058656632900238,
-0.01478551048785448,
-0.013946584425866604,
-0.12994617223739624,
-0.005664733238518238,
-0.045995984226465225,
0.0564180426299572,
-0.01525796577334404,
-0.018299540504813194,
0.038195278495550156,
-0.08348716050386429,
-0.0016703171422705054,
-0.04531668499112129,
0.013033868744969368,
0.00012644378875847906,
0.05814240127801895,
-0.04317070171236992,
-0.02834220789372921,
0.06547710299491882,
0.04617522656917572,
-0.023924225941300392,
-0.052410975098609924,
-0.008366483263671398,
-0.046189628541469574,
0.05602865293622017,
0.009786593727767467,
0.033704400062561035,
-0.028658395633101463,
-0.0207576472312212,
0.052753936499357224,
0.0191158689558506,
0.09784184396266937,
0.08538134396076202,
0.04043794050812721,
-0.000512773753143847,
0.0552336685359478,
0.014168632216751575,
0.033135611563920975,
-0.05475277453660965,
0.08209475129842758,
-0.041822586208581924,
-0.02705732360482216,
0.010727345012128353,
0.003285572864115238,
0.10756084322929382,
-0.036646001040935516,
0.01855812594294548,
0.0446365587413311,
0.021781854331493378,
-0.011322585865855217,
0.028094151988625526,
-0.06565113365650177,
-0.026979362592101097,
-0.018042942509055138,
0.023985175415873528,
-0.1406155526638031,
-0.05018430948257446,
-0.06548085063695908,
0.04898659884929657,
-0.004715575836598873,
-0.03208388015627861,
-0.07467085123062134,
-0.06362692266702652,
-2.1880049405351263e-33,
-0.0032776803709566593,
0.044056519865989685,
-0.03087184578180313,
0.03400355577468872,
-0.050156909972429276,
-0.03095724992454052,
0.022901352494955063,
-0.014523875899612904,
-0.03523470088839531,
-0.0598800890147686,
-0.06255092471837997,
-0.04069867357611656,
-0.03663848713040352,
0.0018447600305080414,
0.03664175793528557,
-0.038019195199012756,
0.11283060908317566,
0.012475905939936638,
-0.06637530773878098,
0.001761903869919479,
0.03996007889509201,
0.08245107531547546,
-0.010045132599771023,
-0.006479256320744753,
-0.08546385169029236,
0.055299170315265656,
-0.12770049273967743,
0.05167654529213905,
-0.015467898920178413,
-0.06662554293870926,
-0.050568822771310806,
0.019786331802606583,
0.008101657032966614,
-0.05823451653122902,
-0.04503784701228142,
0.0414293110370636,
-0.011891435831785202,
0.002919259946793318,
-0.036404769867658615,
0.1551186442375183,
0.05568182095885277,
-0.06767349690198898,
-0.03752032667398453,
0.0010132940951734781,
-0.07099440693855286,
-0.09238990396261215,
-0.08496203273534775,
-0.033302344381809235,
0.007670758292078972,
0.07761598378419876,
-0.004683070350438356,
0.024027878418564796,
-0.050196703523397446,
-0.026320988312363625,
-0.08493106067180634,
-0.04908985272049904,
-0.05476970225572586,
-0.011118462309241295,
0.06712056696414948,
-0.01674772799015045,
0.011101889424026012,
-0.041095539927482605,
-0.047499462962150574,
-0.058008428663015366,
0.0002587259514257312,
-0.024297339841723442,
-0.0008052958874031901,
-0.09590890258550644,
0.08321289718151093,
-0.032909974455833435,
0.06737484782934189,
-0.052346549928188324,
-0.01801067404448986,
0.021049458533525467,
0.022588061168789864,
0.08889937400817871,
0.05062616616487503,
0.014347738586366177,
0.012707185931503773,
-0.047193363308906555,
-0.024344071745872498,
-0.012045173905789852,
0.037892792373895645,
0.02091444842517376,
0.028264859691262245,
0.11902575194835663,
-0.012845422141253948,
0.014906114898622036,
0.02761184237897396,
-0.004388648550957441,
-0.036850593984127045,
0.008624663576483727,
0.018239445984363556,
0.0841580256819725,
0.010019088163971901,
-4.364774142118222e-8,
-0.07786504179239273,
-0.018743019551038742,
-0.00558658130466938,
0.017863590270280838,
0.02746417000889778,
-0.049007195979356766,
-0.02506263181567192,
0.08238954842090607,
-0.020204972475767136,
-0.0277092345058918,
0.040860023349523544,
-0.018418435007333755,
-0.031479354947805405,
-0.02735622599720955,
0.029058309271931648,
0.04254094138741493,
-0.013920754194259644,
0.01710568182170391,
-0.0376037172973156,
-0.04162431135773659,
-0.005956097040325403,
0.023802820593118668,
-0.0014519396936520934,
-0.019325032830238342,
-0.007203320041298866,
-0.010259543545544147,
-0.019045863300561905,
0.09724792838096619,
0.027423948049545288,
-0.02528141625225544,
0.05502167344093323,
0.05357978865504265,
-0.0458306185901165,
-0.062318138778209686,
0.027154656127095222,
0.030345609411597252,
0.0730210542678833,
-0.009291027672588825,
-0.03093722276389599,
0.04820787161588669,
0.09412792325019836,
0.026222115382552147,
-0.006186889950186014,
0.0970548465847969,
0.05087856575846672,
0.01100317481905222,
-0.02900089882314205,
-0.06080905720591545,
0.04290365055203438,
-0.09797633439302444,
0.018634255975484848,
-0.03623636066913605,
-0.038189053535461426,
0.13128617405891418,
0.004221971612423658,
-0.026177432388067245,
0.054638996720314026,
0.07981044054031372,
0.016034189611673355,
0.07400286942720413,
0.05999790132045746,
0.03126604110002518,
-0.03471887484192848,
-0.02146395482122898
] |
Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model | https://openreview.net/forum?id=PplM2kDrl3 | [
"Jincheng Zhong",
"XiangCheng Zhang",
"Jianmin Wang",
"Mingsheng Long"
] | Poster | Recent advancements in diffusion models have revolutionized generative modeling. However, the impressive and vivid outputs they produce often come at the cost of significant model scaling and increased computational demands. Consequently, building personalized diffusion models based on off-the-shelf models has emerged as an appealing alternative. In this paper, we introduce a novel perspective on conditional generation for transferring a pre-trained model. From this viewpoint, we propose *Domain Guidance*, a straightforward transfer approach that leverages pre-trained knowledge to guide the sampling process toward the target domain. Domain Guidance shares a formulation similar to advanced classifier-free guidance, facilitating better domain alignment and higher-quality generations. We provide both empirical and theoretical analyses of the mechanisms behind Domain Guidance. Our experimental results demonstrate its substantial effectiveness across various transfer benchmarks, achieving over a 19.6\% improvement in FID and a 23.4\% improvement in FD$_\text{DINOv2}$ compared to standard fine-tuning. Notably, existing fine-tuned models can seamlessly integrate Domain Guidance to leverage these benefits, without additional training. | transfer learning, diffusion models, fine-tuning, guidance | This paper proposes a conditional generation perspective for the transfer of diffusion models and derives a simple approach named Domain Guidance to enhance transfer learning. | 1,080 | null | [
-0.04170486703515053,
-0.12201423197984695,
0.023601358756422997,
0.042586248368024826,
0.10359067469835281,
-0.05624614655971527,
-0.04577214643359184,
0.0017975772498175502,
0.021007973700761795,
-0.02663395181298256,
-0.023687181994318962,
-0.07412935793399811,
0.024595828726887703,
0.012987742200493813,
0.00359162874519825,
-0.04004779830574989,
0.0490838959813118,
0.05191376432776451,
-0.123954176902771,
-0.022716503590345383,
-0.0036099040880799294,
0.07765446603298187,
0.0014589218189939857,
-0.029269207268953323,
0.0011216100538149476,
0.03504825383424759,
0.07695793360471725,
-0.02554674632847309,
0.02636083774268627,
-0.04734845086932182,
0.06094072386622429,
0.033189330250024796,
-0.09208942949771881,
-0.013293701224029064,
-0.027907712385058403,
0.069379523396492,
-0.10283858329057693,
0.01037390436977148,
0.04142744466662407,
-0.04096309468150139,
0.015524587593972683,
-0.05384302884340286,
0.05129706859588623,
0.044250741600990295,
0.03960698843002319,
0.000372593873180449,
-0.019733887165784836,
0.006500015966594219,
0.002913014031946659,
0.022496018558740616,
-0.03327788785099983,
-0.057136375457048416,
-0.05796758458018303,
0.07069383561611176,
-0.017881952226161957,
0.08661547303199768,
0.08478908985853195,
0.019438832998275757,
-0.008884941227734089,
-0.037823811173439026,
-0.1217174381017685,
-0.12576551735401154,
-0.03955047205090523,
-0.04831001162528992,
-0.05326902121305466,
-0.025868337601423264,
-0.05250786989927292,
0.1307283192873001,
-0.013500138185918331,
0.03845839574933052,
-0.05481073260307312,
0.09970657527446747,
-0.07410544902086258,
0.039029378443956375,
0.060316044837236404,
0.05796405300498009,
0.04717405512928963,
0.015598181635141373,
0.022458776831626892,
-0.09902962297201157,
0.016740255057811737,
0.03091319277882576,
0.10115876793861389,
-0.0013097089249640703,
0.048193417489528656,
0.03005167283117771,
0.03009353205561638,
0.02213316597044468,
0.022479694336652756,
-0.0014509613392874599,
-0.0008123028674162924,
0.0155160678550601,
0.013032416813075542,
0.0046166242100298405,
-0.06154096871614456,
0.07017911970615387,
-0.004971744492650032,
-0.09034571051597595,
-0.01846618577837944,
0.08081140369176865,
-0.011019740253686905,
0.07925093173980713,
0.01921628788113594,
-0.07545872032642365,
0.01818872056901455,
-0.10161449760198593,
0.07804366201162338,
0.10581430792808533,
0.006572504993528128,
-0.050452083349227905,
0.0038958608638495207,
0.06772281974554062,
0.012633669190108776,
-0.06305330991744995,
0.01518819760531187,
-0.07361513376235962,
-0.036879539489746094,
-0.03840562701225281,
0.018437474966049194,
-0.00711146229878068,
-0.056803248822689056,
0.042076192796230316,
-0.0649682804942131,
0.028714120388031006,
0.01461797021329403,
-0.061867065727710724,
-0.051787275820970535,
1.123793629714727e-33,
-0.022370440885424614,
0.019915714859962463,
-0.012451687827706337,
0.08382940292358398,
-0.02902425080537796,
-0.005386160220950842,
0.006935897283256054,
-0.03780525550246239,
0.01408747024834156,
-0.022185860201716423,
-0.026237910613417625,
0.011248590424656868,
-0.13072268664836884,
0.09943707287311554,
0.0430891215801239,
-0.016838734969496727,
-0.052742935717105865,
0.06193402782082558,
0.05889788642525673,
0.04449137672781944,
0.032074905931949615,
0.03346601501107216,
-0.03468703851103783,
-0.029029231518507004,
0.012566019780933857,
0.055346373468637466,
-0.0010441727936267853,
-0.042028699070215225,
-0.006397308316081762,
0.01852426305413246,
-0.09613129496574402,
-0.033156465739011765,
-0.02644548937678337,
-0.011556384153664112,
0.06695135682821274,
0.04774869605898857,
-0.02145286835730076,
-0.06128034368157387,
-0.024572396650910378,
-0.02366151660680771,
-0.017430249601602554,
0.032113853842020035,
0.0179620198905468,
0.04475679248571396,
-0.12595053017139435,
-0.018124980852007866,
0.019619345664978027,
-0.045671891421079636,
-0.022367995232343674,
-0.002515233587473631,
0.04717651754617691,
-0.07656107097864151,
0.024149807170033455,
-0.037003710865974426,
-0.07900360226631165,
-0.024754775688052177,
0.01824588142335415,
-0.035850320011377335,
-0.00046402314910665154,
-0.0015945896739140153,
0.08114677667617798,
0.0708465650677681,
-0.05063612386584282,
0.08775035291910172,
0.020138951018452644,
-0.023520834743976593,
-0.013436353765428066,
0.05583605170249939,
0.05351325497031212,
-0.0035595414228737354,
-0.015649843961000443,
0.00009460018918616697,
-0.05227939039468765,
-0.05861233174800873,
0.09152158349752426,
-0.04519543796777725,
0.0217940304428339,
-0.059889502823352814,
0.026784362271428108,
-0.019310971722006798,
-0.0461094044148922,
0.029284445568919182,
-0.07132583856582642,
-0.08019551634788513,
0.025752095505595207,
-0.010683857835829258,
0.0030819308012723923,
-0.05931250751018524,
-0.10259097069501877,
-0.030247272923588753,
-0.0029531391337513924,
-0.02117541804909706,
0.009456212632358074,
0.06857726722955704,
0.05378329008817673,
-1.1581210034946149e-33,
-0.004346934147179127,
0.07270169258117676,
0.08753640204668045,
0.072873555123806,
-0.0277685709297657,
0.013023220002651215,
-0.0003197038022335619,
0.033103104680776596,
0.04874901473522186,
-0.0878639742732048,
-0.038335464894771576,
-0.03168588504195213,
-0.018616655841469765,
-0.03088894672691822,
-0.019373811781406403,
0.0023090119939297438,
-0.03585980832576752,
-0.015943238511681557,
0.020139722153544426,
0.010895923711359501,
-0.02158423513174057,
0.076553113758564,
-0.08178693801164627,
0.07970127463340759,
-0.0742771178483963,
0.03623383119702339,
-0.02171418070793152,
0.1021319255232811,
-0.007249921094626188,
-0.03493751958012581,
-0.07038365304470062,
0.000826398259960115,
0.031221065670251846,
0.004512199200689793,
-0.08722716569900513,
0.08653479069471359,
0.07049768418073654,
0.023360125720500946,
-0.024430910125374794,
0.08636942505836487,
-0.007895312272012234,
-0.03923513740301132,
-0.09851495176553726,
0.042169198393821716,
-0.05435935780405998,
0.03219084441661835,
-0.024442283436655998,
0.04006389528512955,
0.07167398184537888,
0.018045561388134956,
0.013013784773647785,
0.11142401397228241,
-0.05586422234773636,
-0.04370087385177612,
-0.03359673172235489,
-0.08914533257484436,
0.009836101904511452,
-0.06532087922096252,
0.059365302324295044,
-0.014556776732206345,
-0.03895415738224983,
-0.01728341355919838,
0.03226223587989807,
-0.04490022733807564,
-0.03273583576083183,
-0.01955466717481613,
-0.004712292458862066,
0.04886152967810631,
0.020000800490379333,
0.0673985406756401,
-0.01102241687476635,
0.0010786651400849223,
-0.013590376824140549,
-0.03728137165307999,
-0.0029009226709604263,
-0.1309436708688736,
0.10542246699333191,
-0.04680219665169716,
0.02033214643597603,
-0.036857765167951584,
0.035673875361680984,
0.040455471724271774,
0.018305093050003052,
-0.022944724187254906,
0.07671122252941132,
0.029054680839180946,
0.0842193067073822,
-0.013138733804225922,
-0.030272014439105988,
0.016105541959404945,
-0.07250938564538956,
0.02019857056438923,
0.08376822620630264,
0.030508622527122498,
-0.15062600374221802,
-4.6614914595011214e-8,
-0.07183782756328583,
0.05819942429661751,
0.038665708154439926,
0.08768680691719055,
0.01836722157895565,
0.015210608951747417,
-0.005815594457089901,
0.09540305286645889,
-0.02638409100472927,
0.0022749167401343584,
0.0007620688993483782,
0.0650629997253418,
-0.011925234459340572,
0.009394790045917034,
0.03493808954954147,
0.049718331545591354,
0.026604725047945976,
0.07301843911409378,
-0.03525182604789734,
-0.014324265532195568,
0.02492816187441349,
0.028620930388569832,
0.012383277527987957,
-0.0412030890583992,
0.15498413145542145,
-0.06888457387685776,
0.0011632760288193822,
0.014865419827401638,
0.11532837152481079,
0.001955517567694187,
-0.015630943700671196,
0.008230265229940414,
-0.04068923369050026,
0.04641354829072952,
0.03397909924387932,
0.09217190742492676,
0.02255191095173359,
-0.03491651639342308,
-0.02465701848268509,
-0.039613738656044006,
-0.034153059124946594,
0.09073785692453384,
-0.07482428848743439,
0.01335255429148674,
-0.02412819117307663,
0.01056796032935381,
0.07129539549350739,
-0.03320684656500816,
0.030429579317569733,
-0.007273400668054819,
-0.002074358519166708,
-0.020314078778028488,
0.010587302036583424,
0.01812601089477539,
0.0205274298787117,
-0.003860193770378828,
0.03457672894001007,
0.017035724595189095,
0.03032141923904419,
0.033992934972047806,
0.014786548912525177,
-0.023941218852996826,
-0.02253245934844017,
0.036753129214048386
] |
MLLM can see? Dynamic Correction Decoding for Hallucination Mitigation | https://openreview.net/forum?id=4z3IguA4Zg | [
"Chenxi Wang",
"Xiang Chen",
"Ningyu Zhang",
"Bozhong Tian",
"Haoming Xu",
"Shumin Deng",
"Huajun Chen"
] | Poster | Multimodal Large Language Models (MLLMs) frequently exhibit hallucination phenomena, but the underlying reasons remain poorly understood. In this paper, we present an empirical analysis and find that, although MLLMs incorrectly generate the objects in the final output, they are actually able to recognize visual objects in the preceding layers. We speculate that this may be due to the strong knowledge priors of the language model suppressing the visual information, leading to hallucinations. Motivated by this, we propose a novel dynamic correction decoding method for MLLMs DeCo, which adaptively selects the appropriate preceding layers and proportionally integrates knowledge into the final layer to adjust the output logits. Note that DeCo is model agnostic and can be seamlessly incorporated with various classic decoding strategies and applied to different MLLMs. We evaluate DeCo on widely-used benchmarks, demonstrating that it can reduce hallucination rates by a large margin compared to baselines, highlighting its potential to mitigate hallucinations. Code is available at https://github.com/zjunlp/DeCo. | Hallucination Mitigation, Multimodal Large Language Models, Decoding Strategy | A dynamic correction decoding method for MLLMs (Deco), which adaptively selects the appropriate preceding layers and proportionally integrates knowledge into the final layer to adjust the output logits for hallucination mitigation. | 1,076 | 2410.11779 | [
-0.007023266516625881,
-0.06343734264373779,
0.02079000137746334,
0.02229670062661171,
0.059526339173316956,
-0.033748429268598557,
-0.053669460117816925,
-0.08014829456806183,
0.03683473542332649,
-0.03077581711113453,
-0.08700054883956909,
-0.023331811651587486,
0.04156176373362541,
0.01823112927377224,
-0.043756429105997086,
-0.01520390436053276,
0.023249221965670586,
0.17334522306919098,
-0.07697931677103043,
-0.05742558091878891,
0.04970252141356468,
0.031989824026823044,
-0.008320024237036705,
-0.0011791305150836706,
0.10036525875329971,
0.0294438898563385,
-0.028300803154706955,
-0.016996320337057114,
0.02500472217798233,
-0.08717263489961624,
0.03693864867091179,
0.0707845613360405,
0.01916179060935974,
0.041602443903684616,
-0.038303621113300323,
0.03248758614063263,
-0.03446343168616295,
0.018851149827241898,
0.033995602279901505,
-0.04898017272353172,
-0.060854457318782806,
0.03895971179008484,
0.009102457202970982,
-0.03762059286236763,
0.004654353018850088,
0.011185879819095135,
-0.01798829436302185,
-0.03842933848500252,
0.009575963020324707,
-0.044811028987169266,
-0.035856276750564575,
0.010044465772807598,
0.045235004276037216,
0.04255254939198494,
-0.07243825495243073,
0.039997316896915436,
0.046793799847364426,
0.008612207137048244,
-0.015096045099198818,
0.02512088604271412,
-0.0724833756685257,
-0.029153648763895035,
-0.01751219481229782,
0.008781282231211662,
-0.0059562018141150475,
0.039783403277397156,
0.024721186608076096,
0.022081425413489342,
0.029228361323475838,
0.01734927110373974,
-0.020973747596144676,
0.01832742616534233,
-0.027321364730596542,
0.007189204450696707,
0.06460484862327576,
0.039047084748744965,
-0.018020160496234894,
-0.10947731882333755,
0.06089742109179497,
-0.12046749889850616,
0.03452237322926521,
0.0468762069940567,
-0.003218940459191799,
0.0018761121900752187,
0.13773667812347412,
-0.0396893247961998,
0.009725299663841724,
0.0408763587474823,
0.016657069325447083,
0.0650402382016182,
-0.0822799876332283,
-0.11419840902090073,
0.054605040699243546,
0.01335203368216753,
0.006272430997341871,
0.03528576344251633,
0.04954811558127403,
-0.04765470325946808,
-0.009341901168227196,
0.06334296613931656,
-0.014954381622374058,
0.02474856935441494,
-0.015047743916511536,
-0.07515955716371536,
-0.017628801986575127,
-0.06407129019498825,
0.16020874679088593,
0.12755650281906128,
-0.0042425841093063354,
-0.05290551856160164,
0.028796439990401268,
0.05563242360949516,
-0.010086718015372753,
-0.000028692365958704613,
0.03882633149623871,
-0.02148192934691906,
0.04754069447517395,
-0.031208261847496033,
-0.009899670258164406,
0.051210928708314896,
0.017302563413977623,
0.009257877245545387,
0.0014604413881897926,
-0.02762649580836296,
0.032667603343725204,
-0.050010956823825836,
-0.04577038064599037,
-2.1085106197750825e-34,
0.057861097157001495,
0.04088616371154785,
0.004235214088112116,
0.05796504765748978,
0.061174847185611725,
0.014238594099879265,
-0.04890304058790207,
-0.07014656066894531,
0.05339338630437851,
-0.005713368766009808,
-0.08398975431919098,
-0.014090071432292461,
-0.07809270918369293,
0.07784567773342133,
-0.034138236194849014,
0.043600331991910934,
-0.05397064983844757,
0.03995627909898758,
-0.0893867164850235,
0.032778386026620865,
0.03662198781967163,
-0.013445878401398659,
-0.03237414360046387,
-0.04841084033250809,
-0.015197716653347015,
0.0941973328590393,
0.04348864406347275,
-0.1032174602150917,
-0.0421370230615139,
0.05317194014787674,
-0.1250213086605072,
0.023394476622343063,
0.005437720101326704,
0.010515394620597363,
0.0586906298995018,
-0.02086932212114334,
0.021319735795259476,
-0.028211593627929688,
-0.02690785564482212,
-0.030012719333171844,
-0.04995452240109444,
0.031417082995176315,
-0.06286196410655975,
-0.04409653693437576,
-0.06936334818601608,
0.010366779752075672,
-0.000650474801659584,
0.018568282946944237,
-0.09591200202703476,
-0.019442960619926453,
0.08218485862016678,
0.05079459398984909,
-0.03467473015189171,
-0.017680754885077477,
-0.03797053173184395,
-0.0602702870965004,
0.043390970677137375,
-0.04780267924070358,
0.06956297904253006,
0.03633803501725197,
0.0580657534301281,
0.08650772273540497,
-0.014321654103696346,
0.09770434349775314,
0.03969983756542206,
-0.0071316310204565525,
0.027809536084532738,
-0.016404829919338226,
-0.017427878454327583,
-0.07720374315977097,
-0.09914356470108032,
-0.004819351248443127,
0.021375739946961403,
-0.015931518748402596,
-0.015627916902303696,
-0.01344747468829155,
0.0347757488489151,
-0.03263184800744057,
0.02676924504339695,
0.024214740842580795,
0.04794779047369957,
-0.03821149095892906,
-0.018554527312517166,
-0.07657650858163834,
-0.0032174333464354277,
-0.04725879430770874,
0.022896984592080116,
-0.07272113114595413,
-0.10553368180990219,
-0.07039733231067657,
-0.002668403321877122,
-0.06770738959312439,
0.05930455029010773,
-0.06922394037246704,
-0.020137030631303787,
-1.2086413629700567e-33,
0.02777264267206192,
0.0017712614499032497,
0.051043227314949036,
0.02371012046933174,
-0.05373890697956085,
-0.08209661394357681,
0.015400734730064869,
0.025545969605445862,
0.018475275486707687,
-0.0713464543223381,
-0.01713770255446434,
-0.08091825246810913,
-0.014774165116250515,
0.007114762905985117,
0.091347835958004,
-0.055705320090055466,
0.03333655744791031,
-0.01658177748322487,
0.004062405321747065,
0.06244752183556557,
-0.017267612740397453,
0.0957503467798233,
-0.11761995404958725,
-0.03033985011279583,
-0.002902774838730693,
0.07519755512475967,
-0.08055835962295532,
0.08083615452051163,
0.07016117125749588,
-0.007682173512876034,
0.023251885548233986,
0.07894432544708252,
-0.011768318712711334,
-0.037093598395586014,
-0.013093036599457264,
0.06470610946416855,
0.010758880525827408,
-0.03919364884495735,
-0.06728419661521912,
0.05728700011968613,
0.06401555240154266,
-0.028366683050990105,
-0.08656943589448929,
-0.012636822648346424,
0.0337352491915226,
-0.0792631134390831,
-0.10854160785675049,
-0.011588520370423794,
0.027916254475712776,
0.010369065217673779,
0.03416753187775612,
0.05751071870326996,
-0.10360733419656754,
-0.02767491154372692,
-0.0009347248123958707,
-0.06268322467803955,
-0.11095001548528671,
0.023810213431715965,
0.06573908776044846,
0.031474895775318146,
-0.07145807147026062,
-0.042822111397981644,
-0.00500360457226634,
-0.021039951592683792,
-0.01314141508191824,
0.0012287605786696076,
0.004063344094902277,
-0.027205277234315872,
0.031032171100378036,
-0.00355140189640224,
0.06289058178663254,
-0.0864640325307846,
-0.04333386570215225,
0.08828715980052948,
0.0445544570684433,
-0.04238944500684738,
-0.028811998665332794,
0.004802465904504061,
0.004619872197508812,
-0.024834321811795235,
-0.034404072910547256,
-0.07846947759389877,
0.035515159368515015,
0.08381573110818863,
0.09973753988742828,
0.020876552909612656,
-0.03233752399682999,
0.05276699364185333,
-0.009595728479325771,
-0.017518511041998863,
-0.03643117845058441,
-0.0074254246428608894,
0.024772612378001213,
0.08854274451732635,
-0.02809871733188629,
-4.160985156431707e-8,
-0.09929871559143066,
0.004837284795939922,
-0.007343737408518791,
-0.0012645371025428176,
0.040474504232406616,
-0.03238813951611519,
-0.04640456289052963,
0.07769312709569931,
0.011868923902511597,
-0.08825966715812683,
0.10685930401086807,
-0.02746308036148548,
-0.005604094360023737,
-0.0857924297451973,
-0.007088378071784973,
0.1286820024251938,
-0.05462869256734848,
0.04423138126730919,
-0.022860553115606308,
0.0016225669533014297,
0.06074169650673866,
0.022282065823674202,
-0.04685322567820549,
-0.04905426874756813,
0.0067145791836082935,
-0.028432266786694527,
-0.008108372800052166,
0.0772845596075058,
0.00397899653762579,
-0.04301142320036888,
0.04309656471014023,
0.06583717465400696,
0.06927823275327682,
-0.016195662319660187,
-0.008208030834794044,
0.048229608684778214,
0.03388967737555504,
-0.04503824934363365,
0.007738595362752676,
0.020680582150816917,
0.03513545170426369,
-0.0013315421529114246,
-0.04903416335582733,
0.0075017292983829975,
0.025118310004472733,
-0.015886817127466202,
0.05112815275788307,
-0.07513954490423203,
-0.013895793817937374,
-0.03355114534497261,
-0.004361740313470364,
0.07550676167011261,
0.0027373498305678368,
0.09309044480323792,
0.1004599928855896,
0.02304426021873951,
0.0744326189160347,
-0.010838369838893414,
0.05126470699906349,
0.05055716261267662,
-0.012937015853822231,
0.02315298467874527,
-0.0502762533724308,
-0.008511574938893318
] |
GDrag:Towards General-Purpose Interactive Editing with Anti-ambiguity Point Diffusion | https://openreview.net/forum?id=8G3FyfHIko | [
"Xiaojian Lin",
"Hanhui Li",
"Yuhao Cheng",
"Yiqiang Yan",
"Xiaodan Liang"
] | Poster | Recent interactive point-based image manipulation methods have gained considerable attention for being user-friendly. However, these methods still face two types of ambiguity issues that can lead to unsatisfactory outcomes, namely, intention ambiguity which misinterprets the purposes of users, and content ambiguity where target image areas are distorted by distracting elements. To address these issues and achieve general-purpose manipulations, we propose a novel task-aware, training-free framework called GDrag. Specifically, GDrag defines a taxonomy of atomic manipulations, which can be parameterized and combined unitedly to represent complex manipulations, thereby reducing intention ambiguity. Furthermore, GDrag introduces two strategies to mitigate content ambiguity, including an anti-ambiguity dense trajectory calculation method (ADT) and a self-adaptive motion supervision method (SMS). Given an atomic manipulation, ADT converts the sparse user-defined handle points into a dense point set by selecting their semantic and geometric neighbors, and calculates the trajectory of the point set. Unlike previous motion supervision methods relying on a single global scale for low-rank adaption, SMS jointly optimizes point-wise adaption scales and latent feature biases. These two methods allow us to model fine-grained target contexts and generate precise trajectories. As a result, GDrag consistently produces precise and appealing results in different editing tasks. Extensive experiments on the challenging DragBench dataset demonstrate that GDrag outperforms state-of-the-art methods significantly. The code of GDrag will be released upon acceptance. | Interactive editing, dragging-based image manipulation, diffusion models | null | 1,073 | null | [
-0.010898040607571602,
-0.04717426002025604,
0.03493577614426613,
0.04428314045071602,
0.05259832367300987,
0.0032740181777626276,
0.038025423884391785,
-0.012986931949853897,
0.06071828678250313,
-0.06921546906232834,
0.011731799691915512,
-0.04272552579641342,
-0.0003493427357170731,
0.05289643257856369,
-0.003291092114523053,
-0.06730664521455765,
0.05784182995557785,
0.04534193128347397,
-0.05063549801707268,
-0.006604359019547701,
0.07040774077177048,
-0.0022985159885138273,
0.03976317495107651,
-0.014741254970431328,
0.06300405412912369,
0.01802472211420536,
0.023908449336886406,
-0.06561777740716934,
0.04434506967663765,
-0.030271656811237335,
0.032362200319767,
0.019220881164073944,
-0.02942061424255371,
0.034170884639024734,
-0.06874116510152817,
0.024565506726503372,
-0.029339130967855453,
0.07335522025823593,
-0.0025988591369241476,
-0.07179702818393707,
0.0016659677494317293,
0.03337298333644867,
0.017689984291791916,
0.04511938616633415,
0.0491415336728096,
0.016395216807723045,
0.014402681030333042,
-0.05029486119747162,
-0.048642437905073166,
0.07580359280109406,
-0.02129063382744789,
-0.052634380757808685,
-0.056197941303253174,
-0.059212878346443176,
-0.00028564990498125553,
0.01609555445611477,
0.049457211047410965,
0.008719364181160927,
0.0947103276848793,
-0.05317877605557442,
-0.0022839107550680637,
-0.019268503412604332,
0.010616780258715153,
-0.0376114696264267,
-0.018259890377521515,
-0.05453313887119293,
0.12204290181398392,
-0.027738777920603752,
0.028852315619587898,
-0.021908052265644073,
0.009887664578855038,
-0.015857819467782974,
-0.0006298251682892442,
0.034135688096284866,
-0.05431587994098663,
0.017721205949783325,
-0.06880518794059753,
-0.013662493787705898,
-0.03619750589132309,
-0.0187639482319355,
0.07766216993331909,
-0.07002905011177063,
0.03535807132720947,
0.0352955125272274,
0.06994743645191193,
-0.026885343715548515,
-0.05818238481879234,
0.09853003919124603,
0.05712413415312767,
0.004310502205044031,
-0.019502803683280945,
-0.020001158118247986,
0.0018683617236092687,
-0.067257821559906,
0.04591967910528183,
-0.024506745859980583,
-0.03186415135860443,
-0.09264832735061646,
-0.023239072412252426,
0.08087847381830215,
0.07656777650117874,
0.08914146572351456,
-0.07231463491916656,
0.0059775011613965034,
0.032281067222356796,
0.0599437952041626,
-0.007221590727567673,
-0.022213654592633247,
-0.015527752228081226,
0.014006410725414753,
0.0015499292640015483,
-0.00522115733474493,
0.020008595660328865,
-0.04680102691054344,
0.029649781063199043,
-0.003378933761268854,
0.03190692141652107,
-0.03609824553132057,
0.027296019718050957,
0.020488301292061806,
0.013361587189137936,
-0.05295621231198311,
-0.03940216451883316,
0.0694412961602211,
0.03165234625339508,
-0.08291938155889511,
-0.06516910344362259,
3.186663706855967e-34,
0.07551800459623337,
0.0057961661368608475,
0.09612385928630829,
-0.006337770260870457,
0.014485010877251625,
0.005742525681853294,
-0.04595668986439705,
-0.03603682667016983,
-0.061368390917778015,
-0.04078639671206474,
0.062054216861724854,
-0.09600098431110382,
-0.0054020630195736885,
0.09923980385065079,
0.0914284735918045,
-0.03424821048974991,
-0.03303123638033867,
0.031680308282375336,
0.022229475900530815,
0.03663608431816101,
-0.021968122571706772,
-0.05279576778411865,
-0.06296815723180771,
-0.037640925496816635,
-0.058334141969680786,
0.048319049179553986,
0.02182038687169552,
0.005415833555161953,
0.049100425094366074,
-0.024232784286141396,
-0.10619640350341797,
-0.02458403818309307,
-0.001489519840106368,
0.031161582097411156,
-0.009379292838275433,
-0.024124925956130028,
-0.07163171470165253,
-0.10610145330429077,
0.051132701337337494,
-0.06958398222923279,
0.008604053407907486,
-0.005360369570553303,
-0.024123087525367737,
-0.0014568427577614784,
-0.10085252672433853,
0.03758322447538376,
-0.0002218541339971125,
0.11953413486480713,
-0.10699150711297989,
0.08751334249973297,
0.05679044499993324,
0.032658323645591736,
0.032267794013023376,
-0.06518717855215073,
-0.0067163994535803795,
-0.009458591230213642,
0.06631433218717575,
-0.02298901230096817,
0.02441880851984024,
0.02037348411977291,
0.04691444709897041,
0.05709049105644226,
-0.09454468637704849,
0.08925052732229233,
0.032842375338077545,
0.04540722444653511,
-0.061101026833057404,
0.06068068742752075,
0.09295648336410522,
-0.007324162870645523,
-0.08633924275636673,
-0.019503403455018997,
0.04182567819952965,
-0.04460645839571953,
0.004803014453500509,
-0.03429121896624565,
0.017862949520349503,
-0.06455699354410172,
0.05858641490340233,
-0.006169233471155167,
-0.08173594623804092,
-0.03929347172379494,
0.05784425511956215,
-0.056269314140081406,
-0.09096718579530716,
-0.09875140339136124,
0.041029125452041626,
-0.02839876525104046,
-0.026978297159075737,
-0.04670681804418564,
0.01983451098203659,
-0.007021743804216385,
-0.03515968471765518,
0.09834562242031097,
-0.010832215659320354,
4.532786113652398e-34,
0.037327416241168976,
0.02383737824857235,
-0.00789191946387291,
0.040279142558574677,
0.025770453736186028,
-0.03742661699652672,
-0.009639591909945011,
-0.05890730395913124,
0.10322672873735428,
-0.11607149243354797,
-0.08570500463247299,
-0.017895163968205452,
-0.05038607865571976,
-0.05535009875893593,
0.023613542318344116,
0.05542165786027908,
-0.022894583642482758,
0.025919942185282707,
-0.05855519697070122,
0.00988173671066761,
-0.009511827491223812,
0.00048360435175709426,
0.026013724505901337,
0.0012245182879269123,
-0.016719631850719452,
0.06723926961421967,
0.05315913259983063,
0.03155582398176193,
-0.0036475437227636576,
-0.021108388900756836,
-0.020356345921754837,
0.015354459173977375,
0.07194466888904572,
-0.06479888409376144,
0.04051635414361954,
0.0004049515991937369,
0.029210181906819344,
-0.004590620286762714,
-0.004542241338640451,
0.08206301927566528,
0.03112681210041046,
0.001136073493398726,
-0.005105421878397465,
0.0698499083518982,
-0.03323064744472504,
0.0757712796330452,
-0.06513544172048569,
0.03410781919956207,
-0.0026514045894145966,
0.08841457217931747,
-0.10173064470291138,
0.03386978060007095,
-0.10916547477245331,
-0.06050540506839752,
0.014759751968085766,
-0.06349878758192062,
0.00946879107505083,
-0.032679200172424316,
0.0916190892457962,
-0.06492412090301514,
-0.0039010385517030954,
0.028174826875329018,
-0.04247584566473961,
-0.024827755987644196,
0.023664096370339394,
-0.007865866646170616,
-0.06390614062547684,
-0.05452309176325798,
-0.0010157407959923148,
0.08099682629108429,
0.03284377232193947,
-0.03847340866923332,
0.0295528806746006,
-0.05634267255663872,
0.08681465685367584,
-0.05509520694613457,
0.09190237522125244,
-0.030142592266201973,
-0.056697312742471695,
-0.13586419820785522,
0.06930931657552719,
0.028551628813147545,
0.04956364631652832,
0.0652184933423996,
-0.04280952364206314,
0.1478629857301712,
-0.0485844686627388,
0.10433084517717361,
0.011759900487959385,
0.05161752924323082,
-0.08206330239772797,
0.014702205546200275,
0.04216622933745384,
0.04358139634132385,
-0.08341111987829208,
-4.281468690692236e-8,
-0.10934844613075256,
0.030330820009112358,
-0.017165884375572205,
-0.0006983843632042408,
0.0024829923640936613,
-0.0007356176502071321,
0.02263752557337284,
0.025861643254756927,
0.04912611469626427,
-0.02454313635826111,
0.060444559901952744,
-0.07353602349758148,
-0.006707688793540001,
-0.042539503425359726,
-0.0268434826284647,
0.024174079298973083,
0.007346149999648333,
0.04148634150624275,
-0.051922231912612915,
0.044049713760614395,
-0.04847344011068344,
0.06707270443439484,
-0.05467342212796211,
-0.05437741428613663,
0.019115451723337173,
-0.018979882821440697,
-0.032320596277713776,
0.10827096551656723,
0.0818597674369812,
-0.017176544293761253,
0.06174560263752937,
0.00694532785564661,
0.08012226969003677,
0.1013968214392662,
-0.05313732102513313,
0.03259211778640747,
0.05317161977291107,
-0.022630224004387856,
-0.022005833685398102,
0.004332737531512976,
-0.029646409675478935,
0.06631013751029968,
0.0012226810213178396,
0.016958316788077354,
-0.021811019629240036,
0.033386651426553726,
0.066407211124897,
-0.12539125978946686,
0.008808223530650139,
0.012304523028433323,
-0.014084214344620705,
0.03598211705684662,
-0.046749528497457504,
0.06781710684299469,
0.09130631387233734,
0.05648712441325188,
0.01363879069685936,
-0.016490047797560692,
0.06415696442127228,
0.07464943826198578,
0.007506422698497772,
-0.017592286691069603,
-0.0428055040538311,
0.011401204392313957
] |
AvatarGO: Zero-shot 4D Human-Object Interaction Generation and Animation | https://openreview.net/forum?id=Trf0R8eoGF | [
"Yukang Cao",
"Liang Pan",
"Kai Han",
"Kwan-Yee K. Wong",
"Ziwei Liu"
] | Poster | Recent advancements in diffusion models have led to significant improvements in the generation and animation of 4D full-body human-object interactions (HOI). Nevertheless, existing methods primarily focus on SMPL-based motion generation, which is limited by the scarcity of realistic large-scale interaction data. This constraint affects their ability to create everyday HOI scenes. This paper addresses this challenge using a zero-shot approach with a pre-trained diffusion model. Despite this potential, achieving our goals is difficult due to the diffusion model's lack of understanding of ''where'' and ''how'' objects interact with the human body. To tackle these issues, we introduce **AvatarGO**, a novel framework designed to generate animatable 4D HOI scenes directly from textual inputs. Specifically, **1)** for the ''where'' challenge, we propose **LLM-guided contact retargeting**, which employs Lang-SAM to identify the contact body part from text prompts, ensuring precise representation of human-object spatial relations. **2)** For the ''how'' challenge, we introduce **correspondence-aware motion optimization** that constructs motion fields for both human and object models using the linear blend skinning function from SMPL-X. Our framework not only generates coherent compositional motions, but also exhibits greater robustness in handling penetration issues. Extensive experiments with existing methods validate AvatarGO's superior generation and animation capabilities on a variety of human-object pairs and diverse poses. As the first attempt to synthesize 4D avatars with object interactions, we hope AvatarGO could open new doors for human-centric 4D content creation. | 4D human avatar generation, compositional generation, human-object interaction | an interactable animatable pipeline for generating and animating 4D avatars with object interactions from text prompts. | 1,071 | 2410.07164 | [
-0.015117902308702469,
0.009782467968761921,
0.0714377835392952,
-0.009347227402031422,
0.04397397115826607,
-0.054854102432727814,
-0.01738879457116127,
-0.0504300557076931,
0.03982727229595184,
-0.008255439810454845,
0.016187293455004692,
-0.09108155965805054,
-0.08003561943769455,
0.06529051065444946,
0.0007875270675867796,
-0.05735897272825241,
0.07311967015266418,
0.041569460183382034,
-0.10609221458435059,
0.07035472989082336,
-0.006718121003359556,
-0.048613954335451126,
0.05555170401930809,
-0.05880223959684372,
0.006647518835961819,
-0.010839945636689663,
0.030489178374409676,
-0.10682857036590576,
0.06981221586465836,
-0.013168114237487316,
0.03972289711236954,
0.022423021495342255,
-0.015232255682349205,
0.05149008706212044,
-0.09640050679445267,
0.07552613317966461,
0.044414740055799484,
-0.021725092083215714,
-0.11228111386299133,
-0.0796685740351677,
0.00700449338182807,
0.041330013424158096,
0.024658318608999252,
0.014591117389500141,
0.15736123919487,
0.023851225152611732,
-0.057571977376937866,
-0.04418318346142769,
0.03214877098798752,
-0.002853166311979294,
-0.11470585316419601,
-0.08292766660451889,
0.022576605901122093,
0.06594406068325043,
-0.02352127991616726,
0.04902125522494316,
0.062238045036792755,
-0.03971375152468681,
0.0857926532626152,
-0.00855112075805664,
-0.04287775978446007,
-0.05289839580655098,
0.018145473673939705,
0.005490332376211882,
0.0062356069684028625,
-0.0822007954120636,
0.08623917400836945,
-0.02568983845412731,
-0.04303058609366417,
0.02089426852762699,
-0.047792475670576096,
0.010799605399370193,
-0.04163750261068344,
-0.007311976980417967,
0.0024020452983677387,
0.011775529012084007,
-0.05354122817516327,
-0.05096549913287163,
0.002163480268791318,
-0.020503757521510124,
0.10391626507043839,
-0.05472993850708008,
0.04376234859228134,
-0.02918795868754387,
-0.06467271596193314,
0.044826868921518326,
-0.024883808568120003,
0.061386287212371826,
-0.039492297917604446,
0.03566763177514076,
-0.12691520154476166,
-0.010321724228560925,
0.011470199562609196,
-0.03709915652871132,
-0.021622581407427788,
0.026485519483685493,
-0.037516649812459946,
-0.067581407725811,
0.03904717415571213,
0.07081034034490585,
-0.03875823691487312,
-0.031470220535993576,
-0.0045348056592047215,
0.017735421657562256,
0.02221212536096573,
0.000856570026371628,
0.03987140208482742,
-0.035063888877630234,
-0.03314913809299469,
0.07069530338048935,
0.05679740384221077,
-0.04901900514960289,
-0.01890854723751545,
0.02007545530796051,
0.04594441130757332,
-0.061679404228925705,
0.006015891674906015,
-0.025042841210961342,
0.049768395721912384,
-0.03325560688972473,
0.05157150328159332,
-0.010173680260777473,
-0.06272097676992416,
0.004604781046509743,
0.050767261534929276,
-0.09051026403903961,
-0.03360070660710335,
-2.481180211127913e-33,
0.0908459797501564,
0.019666610285639763,
0.10110382735729218,
0.0816105455160141,
0.005090727936476469,
-0.01947922632098198,
-0.038582246750593185,
0.046220969408750534,
-0.008049389347434044,
-0.0631449893116951,
-0.02461221255362034,
-0.05884676054120064,
-0.060206957161426544,
0.07357999682426453,
0.061055295169353485,
-0.062193211168050766,
0.03345794230699539,
0.022814657539129257,
-0.03265530243515968,
0.07914745807647705,
-0.029405808076262474,
-0.03789462149143219,
-0.048171866685152054,
-0.005872488021850586,
0.018681731075048447,
0.06809241324663162,
0.06094878539443016,
-0.07435554265975952,
0.021561717614531517,
-0.007793241646140814,
-0.11862462013959885,
0.030488288030028343,
-0.007587463594973087,
-0.004956959281116724,
0.022556042298674583,
-0.014509391039609909,
-0.07405855506658554,
-0.044598136097192764,
-0.001546851941384375,
0.03378192335367203,
0.028761034831404686,
0.028434937819838524,
-0.037809278815984726,
-0.03009326569736004,
-0.08449888229370117,
0.03448694944381714,
0.07121788710355759,
0.09864460676908493,
-0.07689923793077469,
0.07335877418518066,
0.014207554049789906,
0.05116363614797592,
-0.022848770022392273,
-0.05924096331000328,
-0.06622838973999023,
-0.08788538724184036,
0.007080491166561842,
-0.09956061094999313,
0.010515403933823109,
0.013323970139026642,
0.020592130720615387,
0.048828668892383575,
-0.02872401289641857,
0.07424117624759674,
0.03945329785346985,
-0.028016716241836548,
-0.002900024177506566,
0.05382182449102402,
0.0226289015263319,
0.014991055242717266,
-0.04052351415157318,
-0.019395790994167328,
-0.0010808982187882066,
-0.042340200394392014,
0.056711915880441666,
-0.008025838993489742,
0.0605182945728302,
-0.019640520215034485,
-0.00994090549647808,
0.04131157696247101,
-0.024525726214051247,
0.0015603241045027971,
-0.04142356663942337,
-0.043748967349529266,
0.034519366919994354,
-0.06046817824244499,
0.03257712349295616,
-0.05729790776968002,
-0.06986066699028015,
-0.09367675334215164,
-0.00874897837638855,
-0.05251939222216606,
-0.06032072380185127,
0.05880993977189064,
0.0029784240759909153,
1.3976724258547964e-33,
0.07550617307424545,
0.03540410101413727,
-0.0705241709947586,
0.024958031252026558,
-0.030621062964200974,
-0.0025052749551832676,
0.04330892860889435,
0.06361310929059982,
0.06446944177150726,
-0.01409722026437521,
-0.04449150338768959,
-0.007928802631795406,
0.03844952583312988,
-0.028682319447398186,
0.06710796058177948,
-0.0031102329958230257,
0.007380289025604725,
-0.06749088317155838,
-0.04085993021726608,
0.04102460667490959,
0.032792091369628906,
0.051931433379650116,
-0.07593425363302231,
-0.010220970027148724,
-0.06757836788892746,
0.056970760226249695,
0.07374009490013123,
0.030666138976812363,
-0.038928546011447906,
0.0023633597884327173,
-0.025781700387597084,
-0.047707997262477875,
0.02861868403851986,
0.01671300083398819,
-0.07247190177440643,
0.012633347883820534,
0.045398544520139694,
0.021756045520305634,
0.026552390307188034,
0.02807384729385376,
0.10186275094747543,
-0.030718782916665077,
-0.06298401206731796,
0.08044955879449844,
-0.08667129278182983,
-0.0067214989103376865,
-0.1144050806760788,
0.04137960076332092,
-0.02717508189380169,
0.04989621415734291,
-0.048370685428380966,
0.05819148197770119,
-0.07097505033016205,
-0.09007836878299713,
-0.037713970988988876,
-0.07244735956192017,
-0.038003742694854736,
-0.03287721797823906,
0.07165493071079254,
-0.03103780932724476,
0.012761360965669155,
-0.015073689632117748,
0.02835129015147686,
0.01991182379424572,
-0.007165099494159222,
0.028412239626049995,
-0.03196825087070465,
-0.03680787235498428,
-0.029734531417489052,
-0.0068492828868329525,
0.047141872346401215,
0.027535365894436836,
0.009511985816061497,
0.005075471010059118,
0.042955007404088974,
-0.01812642253935337,
-0.007656975649297237,
-0.035716280341148376,
0.09246364235877991,
-0.035253170877695084,
0.0022219906095415354,
-0.012093475088477135,
0.03478283807635307,
0.08110565692186356,
0.06979115307331085,
0.09197469800710678,
-0.04704543575644493,
0.10359986126422882,
0.01987026445567608,
0.09730581939220428,
-0.058271270245313644,
0.10662580281496048,
0.059854164719581604,
0.09272468090057373,
-0.05927172303199768,
-4.224254723794729e-8,
-0.09088709950447083,
0.016233062371611595,
0.04637860879302025,
0.05149954929947853,
-0.031087614595890045,
0.03984636068344116,
0.03423243761062622,
0.05561748892068863,
0.00639552203938365,
0.00008950274059316143,
-0.00728748319670558,
0.02445373311638832,
0.01435126457363367,
-0.013307087123394012,
0.012717997655272484,
0.10385257005691528,
-0.06308628618717194,
-0.03693382814526558,
-0.05171367898583412,
0.03614756837487221,
0.049530912190675735,
0.019500553607940674,
-0.015665967017412186,
0.02587353065609932,
0.020272035151720047,
-0.026233183220028877,
-0.035099346190690994,
0.060362618416547775,
0.02135482057929039,
-0.09156421571969986,
-0.01878081075847149,
-0.029703538864850998,
-1.3649216157318733e-7,
0.009347209706902504,
-0.04438399896025658,
0.03309376910328865,
0.007151621859520674,
-0.04178467392921448,
-0.0009343705605715513,
-0.07609555125236511,
0.034725889563560486,
0.032794833183288574,
0.01150781475007534,
-0.0025027666706591845,
-0.008884805254638195,
0.04190417006611824,
0.06452903896570206,
-0.14472952485084534,
0.02710307389497757,
-0.014921862632036209,
-0.0844726487994194,
-0.06610608845949173,
-0.024776577949523926,
0.08146006613969803,
0.028023991733789444,
0.034320227801799774,
0.0640026107430458,
0.027370477095246315,
0.10055600851774216,
0.08060131967067719,
0.023128505796194077,
0.04809018224477768,
-0.07906778901815414,
0.06462110579013824
] |
Calibrating LLMs with Information-Theoretic Evidential Deep Learning | https://openreview.net/forum?id=YcML3rJl0N | [
"Yawei Li",
"David Rügamer",
"Bernd Bischl",
"Mina Rezaei"
] | Poster | Fine-tuned large language models (LLMs) often exhibit overconfidence, particularly when trained on small datasets, resulting in poor calibration and inaccurate uncertainty estimates.
Evidential Deep Learning (EDL), an uncertainty-aware approach, enables uncertainty estimation in a single forward pass, making it a promising method for calibrating fine-tuned LLMs. However, despite its computational efficiency, EDL is prone to overfitting, as its training objective can result in overly concentrated probability distributions.
To mitigate this, we propose regularizing EDL by incorporating an information bottleneck (IB). Our approach IB-EDL suppresses spurious information in the evidence generated by the model and encourages truly predictive information to influence both the predictions and uncertainty estimates. Extensive experiments across various fine-tuned LLMs and tasks demonstrate that IB-EDL outperforms both existing EDL and non-EDL approaches.
By improving the trustworthiness of LLMs, IB-EDL facilitates their broader adoption in domains requiring high levels of confidence calibration. | evidential deep learning; information bottleneck; calibration; large language models | This paper introduces an information-theoretic regularization for evidential deep learning, substantially enhancing the calibration of fine-tuned large language models. | 1,068 | 2502.06351 | [
-0.017414165660738945,
-0.08662337809801102,
0.065714031457901,
0.05125159025192261,
0.09332538396120071,
0.026941806077957153,
-0.009774434380233288,
-0.037996936589479446,
0.015194213017821312,
-0.07784205675125122,
-0.09642870724201202,
-0.0676775649189949,
0.02264631725847721,
-0.018794860690832138,
-0.07303410023450851,
-0.010748903267085552,
0.06015313044190407,
0.08269794285297394,
-0.14470437169075012,
-0.08066128939390182,
0.06120805814862251,
0.06833355128765106,
0.05720138922333717,
0.011927190236747265,
0.05793354660272598,
-0.030300239101052284,
-0.017445849254727364,
-0.01845475845038891,
0.012879740446805954,
-0.06781622022390366,
0.08290210366249084,
-0.028710797429084778,
0.020730994641780853,
-0.044978588819503784,
0.04793550446629524,
0.028144534677267075,
-0.0247033704072237,
0.05128860846161842,
0.03980676457285881,
-0.02249164506793022,
0.07641147822141647,
0.016485251486301422,
0.03149816393852234,
0.06299033015966415,
0.05618567019701004,
-0.022683218121528625,
0.03915422037243843,
-0.0007636824739165604,
-0.005934754386544228,
0.05781887471675873,
0.00042066507739946246,
-0.038083773106336594,
0.02461300790309906,
0.032645005732774734,
-0.03674699738621712,
0.016223305836319923,
0.04191378131508827,
0.08412529528141022,
-0.06751365959644318,
-0.005927423946559429,
-0.016430694609880447,
-0.08074129372835159,
-0.04613730311393738,
0.020868292078375816,
-0.051135264337062836,
0.05413100868463516,
0.009851880371570587,
0.07625341415405273,
-0.03261681646108627,
0.017933906987309456,
-0.007211821153759956,
0.10637868195772171,
-0.05570996180176735,
0.11801464855670929,
0.03878297284245491,
0.006983907427638769,
0.0311425793915987,
-0.04322964698076248,
0.08115209639072418,
-0.009630334563553333,
-0.0117489630356431,
0.07506975531578064,
0.03505934774875641,
-0.028193829581141472,
0.16888821125030518,
-0.011565898545086384,
-0.02248236909508705,
0.018069693818688393,
-0.02623400278389454,
-0.03819001838564873,
-0.05434178560972214,
-0.14178867638111115,
-0.017462605610489845,
0.03736524283885956,
0.02103130891919136,
0.06165406480431557,
-0.031120579689741135,
-0.11327948421239853,
-0.05980105325579643,
0.07395441085100174,
-0.045299503952264786,
0.041793886572122574,
0.022859971970319748,
-0.0014429400907829404,
-0.01862173154950142,
-0.03405717760324478,
-0.0023684061598032713,
0.05135687068104744,
-0.03561406582593918,
-0.01646966114640236,
0.07615248113870621,
0.016057753935456276,
0.052632782608270645,
0.05935250222682953,
0.047690726816654205,
-0.01893959566950798,
-0.008070271462202072,
0.02082202211022377,
-0.041608016937971115,
0.01650281995534897,
-0.12250202149152756,
0.03995990380644798,
0.023600000888109207,
0.015481490641832352,
-0.012226054444909096,
-0.009868727065622807,
-0.00695584574714303,
4.2039237529236085e-33,
0.0510406456887722,
-0.003838572185486555,
-0.01964782550930977,
0.046440817415714264,
-0.00008936919766711071,
0.05847425386309624,
-0.008656284771859646,
0.030468927696347237,
-0.04813876748085022,
0.10219845175743103,
-0.01085966918617487,
-0.0049585821107029915,
-0.05311822518706322,
0.04556844383478165,
0.0360969714820385,
0.07297415286302567,
-0.044139087200164795,
0.05582990497350693,
-0.05501456931233406,
-0.030557477846741676,
0.024912701919674873,
-0.06407860666513443,
-0.025284696370363235,
-0.07980410754680634,
-0.004631648305803537,
0.028553703799843788,
0.08957040309906006,
0.05500445142388344,
-0.006333306897431612,
0.004743027500808239,
-0.1094668060541153,
-0.0721091628074646,
0.08114872872829437,
-0.003951754420995712,
0.0489022396504879,
-0.05211982876062393,
-0.02193322777748108,
-0.02630419284105301,
0.053070202469825745,
-0.01982911303639412,
0.06594303250312805,
0.03841814771294594,
0.028645385056734085,
0.03749508038163185,
-0.0768544152379036,
-0.033808521926403046,
0.0006157840834930539,
-0.08366003632545471,
-0.011725977063179016,
-0.09224524348974228,
0.04648683965206146,
-0.015606597997248173,
-0.04535754397511482,
0.01595824770629406,
-0.005564953666180372,
-0.03809627518057823,
0.03878994658589363,
0.05916989967226982,
-0.027013523504137993,
0.00070237583713606,
0.015252653509378433,
0.07446914911270142,
-0.006500414106994867,
0.06665199995040894,
0.04556473717093468,
0.09770557284355164,
-0.05951031669974327,
0.04688769206404686,
-0.013698738999664783,
-0.0926784947514534,
-0.04299776628613472,
-0.02245762012898922,
-0.005770830437541008,
-0.0751643031835556,
0.05139685049653053,
-0.036285631358623505,
0.04021454602479935,
-0.03546721488237381,
0.09683404117822647,
0.040636297315359116,
0.044435929507017136,
-0.006257605738937855,
-0.00022340836585499346,
0.019857121631503105,
-0.036790281534194946,
-0.04098746180534363,
0.011421390809118748,
0.008865387178957462,
-0.008971929550170898,
0.02146141044795513,
-0.013559868559241295,
-0.011520634405314922,
-0.035279788076877594,
-0.018410179764032364,
-0.05787523463368416,
-3.292840325929486e-33,
-0.06726783514022827,
0.027249325066804886,
0.021734517067670822,
0.09928876161575317,
-0.05874490365386009,
-0.05408534035086632,
-0.02265247330069542,
0.043153271079063416,
0.009001517668366432,
-0.09703559428453445,
-0.020757140591740608,
-0.04439007118344307,
0.06251322478055954,
-0.032469168305397034,
0.0453028529882431,
-0.02172398939728737,
-0.0014737615128979087,
0.040915727615356445,
0.024869855493307114,
0.007577275857329369,
0.01955326646566391,
0.006793572101742029,
-0.0646071657538414,
0.02287386916577816,
-0.025657163932919502,
-0.01242049876600504,
-0.059459198266267776,
0.019075321033596992,
0.006688571069389582,
-0.024019388481974602,
-0.03491811454296112,
-0.014598516747355461,
-0.02112683095037937,
-0.04474734514951706,
-0.07701878994703293,
0.04828743636608124,
0.037306711077690125,
0.024987904354929924,
-0.07779919356107712,
0.02655680850148201,
0.10691919177770615,
-0.030360212549567223,
-0.10812167823314667,
-0.06065988168120384,
-0.05627211928367615,
-0.028743693605065346,
-0.040779273957014084,
-0.04447338357567787,
0.061482980847358704,
0.07136020064353943,
0.003161644795909524,
0.07054191827774048,
-0.06191478669643402,
0.09269975125789642,
-0.060835033655166626,
-0.05847685784101486,
-0.05410151928663254,
-0.013198534958064556,
0.05915598198771477,
0.03301367536187172,
-0.08570509403944016,
0.023090390488505363,
-0.009012393653392792,
-0.09748365730047226,
0.0360628105700016,
0.012413243763148785,
-0.06398998200893402,
0.06550996005535126,
0.07365504652261734,
0.0006985128275118768,
0.08795362710952759,
-0.02379206009209156,
-0.004066064488142729,
0.03168993443250656,
-0.0269894078373909,
0.01138315349817276,
0.006764410994946957,
-0.026071688160300255,
0.028031202033162117,
0.0015406482852995396,
0.1059223860502243,
0.03793448209762573,
0.005158918909728527,
0.07381148636341095,
0.09087672084569931,
0.05736918747425079,
0.034245170652866364,
0.00889032892882824,
0.004613871220499277,
0.0615946426987648,
0.005723322741687298,
-0.006652976851910353,
-0.031184278428554535,
0.030709808692336082,
-0.03756815195083618,
-4.349214677290547e-8,
-0.06072995066642761,
0.06947918981313705,
0.016025785356760025,
0.0008557880064472556,
-0.005556012038141489,
-0.040159814059734344,
-0.03335605934262276,
0.09570321440696716,
0.022654645144939423,
-0.05725632235407829,
0.08773218840360641,
0.019566381350159645,
-0.08638804405927658,
-0.04661109670996666,
-0.021880779415369034,
0.0878174901008606,
0.0088995061814785,
0.04981409013271332,
0.0036438710521906614,
-0.05059082433581352,
0.0714850202202797,
-0.00991460308432579,
-0.0252358540892601,
-0.06895297765731812,
0.02968733385205269,
-0.11764425039291382,
-0.02820288948714733,
0.11925075948238373,
0.003489610506221652,
-0.03444325551390648,
-0.016105301678180695,
-0.08543180674314499,
-0.04584861546754837,
-0.020839747041463852,
0.018359629437327385,
0.137496218085289,
-0.010983819141983986,
-0.04541624337434769,
0.05402615666389465,
-0.04502764344215393,
-0.035265449434518814,
0.06408550590276718,
-0.039580266922712326,
0.05407661199569702,
-0.013502351939678192,
-0.019626444205641747,
-0.0017648790962994099,
-0.0719454362988472,
0.0150063531473279,
-0.07893720269203186,
0.037191156297922134,
-0.0016454641008749604,
0.03421660140156746,
0.04001938924193382,
-0.021009046584367752,
0.06956220418214798,
-0.0038691037334501743,
-0.059127070009708405,
-0.06152154877781868,
0.06172679364681244,
0.004070261958986521,
-0.006505836267024279,
-0.06705532222986221,
-0.016194626688957214
] |
TIGeR: Unifying Text-to-Image Generation and Retrieval with Large Multimodal Models | https://openreview.net/forum?id=mr2icR6dpD | [
"Leigang Qu",
"Haochuan Li",
"Tan Wang",
"Wenjie Wang",
"Yongqi Li",
"Liqiang Nie",
"Tat-Seng Chua"
] | Poster | How humans can effectively and efficiently acquire images has always been a perennial question. A classic solution is *text-to-image retrieval* from an existing database; however, the limited database typically lacks creativity. By contrast, recent breakthroughs in *text-to-image generation* have made it possible to produce attractive and counterfactual visual content, but it faces challenges in synthesizing knowledge-intensive images. In this work, we rethink the relationship between text-to-image generation and retrieval, proposing a *unified* framework for both tasks with one single Large Multimodal Model (LMM). Specifically, we first explore the intrinsic discriminative abilities of LMMs and introduce an efficient generative retrieval method for text-to-image retrieval in a training-free manner. Subsequently, we unify generation and retrieval autoregressively and propose an autonomous decision mechanism to choose the best-matched one between generated and retrieved images as the response to the text prompt. To standardize the evaluation of unified text-to-image generation and retrieval, we construct TIGeR-Bench, a benchmark spanning both creative and knowledge-intensive domains. Extensive experiments on TIGeR-Bench and two retrieval benchmarks, *i.e.*, Flickr30K and MS-COCO, demonstrate the superiority of our proposed framework. | Multimodal Large Language Models, Text-to-Image Generation, Cross-Modal Retrieval | This paper explores the intrinsic discriminative ability of multimodal foundation models, and proposes a unified framework combining text-to-image generation and retrieval. | 1,065 | 2406.05814 | [
-0.03176933526992798,
-0.09388265013694763,
0.009308774955570698,
0.07839122414588928,
0.0938880518078804,
0.009602367877960205,
-0.05677857622504234,
-0.0008381961961276829,
0.01779320277273655,
-0.030716383829712868,
0.023826992139220238,
-0.07949046045541763,
0.09810224175453186,
0.05956069752573967,
-0.03518545627593994,
0.045075271278619766,
0.07577958703041077,
0.03259557858109474,
-0.10527532547712326,
0.016791056841611862,
0.016591493040323257,
-0.011079699732363224,
0.1244291141629219,
-0.046995148062705994,
0.03685295954346657,
0.02206340618431568,
-0.008181264623999596,
-0.026583269238471985,
0.014503523707389832,
-0.05463283881545067,
0.014153634198009968,
0.01428604032844305,
-0.02315160073339939,
0.09303915500640869,
-0.02025231532752514,
0.07749190926551819,
-0.06098301336169243,
0.053123507648706436,
0.037016868591308594,
-0.05267142876982689,
0.019038859754800797,
-0.01545774657279253,
0.006502056960016489,
0.0544615164399147,
0.10646647959947586,
0.02922520600259304,
-0.06239708513021469,
-0.05278342589735985,
0.0731055736541748,
0.032424453645944595,
-0.07952485233545303,
-0.021435756236314774,
-0.05654555559158325,
0.039793550968170166,
0.022213224321603775,
0.06204769015312195,
0.037059951573610306,
-0.058771874755620956,
0.008469062857329845,
-0.0299331396818161,
-0.04420506954193115,
-0.090769462287426,
-0.04938265681266785,
-0.05420614778995514,
0.056578416377305984,
-0.07334782928228378,
0.04468892142176628,
-0.04189656302332878,
0.028133930638432503,
-0.033230990171432495,
0.00007847797678550705,
0.03433052450418472,
-0.033669814467430115,
0.07907743752002716,
0.009862850420176983,
0.09610097110271454,
0.03599860146641731,
-0.06323777139186859,
0.04004073888063431,
-0.045869771391153336,
0.06259379535913467,
0.0077714938670396805,
0.0723232850432396,
-0.015206804499030113,
0.10730632394552231,
-0.0049461186863482,
0.010129894129931927,
0.06061546131968498,
0.02523137629032135,
0.037154700607061386,
-0.09526816010475159,
-0.012547626160085201,
0.06311293691396713,
-0.021730849519371986,
0.004814709536731243,
0.04349754750728607,
0.092837393283844,
-0.07081501185894012,
-0.034189797937870026,
0.06674247235059738,
0.059549108147621155,
0.056937865912914276,
-0.04327467456459999,
-0.020088987424969673,
0.01995231583714485,
-0.025390351191163063,
0.0767688974738121,
0.04080316424369812,
0.03125608339905739,
-0.13606762886047363,
0.09363272041082382,
0.006352819036692381,
0.004917492158710957,
0.011738198809325695,
-0.000898087106179446,
-0.009691482409834862,
-0.005889326799660921,
-0.002632057759910822,
0.06480973958969116,
-0.026297619566321373,
0.012002113275229931,
0.005729671102017164,
-0.035504795610904694,
0.009708086028695107,
-0.04576069116592407,
-0.05363626405596733,
-0.015540680848062038,
-1.5865751945514035e-33,
0.05339581519365311,
0.015251534059643745,
0.08596350252628326,
0.09507028758525848,
0.06760398298501968,
-0.056216128170490265,
-0.10287124663591385,
0.044812675565481186,
-0.10461675375699997,
-0.008191365748643875,
-0.050516244024038315,
0.015116256661713123,
-0.06395135819911957,
0.10969702899456024,
0.030442386865615845,
-0.009959433227777481,
-0.09327057749032974,
-0.01437987107783556,
0.041172273457050323,
0.06319593638181686,
0.025406088680028915,
-0.015747232362627983,
0.011813923716545105,
-0.007577172946184874,
-0.013499973341822624,
-0.021830175071954727,
0.07132327556610107,
-0.08672039955854416,
-0.023656025528907776,
-0.013059872202575207,
-0.15633650124073029,
-0.013805913738906384,
0.097029909491539,
0.022874262183904648,
0.007867256179451942,
-0.0830429345369339,
-0.05509209632873535,
-0.022774890065193176,
-0.01961633376777172,
-0.0038674212992191315,
-0.011144047603011131,
0.02499057725071907,
-0.048698242753744125,
-0.053699251264333725,
-0.0401538647711277,
-0.0056552886962890625,
-0.0007604877464473248,
0.04083118587732315,
-0.07342050224542618,
-0.012878159061074257,
0.06793811917304993,
0.03247448429465294,
0.004688137210905552,
-0.04549583047628403,
-0.021366821601986885,
-0.038556862622499466,
0.0555274598300457,
0.014559660106897354,
0.07532159239053726,
0.005220728926360607,
-0.00748575059697032,
0.045095887035131454,
-0.007967950776219368,
0.08766070753335953,
0.020439090207219124,
0.006262239534407854,
0.041895102709531784,
0.00711891008540988,
0.06364350020885468,
0.007390980143100023,
0.003765186993405223,
0.013680027797818184,
0.025138232856988907,
-0.11471523344516754,
0.032824382185935974,
0.006063726730644703,
0.012215305119752884,
-0.01651563122868538,
-0.009641374461352825,
0.03628045320510864,
-0.0066802650690078735,
0.01866820454597473,
0.0010640211403369904,
-0.07959765195846558,
-0.01627839356660843,
0.022901048883795738,
0.03730819746851921,
-0.1099497377872467,
0.025670126080513,
-0.04395054653286934,
-0.04287603124976158,
0.009411555714905262,
-0.028704369440674782,
-0.06237274408340454,
0.03168528899550438,
3.0506861019371948e-33,
0.06882718205451965,
0.0021013047080487013,
-0.05342858284711838,
-0.01053268276154995,
0.026230821385979652,
-0.016756068915128708,
0.007234483491629362,
0.03223583474755287,
-0.024922583252191544,
-0.06035817787051201,
-0.0020965656731277704,
0.009314442053437233,
-0.01653486117720604,
-0.0903949961066246,
0.012937282212078571,
0.001477716607041657,
0.04196866229176521,
-0.016868283972144127,
-0.13994117081165314,
0.03553604334592819,
-0.019133735448122025,
0.04499274492263794,
-0.04671873897314072,
0.06165461614727974,
-0.05701432004570961,
0.11190188676118851,
-0.04354432597756386,
0.0043419767171144485,
-0.05642088130116463,
-0.021066946908831596,
0.012185348197817802,
-0.008937285281717777,
0.04899594932794571,
-0.06639079749584198,
-0.08787184208631516,
0.03793058171868324,
0.04031119868159294,
0.009324257262051105,
-0.03764867037534714,
0.20795299112796783,
0.05766953155398369,
-0.011612022295594215,
-0.044502221047878265,
-0.012499193660914898,
-0.03616233542561531,
-0.06394806504249573,
-0.0830664336681366,
-0.051214154809713364,
0.07295543700456619,
0.08400530368089676,
0.023538799956440926,
-0.03427649661898613,
-0.10071767866611481,
-0.004750359337776899,
0.007361576426774263,
-0.12820802628993988,
-0.05242922902107239,
-0.005547261331230402,
0.07416601479053497,
0.024467833340168,
-0.024743877351284027,
-0.07236141711473465,
-0.06405946612358093,
-0.02876296080648899,
-0.015698930248618126,
-0.0245805773884058,
-0.05607591196894646,
-0.04423126205801964,
-0.006404226180166006,
0.026796506717801094,
-0.009172884747385979,
-0.010950379073619843,
0.04779541864991188,
0.02911529317498207,
0.05746553838253021,
0.02677459828555584,
0.022691046819090843,
0.04639256373047829,
0.03470991924405098,
-0.0947653204202652,
-0.020345080643892288,
-0.029181409627199173,
0.02522188238799572,
0.08984141796827316,
0.04816148430109024,
0.04224663972854614,
-0.009510251693427563,
0.05781971663236618,
0.053802236914634705,
0.014347477816045284,
-0.031848449259996414,
-0.021663496270775795,
-0.01367194950580597,
0.030054593458771706,
0.03384561091661453,
-3.442092477712322e-8,
-0.05459895357489586,
0.01445652637630701,
-0.055035222321748734,
-0.009190671145915985,
-0.011191644705832005,
-0.0038186763413250446,
-0.007048168685287237,
0.07320628315210342,
0.03131779283285141,
-0.030168920755386353,
0.07201562821865082,
-0.056119658052921295,
-0.035082921385765076,
0.0025732789654284716,
-0.021810676902532578,
0.0447646826505661,
0.004920937120914459,
0.04926508292555809,
0.021574726328253746,
-0.007482993882149458,
0.02274271845817566,
0.033775974065065384,
0.023918891325592995,
-0.049828559160232544,
0.01329298410564661,
-0.017320094630122185,
-0.09135282784700394,
0.06485066562891006,
0.03496764600276947,
-0.02981426566839218,
-0.011405343189835548,
-0.035329610109329224,
0.017824437469244003,
-0.07299597561359406,
0.026605064049363136,
0.02985641360282898,
0.02685924433171749,
-0.07002239674329758,
-0.028543313965201378,
-0.05021629109978676,
0.06389936804771423,
0.09010902792215347,
-0.02368067018687725,
0.0391094796359539,
-0.008867613971233368,
-0.02090761996805668,
0.05783470347523689,
-0.028294038027524948,
0.04238179698586464,
0.010140735656023026,
-0.04753730818629265,
-0.11044174432754517,
0.044718097895383835,
0.0601160041987896,
0.04864800348877907,
-0.049645014107227325,
0.02121128886938095,
0.0232976246625185,
0.12060701102018356,
0.05746329948306084,
0.0742313489317894,
0.021201597526669502,
-0.024554667994379997,
-0.01759255677461624
] |
Cross-Embodiment Dexterous Grasping with Reinforcement Learning | https://openreview.net/forum?id=twIPSx9qHn | [
"Haoqi Yuan",
"Bohan Zhou",
"Yuhui Fu",
"Zongqing Lu"
] | Poster | Dexterous hands exhibit significant potential for complex real-world grasping tasks. While recent studies have primarily focused on learning policies for specific robotic hands, the development of a universal policy that controls diverse dexterous hands remains largely unexplored.
In this work, we study the learning of cross-embodiment dexterous grasping policies using reinforcement learning (RL). Inspired by the capability of human hands to control various dexterous hands through teleoperation, we propose a universal action space based on the human hand's eigengrasps. The policy outputs eigengrasp actions that are then converted into specific joint actions for each robot hand through a retargeting mapping. We simplify the robot hand's proprioception to include only the positions of fingertips and the palm, offering a unified observation space across different robot hands. Our approach demonstrates an 80\% success rate in grasping objects from the YCB dataset across four distinct embodiments using a single vision-based policy. Additionally, our policy exhibits zero-shot generalization to two previously unseen embodiments and significant improvement in efficient finetuning. For further details and videos, visit our project page (https://sites.google.com/view/crossdex). | dexterous grasping, cross-embodiment learning, reinforcement learning | null | 1,064 | 2410.02479 | [
-0.05897464230656624,
-0.11385957151651382,
0.020958153530955315,
-0.03702135011553764,
-0.04302855208516121,
0.023589229211211205,
0.07236484438180923,
-0.06072002649307251,
-0.02557668276131153,
0.05519988015294075,
0.09263210743665695,
-0.0607050396502018,
-0.024313785135746002,
0.06289172917604446,
0.05194186419248581,
0.0211208313703537,
0.07938092201948166,
0.013669691048562527,
-0.08086594194173813,
0.0756082534790039,
0.04916162043809891,
0.03205123916268349,
0.11013215035200119,
-0.011886822991073132,
-0.0687873587012291,
-0.02115008793771267,
0.025615384802222252,
-0.051056157797575,
0.0347595177590847,
-0.060746025294065475,
-0.036723941564559937,
-0.02644970454275608,
-0.05179142579436302,
-0.09467119723558426,
-0.058587294071912766,
0.07830644398927689,
-0.041346289217472076,
0.008745687082409859,
-0.018920138478279114,
-0.033667005598545074,
-0.011037150397896767,
0.010086381807923317,
0.04958147183060646,
0.05799323692917824,
0.0776391252875328,
0.08293946087360382,
0.018506968393921852,
-0.02982317842543125,
0.08822104334831238,
-0.022677123546600342,
0.005305174272507429,
-0.016011858358979225,
0.05315098538994789,
0.028891783207654953,
0.047150254249572754,
0.033706095069646835,
-0.017488980665802956,
-0.06348428875207901,
0.05475243180990219,
-0.010082884691655636,
0.0163352582603693,
-0.06612653285264969,
-0.01663271337747574,
-0.0032346369698643684,
-0.04825688898563385,
-0.09277629107236862,
0.04586879909038544,
-0.043070536106824875,
-0.034800950437784195,
0.0027842882554978132,
-0.03829016909003258,
-0.013180709443986416,
-0.007386133074760437,
0.036886706948280334,
0.025825561955571175,
0.02077537775039673,
-0.008435823954641819,
-0.021740837022662163,
-0.024111710488796234,
-0.00299968128092587,
0.013461893424391747,
0.003104225266724825,
-0.03076738491654396,
0.06107522174715996,
0.03238943964242935,
-0.04387042671442032,
-0.08376076817512512,
0.04193910211324692,
0.024001125246286392,
-0.025480354204773903,
-0.02916952781379223,
-0.03728214651346207,
-0.02053247205913067,
-0.08463732898235321,
-0.029738396406173706,
-0.017266033217310905,
-0.0009048954816535115,
-0.10439573973417282,
-0.12136706709861755,
0.030184727162122726,
-0.003371825208887458,
0.0024915237445384264,
-0.0016809396911412477,
0.018995027989149094,
0.019143052399158478,
0.002759459661319852,
-0.04046747088432312,
-0.05634700879454613,
0.0037283189594745636,
-0.038045674562454224,
0.019383341073989868,
-0.06720401346683502,
-0.03986899182200432,
0.05874110013246536,
0.0054863509722054005,
-0.05246887728571892,
-0.03333727642893791,
-0.008554964326322079,
0.051823534071445465,
0.04228639602661133,
-0.033759504556655884,
-0.028639869764447212,
0.018959756940603256,
0.03193444013595581,
0.03957065939903259,
-0.03530459851026535,
-0.02528674528002739,
1.8237513836389547e-33,
0.11891621351242065,
0.03893520310521126,
0.045586373656988144,
-0.005448560696095228,
0.005598719231784344,
0.0010378564475104213,
0.004367797635495663,
0.03330571576952934,
-0.046671364456415176,
0.0770958662033081,
0.05783616378903389,
-0.005578687880188227,
-0.023194490000605583,
0.13333760201931,
0.015602834522724152,
0.011694183573126793,
-0.05564650520682335,
0.05190686136484146,
-0.05176299810409546,
0.0639946460723877,
0.03940141201019287,
0.027760252356529236,
-0.010361616499722004,
-0.07021820545196533,
-0.005487528163939714,
-0.02627520076930523,
-0.005022791214287281,
0.01918305642902851,
0.08203081786632538,
0.015349394641816616,
-0.020801832899451256,
0.06213834881782532,
-0.06927095353603363,
0.11382383108139038,
0.03302006796002388,
0.03644411265850067,
-0.03426291048526764,
-0.03723599761724472,
0.003007258754223585,
-0.05189712718129158,
0.018217535689473152,
0.011224760673940182,
-0.030738063156604767,
-0.02619771659374237,
-0.029540877789258957,
-0.08899369835853577,
0.04246586933732033,
0.03256123140454292,
-0.062355753034353256,
-0.03833252936601639,
0.04119852930307388,
0.02676471881568432,
-0.02857772260904312,
-0.05989968031644821,
-0.044957950711250305,
-0.001617635483853519,
-0.039071593433618546,
0.09790335595607758,
-0.004415728617459536,
0.007347182370722294,
-0.013903811573982239,
0.014796720817685127,
-0.054789651185274124,
0.0850793719291687,
-0.014216611161828041,
0.1047447919845581,
0.030048595741391182,
0.049218956381082535,
0.0870152935385704,
-0.023665012791752815,
-0.021025486290454865,
0.0007911434513516724,
0.02180291898548603,
-0.06728321313858032,
0.0954231470823288,
0.014217568561434746,
0.03748907521367073,
-0.0707402229309082,
0.01264174934476614,
0.08175551891326904,
0.007301909849047661,
0.06620081514120102,
0.009796792641282082,
-0.006489455699920654,
0.01708272472023964,
-0.11651922762393951,
0.02200910449028015,
-0.04339355230331421,
-0.042712029069662094,
-0.01990773342549801,
-0.04545440152287483,
-0.044094447046518326,
-0.06311958283185959,
0.02993990294635296,
-0.04987606033682823,
-2.8552357378710115e-33,
0.002284395042806864,
-0.008349996991455555,
0.05589217692613602,
0.01834896020591259,
-0.06510796397924423,
-0.11088171601295471,
0.028515785932540894,
-0.04626638814806938,
-0.04387075453996658,
-0.10496938228607178,
-0.08519275486469269,
0.022375425323843956,
-0.012053156271576881,
-0.026415636762976646,
-0.0023585641756653786,
0.04246634244918823,
0.001436657621525228,
0.04304495453834534,
0.008623692207038403,
-0.05214677378535271,
0.06604941934347153,
0.05254807695746422,
0.06250204890966415,
0.07868797332048416,
-0.05977727100253105,
0.023723023012280464,
0.03773881122469902,
-0.004472552333027124,
-0.026227666065096855,
0.03204444423317909,
-0.0569361075758934,
-0.05665503069758415,
-0.013726944103837013,
0.051367394626140594,
-0.08567949384450912,
0.03542815148830414,
-0.04279569163918495,
0.07257869839668274,
0.042326364666223526,
0.11397486925125122,
0.043804578483104706,
-0.07335738837718964,
-0.024165349081158638,
0.0015022095758467913,
0.00903855636715889,
-0.0464964285492897,
-0.10527011007070541,
0.022351732477545738,
-0.026430487632751465,
0.03175940364599228,
-0.06576419621706009,
0.07038804888725281,
-0.10304680466651917,
-0.09023423492908478,
-0.013717984780669212,
-0.058094315230846405,
-0.009670533239841461,
-0.03400523588061333,
0.10363731533288956,
-0.016774358227849007,
-0.0186311025172472,
0.02775089628994465,
-0.019248588010668755,
0.12349973618984222,
0.041704677045345306,
0.05152364820241928,
0.02236267738044262,
0.0074128657579422,
0.024189641699194908,
0.009395604021847248,
0.05180272087454796,
0.028339819982647896,
0.07048395276069641,
-0.11609191447496414,
0.10264486074447632,
0.016589974984526634,
-0.0017161982832476497,
-0.07548613846302032,
0.015438208356499672,
-0.009617353789508343,
0.03585582599043846,
-0.0658225268125534,
0.07082317769527435,
0.1418958306312561,
0.008543949574232101,
0.12250849604606628,
-0.018724454566836357,
0.01977463811635971,
0.0030976205598562956,
0.01895458996295929,
-0.04061722010374069,
0.0471777506172657,
-0.00008011373574845493,
0.06905768811702728,
0.0705801472067833,
-4.086130900304852e-8,
-0.02102108672261238,
0.09909240901470184,
0.09420863538980484,
0.0036168114747852087,
0.03516654297709465,
0.058085206896066666,
-0.040876276791095734,
-0.038506682962179184,
-0.04085791856050491,
-0.01976047269999981,
0.018327457830309868,
0.046810001134872437,
0.046462107449769974,
-0.05173925682902336,
-0.04820423945784569,
0.018131837248802185,
0.015412131324410439,
-0.04992511495947838,
-0.003065347671508789,
0.009250607341527939,
0.0842270702123642,
-0.020398560911417007,
0.09384585171937943,
-0.0593903511762619,
0.004736708011478186,
-0.1006285697221756,
-0.14145396649837494,
0.019808921962976456,
-0.0021912145894020796,
-0.022685959935188293,
0.05099193751811981,
-0.06516172736883163,
0.005084999371320009,
0.003993920981884003,
0.014033731073141098,
-0.015772022306919098,
-0.03604889288544655,
-0.03090130351483822,
0.016574183478951454,
-0.004562066402286291,
-0.02334451489150524,
0.07538022100925446,
-0.0069384491071105,
0.024518925696611404,
-0.005544919520616531,
0.0765131488442421,
-0.005144888069480658,
-0.10618303716182709,
0.011047509498894215,
-0.04715660214424133,
0.024230802431702614,
-0.0023855443578213453,
0.011734583415091038,
0.0696316808462143,
0.08669841289520264,
0.014445111155509949,
-0.02220894955098629,
-0.054367855191230774,
0.013389396481215954,
0.10011444240808487,
-0.03166801109910011,
-0.0009091263636946678,
-0.11086671054363251,
0.004633069038391113
] |
VEDIT: Latent Prediction Architecture For Procedural Video Representation Learning | https://openreview.net/forum?id=LDAj4UJ4aL | [
"Han Lin",
"Tushar Nagarajan",
"Nicolas Ballas",
"Mido Assran",
"Mojtaba Komeili",
"Mohit Bansal",
"Koustuv Sinha"
] | Poster | Procedural video representation learning is an active research area where the objective is to learn an agent which can anticipate and forecast the future given the present video input, typically in conjunction with textual annotations. Prior works often rely on large-scale pretraining of visual encoders and prediction models with language supervision. However, the necessity and effectiveness of extending compute intensive pretraining to learn video clip sequences with noisy text supervision have not yet been fully validated by previous works. In this work, we show that a strong off-the-shelf frozen pretrained visual encoder, along with a well designed prediction model, can achieve state-of-the-art (SoTA) performance in forecasting and procedural planning without the need for pretraining the prediction model, nor requiring additional supervision from language or ASR. Instead of learning representations from pixel space, our method utilizes the latent embedding space of publicly available vision encoders. By conditioning on frozen clip-level embeddings from observed steps to predict the actions of unseen steps, our prediction model is able to learn robust representations for forecasting through iterative denoising —leveraging the recent advances in diffusion transformers (Peebles & Xie, 2023). Empirical studies over a total of five procedural learning tasks across four datasets (NIV, CrossTask, COIN and Ego4D-v2) show that our model advances the strong baselines in long-horizon action anticipation (+2.6% in Verb ED@20, +3.1% in Noun ED@20), and significantly improves the SoTA in step forecasting (+5.0%), task classification (+3.8%), and procedure planning tasks (up to +2.28% in success rate, +3.39% in mAcc, and +0.90% in mIoU). | Procedural Learning from Videos, Representation Learning, Diffusion Transformer | null | 1,060 | 2410.03478 | [
-0.04154942184686661,
-0.1626516878604889,
0.03875666856765747,
0.031423937529325485,
0.05791661515831947,
0.07148925215005875,
-0.08141294121742249,
-0.06823619455099106,
0.01936187408864498,
-0.07350166887044907,
-0.05126681551337242,
-0.05499028414487839,
-0.028450829908251762,
0.11058545112609863,
-0.056530971080064774,
0.00551238376647234,
0.08030160516500473,
0.07925950735807419,
-0.04266258701682091,
-0.0738755539059639,
0.057632699608802795,
-0.026714395731687546,
0.022999165579676628,
0.04322069138288498,
0.11425267904996872,
0.052716828882694244,
0.014848788268864155,
-0.021376866847276688,
0.03213300555944443,
0.00956661719828844,
0.04458573833107948,
-0.03573956340551376,
0.0006975495489314198,
-0.0017021950334310532,
-0.08895990252494812,
0.08022201061248779,
-0.04331820085644722,
-0.012281820178031921,
-0.025295432657003403,
-0.020305030047893524,
0.0046211970038712025,
0.041093289852142334,
0.018981941044330597,
-0.002302220556885004,
0.05721656233072281,
-0.025080811232328415,
0.024023503065109253,
-0.09308552742004395,
-0.043831463903188705,
0.0016930189449340105,
-0.07823183387517929,
-0.02944343350827694,
-0.024992037564516068,
-0.021977664902806282,
-0.03554948791861534,
0.03977624699473381,
0.10477056354284286,
-0.01710386388003826,
0.02008674293756485,
-0.022199710831046104,
-0.002059198683127761,
-0.11911248415708542,
0.0310992244631052,
-0.03541053086519241,
-0.031533382833004,
-0.03548845276236534,
0.057105932384729385,
0.01299482025206089,
-0.04214790463447571,
0.007205136585980654,
-0.06102333962917328,
0.016231633722782135,
-0.0025210902094841003,
0.05512464419007301,
-0.0009178672335110605,
0.030097683891654015,
0.06159226596355438,
-0.04033493250608444,
0.062777079641819,
-0.06963934004306793,
0.07108883559703827,
0.001964246155694127,
0.05377943441271782,
0.01725999265909195,
0.05460710451006889,
0.05256408452987671,
0.04623382166028023,
0.10121909528970718,
0.03148353472352028,
-0.025416351854801178,
-0.03015902079641819,
-0.03628595173358917,
-0.000934544310439378,
0.0030433342326432467,
0.026651015505194664,
0.07119633257389069,
-0.02153475396335125,
-0.09380015730857849,
-0.012787417508661747,
0.017269304022192955,
0.009594461880624294,
0.03111501783132553,
0.021534616127610207,
0.013350355438888073,
0.024303598329424858,
-0.006785203702747822,
0.13297787308692932,
0.004797239322215319,
-0.027924491092562675,
-0.07548026740550995,
0.0395437590777874,
-0.009866999462246895,
0.046045586466789246,
0.05024499446153641,
0.04485803097486496,
0.0229751318693161,
0.0042734392918646336,
0.024674084037542343,
0.008960900828242302,
0.06688883155584335,
0.008433813229203224,
0.020932577550411224,
0.005567663814872503,
0.013786064460873604,
0.10018070042133331,
-0.047787200659513474,
-0.03756950423121452,
1.5942801697201849e-34,
0.04005395621061325,
0.0043001496233046055,
0.11082162708044052,
0.012831643223762512,
0.08131516724824905,
-0.005334979854524136,
0.0015363296261057258,
0.06604474782943726,
-0.024765852838754654,
-0.040423814207315445,
0.03416679799556732,
-0.09346376359462738,
-0.08223380893468857,
0.11305350065231323,
0.006097479723393917,
-0.0354999341070652,
-0.04568493366241455,
-0.011220562271773815,
-0.043038275092840195,
0.04253622144460678,
0.001626783749088645,
-0.006530458107590675,
-0.03319105505943298,
-0.05533917620778084,
-0.02388598956167698,
-0.0006683441461063921,
0.07211535423994064,
-0.09420594573020935,
0.00997812207788229,
-0.03197186067700386,
-0.09864092618227005,
0.0206480473279953,
0.03182557597756386,
0.02804535999894142,
0.012378526851534843,
-0.02792721427977085,
-0.0319904163479805,
-0.025837430730462074,
0.02958143875002861,
-0.012697821483016014,
0.05353627726435661,
0.055061303079128265,
-0.014369743876159191,
-0.06588752567768097,
-0.08504586666822433,
-0.08078862726688385,
0.035949669778347015,
0.07783437520265579,
-0.05295971781015396,
-0.014383969828486443,
0.09426724165678024,
0.036906808614730835,
-0.06306712329387665,
-0.12246225029230118,
-0.046206776052713394,
-0.009584221057593822,
0.09623324126005173,
0.002459948416799307,
-0.0018558085430413485,
-0.01923675648868084,
-0.03348314017057419,
0.0043084826320409775,
0.01649707555770874,
0.06248104199767113,
-0.007780604995787144,
-0.015560522675514221,
0.01948641985654831,
0.05516399070620537,
0.024395916610956192,
-0.06618989259004593,
-0.05225469172000885,
0.07179300487041473,
-0.029822643846273422,
-0.09856347739696503,
0.006977173965424299,
0.01948658749461174,
-0.031188135966658592,
-0.10391415655612946,
0.01811210811138153,
0.05680624395608902,
-0.021945985034108162,
-0.02933775633573532,
-0.004418586380779743,
-0.020595232024788857,
0.013776512816548347,
-0.04246386140584946,
0.05127529054880142,
-0.05371086299419403,
0.043640051037073135,
0.017233990132808685,
-0.07330042868852615,
-0.09348578006029129,
-0.03954459726810455,
-0.0017581006977707148,
0.06256882101297379,
-6.4532780191346385e-34,
0.01988946460187435,
0.05231177806854248,
-0.04690662771463394,
0.051183827221393585,
0.0455031618475914,
-0.04775994271039963,
-0.012996350415050983,
0.033589918166399,
0.046102192252874374,
-0.07627817988395691,
-0.00972179975360632,
0.0012747463770210743,
0.002569475444033742,
-0.007553989998996258,
0.029134780168533325,
-0.04389815777540207,
-0.0070553426630795,
0.03813629597425461,
-0.03326157480478287,
-0.05514631047844887,
0.01754804700613022,
0.010480468161404133,
-0.0143435038626194,
0.06334808468818665,
-0.02944245934486389,
0.02293555811047554,
0.06877852231264114,
0.06886246055364609,
-0.013780219480395317,
0.030263788998126984,
0.009034178219735622,
-0.06959740817546844,
0.013515529222786427,
-0.026592113077640533,
-0.06755007803440094,
0.049857303500175476,
0.06841951608657837,
-0.02826201170682907,
-0.09041862189769745,
0.05027499794960022,
0.1022355854511261,
-0.023172562941908836,
-0.0436546765267849,
0.014234748668968678,
-0.054987482726573944,
-0.0498117133975029,
-0.03402760997414589,
0.0009468717616982758,
0.03824010118842125,
0.0941319614648819,
-0.07021842151880264,
0.06634434312582016,
-0.12593789398670197,
-0.026914427056908607,
-0.039732735604047775,
-0.07017318904399872,
-0.06651604920625687,
-0.004816317930817604,
0.083726666867733,
-0.010792478919029236,
-0.03396448493003845,
0.060685135424137115,
-0.027085542678833008,
-0.04340401291847229,
0.018256977200508118,
-0.06366563588380814,
-0.09428425878286362,
0.031146079301834106,
0.06765230745077133,
-0.034754425287246704,
0.05607222020626068,
-0.014610446989536285,
-0.02284565009176731,
0.0524006113409996,
0.04329012334346771,
-0.023881545290350914,
0.028050091117620468,
-0.0013259821571409702,
0.02960345149040222,
-0.03916056081652641,
-0.014410510659217834,
-0.03723615035414696,
0.0020611542277038097,
0.04001264274120331,
0.10208208113908768,
0.09677103161811829,
0.045085906982421875,
0.03396982327103615,
0.01992405392229557,
0.009098296985030174,
-0.04245853051543236,
0.028710249811410904,
-0.019716033712029457,
0.05610328167676926,
-0.044897403568029404,
-3.399411596660684e-8,
-0.12296950817108154,
0.006546095944941044,
0.01263226568698883,
0.0003706463030539453,
0.010024726390838623,
-0.07641345262527466,
0.08492328226566315,
0.04129193723201752,
0.008668950758874416,
-0.008623087778687477,
0.13166135549545288,
-0.033822208642959595,
0.000772605650126934,
-0.06529330462217331,
0.01691126823425293,
0.020474771037697792,
0.06878736615180969,
0.007599464152008295,
-0.009487012401223183,
0.0012741467216983438,
0.01465162169188261,
-0.024743271991610527,
-0.004946527536958456,
-0.03938642516732216,
-0.031831081956624985,
-0.040798865258693695,
-0.04170287400484085,
0.10775551199913025,
0.08930973708629608,
0.016980133950710297,
0.05159151181578636,
-0.008186989463865757,
0.09705573320388794,
-0.07782692462205887,
-0.04907514899969101,
0.005592448636889458,
0.018623115494847298,
-0.06356440484523773,
0.09351758658885956,
-0.047629088163375854,
0.020298946648836136,
0.017755184322595596,
0.032134369015693665,
-0.027741562575101852,
-0.014398780651390553,
0.03227628767490387,
-0.014291868545114994,
-0.14549590647220612,
0.04375654086470604,
-0.04331506788730621,
-0.06354089081287384,
-0.049753282219171524,
0.0014450537273660302,
0.10320181399583817,
0.1063748374581337,
0.010815093293786049,
0.034142252057790756,
0.0018508554203435779,
0.043598029762506485,
0.0741768404841423,
-0.04070597514510155,
0.01171040441840887,
0.015240278095006943,
0.03866373747587204
] |
Refine-by-Align: Reference-Guided Artifacts Refinement through Semantic Alignment | https://openreview.net/forum?id=D9CRb1KZQc | [
"Yizhi Song",
"Liu He",
"Zhifei Zhang",
"Soo Ye Kim",
"He Zhang",
"Wei Xiong",
"Zhe Lin",
"Brian L. Price",
"Scott Cohen",
"Jianming Zhang",
"Daniel Aliaga"
] | Poster | Personalized image generation has emerged from the recent advancements in generative models. However, these generated personalized images often suffer from localized artifacts such as incorrect logos, reducing fidelity and fine-grained identity details of the generated results. Furthermore, there is little prior work tackling this problem. To help improve these identity details in the personalized image generation, we introduce a new task: reference-guided artifacts refinement. We present Refine-by-Align, a first-of-its-kind model that employs a diffusion-based framework to address this challenge. Our model consists of two stages: Alignment Stage and Refinement Stage, which share weights of a unified neural network model. Given a generated image, a masked artifact region, and a reference image, the alignment stage identifies and extracts the corresponding regional features in the reference, which are then used by the refinement stage to fix the artifacts. Our model-agnostic pipeline requires no test-time tuning or optimization. It automatically enhances image fidelity and reference identity in the generated image, generalizing well to existing models on various tasks including but not limited to customization, generative compositing, view synthesis, and virtual try-on. Extensive experiments and comparisons demonstrate that our pipeline greatly pushes the boundary of fine details in the image synthesis models. | diffusion model; inpainting; generative artifacts; image editing; image synthesis; artifacts refinement | null | 1,040 | null | [
-0.024113109335303307,
-0.047881830483675,
0.06148035451769829,
0.03384718298912048,
0.04222867637872696,
0.003096402855589986,
-0.05045284330844879,
-0.08876359462738037,
0.008277964778244495,
-0.0852627232670784,
-0.036733515560626984,
-0.04341720789670944,
0.02980566956102848,
0.010620471090078354,
-0.09640546888113022,
-0.023569315671920776,
0.05462653189897537,
0.07500233501195908,
-0.07812907546758652,
-0.05918106064200401,
0.012421227060258389,
0.011503662914037704,
0.010668589733541012,
-0.038810137659311295,
0.010119681246578693,
0.031213583424687386,
-0.00022886872466187924,
-0.014695332385599613,
0.061520952731370926,
-0.10771550983190536,
0.041745010763406754,
0.05222547426819801,
-0.10724195837974548,
0.02923356555402279,
-0.037410054355859756,
0.13573729991912842,
-0.047874461859464645,
0.05764388665556908,
0.061263520270586014,
-0.04917670413851738,
0.04445446655154228,
0.023540183901786804,
0.014926863834261894,
0.022919898852705956,
0.119985431432724,
-0.002845062641426921,
0.02774828113615513,
-0.006287934258580208,
0.024518167600035667,
0.012495412491261959,
-0.06690441817045212,
-0.018234306946396828,
-0.04023072123527527,
0.07383906096220016,
0.021169815212488174,
0.09195253998041153,
0.07807474583387375,
-0.02394096739590168,
0.01689080148935318,
0.003363001858815551,
-0.05561748892068863,
-0.03056268021464348,
-0.01261869166046381,
0.009210558608174324,
-0.022306984290480614,
-0.024463390931487083,
-0.023946933448314667,
-0.00011506499140523374,
-0.009628665633499622,
-0.0013645989820361137,
0.03852037712931633,
0.03061019815504551,
0.01648533158004284,
-0.004853391088545322,
0.03164542093873024,
0.11488610506057739,
0.026704082265496254,
0.007773430086672306,
0.0635199099779129,
-0.07421420514583588,
0.016607802361249924,
0.018397381529211998,
0.07906992733478546,
-0.0343661792576313,
0.0561290979385376,
-0.004662869963794947,
-0.006769081577658653,
0.03286846727132797,
0.06463947147130966,
0.006706735584884882,
-0.07509659230709076,
-0.04547874629497528,
-0.008150622248649597,
-0.029992520809173584,
0.057207152247428894,
0.05287208408117294,
0.042457450181245804,
-0.07214535027742386,
-0.043627601116895676,
0.042372673749923706,
-0.033395007252693176,
0.059906382113695145,
0.023922089487314224,
-0.0681607648730278,
0.0324096642434597,
0.00036302642547525465,
0.010844668373465538,
0.06069142371416092,
-0.01888715662062168,
-0.0579998679459095,
0.043502625077962875,
-0.02359919063746929,
-0.032722268253564835,
-0.026365108788013458,
0.023668020963668823,
-0.04820883646607399,
-0.018936999142169952,
-0.07925257831811905,
0.018983395770192146,
-0.04205944761633873,
-0.0439191535115242,
0.07824310660362244,
-0.09964330494403839,
0.04918734356760979,
0.0032218743581324816,
-0.05626323074102402,
-0.039441127330064774,
1.207887669053088e-33,
-0.005165751092135906,
0.019214127212762833,
0.03454793617129326,
0.08614213019609451,
0.10367945581674576,
-0.033295888453722,
-0.04554218053817749,
0.014947984367609024,
-0.04195011779665947,
-0.038177814334630966,
-0.040534503757953644,
-0.01256575994193554,
-0.1064886674284935,
0.11073217540979385,
-0.04571305960416794,
-0.00847671926021576,
-0.10263649374246597,
0.08070050179958344,
-0.05635726824402809,
0.03712201863527298,
-0.07754267752170563,
0.009459002874791622,
-0.048239029943943024,
-0.028367334976792336,
-0.017375560477375984,
-0.06197385489940643,
0.08146966993808746,
-0.01389540545642376,
-0.07341094315052032,
-0.018661146983504295,
-0.0846814215183258,
-0.012215303257107735,
0.0666067823767662,
0.03466244414448738,
-0.06522253155708313,
-0.033477526158094406,
0.035827480256557465,
-0.01409994252026081,
-0.006323880050331354,
0.07686172425746918,
0.05982259288430214,
0.09245635569095612,
-0.0751497894525528,
0.05419682338833809,
-0.04987849295139313,
-0.047461289912462234,
0.04205437749624252,
0.038946304470300674,
0.03154776245355606,
-0.02629014104604721,
0.031371861696243286,
0.05109198018908501,
0.020387077704072,
-0.04339105263352394,
-0.12943829596042633,
-0.03761812672019005,
0.044467613101005554,
-0.04687698185443878,
0.02244909666478634,
-0.015304249711334705,
0.05524342879652977,
0.035717904567718506,
-0.06858029961585999,
0.14288835227489471,
0.04303702712059021,
-0.0035559225361794233,
0.06988025456666946,
0.0657147690653801,
0.04903067275881767,
0.04608113318681717,
-0.057711247354745865,
0.009902175515890121,
-0.008059769868850708,
-0.004860279615968466,
0.060302454978227615,
-0.004123289603739977,
0.03789643570780754,
0.00466093048453331,
-0.027746403589844704,
0.031229760497808456,
-0.03980515897274017,
0.047665249556303024,
-0.02883618324995041,
-0.0018637587781995535,
-0.09551498293876648,
-0.0343276783823967,
0.07291728258132935,
-0.03232935443520546,
-0.010040681809186935,
-0.028520921245217323,
0.018491992726922035,
0.012135137803852558,
0.014561118558049202,
0.03065006248652935,
0.014664226211607456,
4.005877256778224e-35,
0.03530323505401611,
0.025726063176989555,
0.09656956791877747,
0.03943490609526634,
-0.0161919966340065,
-0.003887384431436658,
0.008435231633484364,
0.04225163534283638,
-0.015412168577313423,
-0.06462883949279785,
-0.019535595551133156,
-0.07239851355552673,
-0.020659999921917915,
-0.05854141339659691,
-0.061643242835998535,
-0.06400997191667557,
-0.023059073835611343,
0.050262078642845154,
-0.008057519793510437,
0.05303560942411423,
0.08851649612188339,
0.11108643561601639,
-0.08277591317892075,
0.013636283576488495,
-0.08016036450862885,
0.07745963335037231,
-0.016921522095799446,
0.026366958394646645,
0.016817128285765648,
-0.036008644849061966,
-0.01272283960133791,
-0.022228769958019257,
0.05304795876145363,
-0.0779835507273674,
-0.05378623679280281,
0.09254071116447449,
0.044371627271175385,
-0.02441389113664627,
0.01568755693733692,
0.04742824658751488,
0.024398934096097946,
-0.007580578327178955,
-0.1291782259941101,
0.03834260627627373,
0.0016592949395999312,
-0.06716036051511765,
-0.03743386268615723,
-0.032988082617521286,
-0.01417990680783987,
0.07348475605249405,
-0.07571935653686523,
0.05131054297089577,
-0.12768426537513733,
-0.009026214480400085,
-0.05982913821935654,
-0.06269709765911102,
-0.05675320699810982,
0.016679514199495316,
0.09516478329896927,
0.010999116115272045,
-0.01714588701725006,
-0.08653164654970169,
0.012446935288608074,
-0.1068166047334671,
0.02973695658147335,
-0.017067935317754745,
-0.022870376706123352,
0.009247427806258202,
-0.016726426780223846,
0.0501079261302948,
-0.010807051323354244,
-0.0056723556481301785,
0.002162219723686576,
0.04730583354830742,
0.017447063699364662,
-0.09439914673566818,
0.07101050764322281,
-0.01411151047796011,
0.034894589334726334,
-0.0806259959936142,
0.004163559526205063,
0.0016989316791296005,
-0.04679383337497711,
0.09379209578037262,
0.09932263195514679,
0.05201934278011322,
0.015431595034897327,
0.03331785649061203,
0.04182210564613342,
0.057646121829748154,
-0.07196176052093506,
0.007539309561252594,
-0.007826658897101879,
0.01194055750966072,
-0.054460544139146805,
-4.3235459656898456e-8,
-0.04700176790356636,
0.10645667463541031,
0.002828051568940282,
0.08944756537675858,
-0.045887541025877,
-0.09609764069318771,
0.05852077156305313,
0.05326676741242409,
0.02192801609635353,
-0.08737222850322723,
0.009973873384296894,
-0.042269911617040634,
-0.0291021429002285,
-0.06734368205070496,
-0.021256212145090103,
0.031193876639008522,
-0.04379705712199211,
0.047743067145347595,
-0.008084225468337536,
-0.007135393563657999,
-0.02825983241200447,
0.054471760988235474,
0.016471017152071,
-0.04584617540240288,
0.08433368057012558,
-0.019224777817726135,
-0.03666236624121666,
0.05564044043421745,
0.029745206236839294,
-0.027469508349895477,
0.021300746127963066,
-0.0031607376877218485,
0.074090376496315,
-0.037217505276203156,
-0.040508732199668884,
0.03723806142807007,
0.04519704356789589,
0.017912495881319046,
-0.03043912909924984,
-0.08583773672580719,
0.018199214711785316,
0.05874769762158394,
0.02948135882616043,
0.005578599404543638,
0.01627659611403942,
-0.043343886733055115,
0.0254973191767931,
-0.005625884514302015,
0.01097372267395258,
-0.019304949790239334,
0.050858672708272934,
-0.03314331918954849,
0.02402441017329693,
-0.0008002559188753366,
0.06698919087648392,
-0.0532088428735733,
0.03710188716650009,
0.02127840556204319,
0.11304420232772827,
0.07626422494649887,
0.0653848722577095,
0.06023730710148811,
-0.042302150279283524,
-0.02214057557284832
] |
FreCaS: Efficient Higher-Resolution Image Generation via Frequency-aware Cascaded Sampling | https://openreview.net/forum?id=TsBDfe8Ra5 | [
"Zhengqiang ZHANG",
"Ruihuang Li",
"Lei Zhang"
] | Poster | While image generation with diffusion models has achieved a great success, generating images of higher resolution than the training size remains a challenging task due to the high computational cost. Current methods typically perform the entire sampling process at full resolution and process all frequency components simultaneously, contradicting with the inherent coarse-to-fine nature of latent diffusion models and wasting computations on processing premature high-frequency details at early diffusion stages. To address this issue, we introduce an efficient $\textbf{Fre}$quency-aware $\textbf{Ca}$scaded $\textbf{S}$ampling framework, $\textbf{FreCaS}$ in short, for higher-resolution image generation. FreCaS decomposes the sampling process into cascaded stages with gradually increased resolutions, progressively expanding frequency bands and refining the corresponding details. We propose an innovative frequency-aware classifier-free guidance (FA-CFG) strategy to assign different guidance strengths for different frequency components, directing the diffusion model to add new details in the expanded frequency domain of each stage. Additionally, we fuse the cross-attention maps of previous and current stages to avoid synthesizing unfaithful layouts. Experiments demonstrate that FreCaS significantly outperforms state-of-the-art methods in image quality and generation speed. In particular, FreCaS is about 2.86$\times$ and 6.07$\times$ faster than ScaleCrafter and DemoFusion in generating a 2048$\times$2048 image using a pretrained SDXL model and achieves an $\text{FID}_b$ improvement of 11.6 and 3.7, respectively. FreCaS can be easily extended to more complex models such as SD3. The source code of FreCaS can be found at https://github.com/xtudbxk/FreCaS. | generative models, diffusion models, training-free | A frequency-aware cascaded sampling framework for higher-resolution image synthesis using pre-trained diffusion models. | 1,039 | 2410.18410 | [
0.028922174125909805,
-0.05826893821358681,
-0.01309815514832735,
0.05906270816922188,
0.05194167047739029,
-0.04577254503965378,
-0.04568055644631386,
-0.04122079908847809,
0.015501315705478191,
-0.012616370804607868,
-0.032895639538764954,
-0.07413164526224136,
-0.00458685215562582,
-0.024518843740224838,
0.033750325441360474,
-0.040783196687698364,
-0.026145411655306816,
0.0969795510172844,
-0.09263980388641357,
-0.08199644088745117,
0.025968190282583237,
-0.06878474354743958,
-0.010298074223101139,
-0.06517305225133896,
-0.009503157809376717,
0.01086344476789236,
0.004750068299472332,
-0.011953787878155708,
0.04929180070757866,
-0.04704778641462326,
0.03911880776286125,
0.09808304160833359,
-0.021104268729686737,
-0.007085341960191727,
-0.0570102296769619,
0.02888198010623455,
0.01801730878651142,
0.04518602788448334,
-0.0017364714294672012,
-0.01887037418782711,
0.061508022248744965,
0.09231409430503845,
0.01507982611656189,
0.033838819712400436,
0.040137454867362976,
0.0037647911813110113,
0.014237359166145325,
0.000067124143242836,
-0.033449236303567886,
-0.03120509721338749,
-0.013368726707994938,
-0.0722176805138588,
-0.07555360347032547,
0.07531912624835968,
-0.022982846945524216,
0.08853497356176376,
0.03324310481548309,
-0.029921069741249084,
0.056481342762708664,
0.013301484286785126,
-0.1274322271347046,
-0.11250215023756027,
0.006287389434874058,
-0.06380565464496613,
-0.021364683285355568,
0.018701842054724693,
0.0077446200884878635,
0.007109370548278093,
0.0036721143405884504,
0.0017977518728002906,
-0.10745052993297577,
0.027093324810266495,
0.013973831199109554,
0.05578553304076195,
0.048190657049417496,
0.06915247440338135,
-0.005568763706833124,
0.04291195794939995,
0.07332520186901093,
-0.003026873106136918,
0.009978232905268669,
0.028066832572221756,
0.03235189616680145,
-0.05006403848528862,
0.09774299710988998,
0.017750747501850128,
0.041283633559942245,
0.004221567418426275,
0.016982827335596085,
-0.02932882122695446,
-0.044239819049835205,
-0.0032964651472866535,
-0.0944395512342453,
-0.04831041395664215,
-0.11636149883270264,
0.020736556500196457,
0.052690621465444565,
-0.05080335587263107,
0.0064543006010353565,
0.10608551651239395,
0.054255034774541855,
-0.02494949661195278,
0.002024388173595071,
0.0774419903755188,
0.018559394404292107,
-0.03962231054902077,
0.08813890814781189,
0.13440541923046112,
-0.026637952774763107,
-0.03513754531741142,
0.028338951990008354,
0.0415438748896122,
0.025491448119282722,
-0.02832053229212761,
0.10952598601579666,
0.004554538521915674,
-0.03807814419269562,
-0.03065091371536255,
0.05068141594529152,
-0.038279939442873,
-0.02884003333747387,
-0.018842758610844612,
-0.13803908228874207,
0.0014449075097218156,
0.020384909585118294,
-0.046032510697841644,
-0.1143701821565628,
2.1354790774945863e-33,
-0.01938333548605442,
0.05784543231129646,
0.08415276557207108,
-0.0223484355956316,
0.0904986560344696,
0.003367822617292404,
0.01774178072810173,
0.008031344041228294,
0.030662333592772484,
0.08566268533468246,
0.017184942960739136,
-0.037991106510162354,
-0.07268528640270233,
0.04903460294008255,
0.013663426041603088,
-0.04981832951307297,
-0.06942768394947052,
0.08012346923351288,
-0.003693569451570511,
-0.007352989166975021,
0.04569327086210251,
-0.041518375277519226,
0.08759792149066925,
-0.032770030200481415,
0.003487285226583481,
0.007700700778514147,
0.03344292193651199,
-0.023503897711634636,
-0.008560458198189735,
-0.000592225173022598,
-0.05720443278551102,
-0.006866670213639736,
-0.000932740920688957,
0.0034275532234460115,
-0.03376567363739014,
-0.04938057065010071,
0.05054184049367905,
-0.023013757541775703,
-0.054003700613975525,
0.04780629277229309,
0.04585346579551697,
0.06064329296350479,
0.03779228404164314,
-0.0035718586295843124,
-0.1327439844608307,
0.008870103396475315,
0.014766410924494267,
0.025279637426137924,
0.01060689426958561,
0.061279598623514175,
-0.006923535373061895,
-0.07726399600505829,
-0.003708382835611701,
-0.07639022171497345,
0.03026258759200573,
0.029144374653697014,
0.046833429485559464,
-0.05129319429397583,
0.002063746564090252,
0.034300949424505234,
0.022562362253665924,
-0.021171968430280685,
-0.0914687067270279,
0.13013046979904175,
-0.005711966194212437,
-0.0050155604258179665,
-0.012674525380134583,
0.08755076676607132,
-0.01682853326201439,
-0.02132660150527954,
-0.014267208985984325,
0.0567040853202343,
0.0032713054679334164,
-0.07590687274932861,
0.10455360263586044,
-0.011551018804311752,
0.07367386668920517,
0.02117774821817875,
0.0330832339823246,
0.031476736068725586,
-0.024398626759648323,
0.03270910680294037,
-0.05588006228208542,
-0.05870494246482849,
-0.04098429158329964,
0.011199717409908772,
-0.0004644831642508507,
-0.0355968251824379,
-0.03553515672683716,
-0.11878672242164612,
-0.07105526328086853,
0.037788670510053635,
-0.010991535149514675,
0.0262929555028677,
-0.06908053904771805,
-7.594249570241214e-34,
0.12816821038722992,
0.04130910336971283,
0.06382115930318832,
-0.007646825164556503,
-0.006499733775854111,
0.011272761970758438,
0.024477701634168625,
0.009658750146627426,
0.092881940305233,
-0.1349475085735321,
-0.019095519557595253,
-0.04529532045125961,
-0.034166038036346436,
-0.07038789242506027,
0.004863149952143431,
-0.01239075232297182,
-0.00952252745628357,
-0.03226405009627342,
0.034775927662849426,
-0.005452654790133238,
-0.060686271637678146,
0.03269720822572708,
0.010867885313928127,
0.022532541304826736,
-0.0381896048784256,
0.0036549607757478952,
-0.008153769187629223,
0.07885169982910156,
0.009370074607431889,
-0.04526043310761452,
-0.01933690719306469,
-0.040275659412145615,
0.10279867053031921,
-0.03600100427865982,
-0.07119858264923096,
0.055339209735393524,
0.023796532303094864,
-0.015220117755234241,
0.0010941078653559089,
0.1158343181014061,
0.024852478876709938,
-0.017883891239762306,
-0.049271754920482635,
0.03134765103459358,
-0.026447497308254242,
-0.030352283269166946,
-0.03982742503285408,
0.03633306547999382,
0.06069429591298103,
0.06305480748414993,
0.013285361230373383,
-0.013168403878808022,
-0.09751774370670319,
-0.031155895441770554,
-0.03138819709420204,
0.004412943031638861,
-0.09485409408807755,
-0.011363359168171883,
0.0658021792769432,
-0.031263336539268494,
-0.0028906078077852726,
-0.04049447551369667,
-0.08879321813583374,
-0.1595614105463028,
0.03361177816987038,
-0.016524529084563255,
-0.04614619165658951,
0.009357807226479053,
0.0803639143705368,
0.1137804239988327,
0.00839681550860405,
-0.04459802806377411,
-0.02318808250129223,
0.0038441296201199293,
0.0007472788565792143,
-0.08191894739866257,
-0.03762724623084068,
-0.08037066459655762,
0.0004972674651071429,
-0.014289647340774536,
-0.008363891392946243,
0.0185424592345953,
-0.017156723886728287,
0.0950094610452652,
0.060642391443252563,
0.030606254935264587,
0.04970315471291542,
-0.005569457076489925,
0.09416051208972931,
-0.005948278587311506,
-0.03187287226319313,
0.05423698574304581,
0.05317222699522972,
0.09600088000297546,
-0.028775298967957497,
-4.671148090551469e-8,
-0.056531134992837906,
0.05367562174797058,
-0.08483138680458069,
-0.002891555428504944,
0.03529287502169609,
-0.08420226722955704,
0.050152551382780075,
0.02098148502409458,
0.02166213095188141,
-0.05397044122219086,
0.09440650045871735,
-0.023249011486768723,
0.006088726222515106,
0.05332212522625923,
0.00822440069168806,
0.015221603214740753,
0.04215816408395767,
0.05977429449558258,
0.010238895192742348,
-0.04056180268526077,
-0.14112064242362976,
0.04169956594705582,
0.06436105072498322,
-0.05454057827591896,
0.060259103775024414,
-0.059534672647714615,
-0.011005441658198833,
0.06913948804140091,
0.08065951615571976,
-0.004211295861750841,
-0.03165328502655029,
-0.04555301368236542,
0.06337902694940567,
0.0011323533253744245,
0.046109892427921295,
0.04747246578335762,
-0.03729768097400665,
-0.016366994008421898,
-0.011400076560676098,
-0.03593175485730171,
-0.007272688206285238,
0.08964241296052933,
-0.04914729297161102,
0.01793634332716465,
-0.01968974806368351,
0.000540560286026448,
0.059928763657808304,
-0.0059747276827692986,
-0.005727531388401985,
-0.039795584976673126,
0.05407429486513138,
0.0027187431696802378,
-0.013980790041387081,
-0.006623989436775446,
0.053916383534669876,
0.0071533601731061935,
0.034406132996082306,
0.01517391949892044,
0.008365336805582047,
0.05723976716399193,
-0.0009504914632998407,
0.02686358243227005,
0.009459789842367172,
-0.02406318485736847
] |
Transformers Learn Low Sensitivity Functions: Investigations and Implications | https://openreview.net/forum?id=4ikjWBs3tE | [
"Bhavya Vasudeva",
"Deqing Fu",
"Tianyi Zhou",
"Elliott Kau",
"Youqi Huang",
"Vatsal Sharan"
] | Poster | Transformers achieve state-of-the-art accuracy and robustness across many tasks, but an understanding of their inductive biases and how those biases differ from other neural network architectures remains elusive. In this work, we identify the sensitivity of the model to token-wise random perturbations in the input as a unified metric which explains the inductive bias of transformers across different data modalities and distinguishes them from other architectures. We show that transformers have lower sensitivity than MLPs, CNNs, ConvMixers and LSTMs, across both vision and language tasks. We also show that this low-sensitivity bias has important implications: i) lower sensitivity correlates with improved robustness; it can also be used as an efficient intervention to further improve the robustness of transformers; ii) it corresponds to flatter minima in the loss landscape; and iii) it can serve as a progress measure for grokking. We support these findings with theoretical results showing (weak) spectral bias of transformers in the NTK regime, and improved robustness due to the lower sensitivity. | transformers, sensitivity, grokking | Transformers have lower sensitivity than alternative architectures, such as LSTMs, MLPs, ConvMixers, and CNNs. Low-sensitivity bias correlates with improved robustness and can serve as a progress measure for grokking. | 1,024 | 2403.06925 | [
-0.08611180633306503,
-0.07948444038629532,
0.016835736110806465,
0.019393956288695335,
0.03188638761639595,
0.038972098380327225,
-0.015590237453579903,
0.010338133201003075,
0.014957821927964687,
-0.07165057212114334,
-0.08685815334320068,
-0.04500366374850273,
0.06549113988876343,
0.020623818039894104,
-0.05158466100692749,
-0.005110020749270916,
0.1150934174656868,
0.10911192744970322,
-0.13008761405944824,
-0.06712240725755692,
0.021220991387963295,
0.050208304077386856,
0.007611778099089861,
0.002209680387750268,
0.08254220336675644,
0.016035296022892,
-0.012756379321217537,
-0.05268453434109688,
-0.007081720512360334,
-0.03286902979016304,
-0.04486057907342911,
-0.04239249229431152,
-0.08294065296649933,
0.015849333256483078,
-0.04437042772769928,
0.04710578918457031,
-0.045149289071559906,
-0.025661272928118706,
0.07733665406703949,
-0.028993062674999237,
0.0224238783121109,
0.0009732445469126105,
0.011701386421918869,
0.002934042364358902,
0.05676034837961197,
0.02611495554447174,
0.06748665124177933,
-0.09593662619590759,
-0.08647565543651581,
-0.0167611725628376,
0.026110662147402763,
-0.0393417552113533,
0.03115021623671055,
0.09554530680179596,
-0.08139558881521225,
0.050200752913951874,
0.00557524710893631,
0.042630020529031754,
-0.009863962419331074,
0.03030114248394966,
-0.019273454323410988,
-0.09045109152793884,
-0.016701672226190567,
0.02762521244585514,
-0.0443374328315258,
-0.020747968927025795,
0.0780612975358963,
-0.0195767804980278,
-0.02779795229434967,
-0.04727691039443016,
-0.026899026706814766,
0.0923113077878952,
-0.017853088676929474,
0.047492388635873795,
0.09303156286478043,
-0.00111112289596349,
0.14142486453056335,
-0.041641149669885635,
0.055673737078905106,
-0.09353578090667725,
0.09050647169351578,
0.05145997926592827,
0.009448654018342495,
-0.001514202682301402,
0.1164432018995285,
-0.049913957715034485,
-0.0284284558147192,
0.04669693112373352,
-0.06059441715478897,
-0.02305254153907299,
-0.04183890298008919,
-0.07689664512872696,
0.0005527600296773016,
0.03219292685389519,
-0.0037631571758538485,
0.03107473812997341,
-0.018347516655921936,
-0.02191564068198204,
-0.06604711711406708,
0.07514581829309464,
-0.021491477265954018,
-0.02452334389090538,
0.042185213416814804,
0.028211072087287903,
0.00567588908597827,
-0.04140491783618927,
0.04297446087002754,
-0.01755906641483307,
-0.028172610327601433,
-0.044302165508270264,
0.05351132154464722,
0.02741737850010395,
0.018782852217555046,
0.012580540962517262,
0.00963563285768032,
-0.026721686124801636,
-0.021766385063529015,
-0.042714521288871765,
-0.020937686786055565,
0.03938109055161476,
-0.05184255540370941,
-0.029281634837388992,
-0.023849645629525185,
0.07260728627443314,
0.016750330105423927,
0.06921736896038055,
-0.03458372503519058,
2.6412301954770983e-33,
-0.012844858691096306,
0.04619379714131355,
0.009989012964069843,
-0.007680568844079971,
0.052043136209249496,
-0.05485748127102852,
0.05567868798971176,
0.012705140747129917,
0.03902334347367287,
0.004286773502826691,
-0.051416970789432526,
0.06143835932016373,
-0.0888066291809082,
0.024152545258402824,
-0.022433875128626823,
0.032459806650877,
-0.013871632516384125,
0.002537840511649847,
-0.02719436213374138,
-0.026869114488363266,
0.018860695883631706,
-0.01742432825267315,
0.0024237774778157473,
-0.04308697208762169,
-0.034822672605514526,
-0.017070092260837555,
0.06318158656358719,
-0.01728476583957672,
-0.0569976270198822,
0.004714821465313435,
-0.06637682765722275,
0.004836501087993383,
0.01978996768593788,
-0.03197430446743965,
0.017277143895626068,
-0.04544750973582268,
-0.011163265444338322,
0.0011107809841632843,
-0.00819670781493187,
-0.052995890378952026,
-0.0237989891320467,
0.08338195085525513,
-0.005740337539464235,
-0.0403430201113224,
-0.025853946805000305,
-0.01920979842543602,
0.04685273766517639,
-0.060760874301195145,
-0.039365414530038834,
-0.04701412469148636,
0.03151663392782211,
0.0157050471752882,
-0.06727544218301773,
0.0115466034039855,
0.08732081949710846,
0.008487979881465435,
0.04044853895902634,
0.027368776500225067,
0.028647100552916527,
0.029489126056432724,
-0.04805602878332138,
0.032899610698223114,
0.0019002245971933007,
0.02572411671280861,
0.01971961185336113,
0.03022698312997818,
0.03547542914748192,
0.0841514840722084,
-0.033951446413993835,
-0.025766877457499504,
-0.030100474134087563,
0.003635501256212592,
-0.0277155302464962,
-0.08444192260503769,
0.046128395944833755,
0.022086091339588165,
0.07213809341192245,
-0.04219818487763405,
-0.02247736230492592,
0.041575510054826736,
0.01256075594574213,
0.0202223751693964,
0.03795502707362175,
-0.010461954399943352,
-0.06559105217456818,
0.011607341468334198,
0.01768605411052704,
-0.0801328793168068,
-0.018328795209527016,
0.01769286021590233,
-0.014526434242725372,
0.023460272699594498,
0.0001988790463656187,
0.004833676852285862,
-0.03835787624120712,
-2.373441929442227e-33,
-0.08367542177438736,
0.08438051491975784,
-0.06689206510782242,
0.11417887359857559,
-0.07697619497776031,
-0.050318196415901184,
0.03405485302209854,
0.06477642804384232,
0.00759948929771781,
0.0008987123728729784,
-0.029749123379588127,
-0.012920976616442204,
-0.02290206030011177,
0.004727715160697699,
0.0797787457704544,
-0.040501948446035385,
-0.048190999776124954,
0.049365609884262085,
0.05146109312772751,
-0.006273393984884024,
0.05406787618994713,
0.17265067994594574,
-0.15085244178771973,
0.004049550276249647,
-0.07689519226551056,
0.05319225788116455,
-0.18656256794929504,
0.06577319651842117,
0.02114737406373024,
-0.060599517077207565,
-0.007465245667845011,
-0.0029034202452749014,
0.00692147808149457,
0.03402433916926384,
-0.007976596243679523,
0.0066643571481108665,
-0.032759469002485275,
-0.01155118178576231,
-0.015872772783041,
0.06598468869924545,
0.039643920958042145,
0.019481884315609932,
-0.04311511665582657,
0.041178878396749496,
-0.01334445457905531,
-0.06455154716968536,
-0.007257383782416582,
-0.08008408546447754,
0.007344966288655996,
0.032592568546533585,
0.04209132865071297,
0.047645002603530884,
-0.11438770592212677,
0.026655644178390503,
-0.0182281956076622,
-0.0472792387008667,
-0.011030401103198528,
-0.015525149181485176,
0.10732468217611313,
0.046134963631629944,
-0.07102923840284348,
-0.030247798189520836,
-0.03182940185070038,
-0.08415189385414124,
-0.025972407311201096,
0.07553106546401978,
-0.06382355839014053,
-0.011612311005592346,
0.12760233879089355,
0.08618371933698654,
0.0653662383556366,
0.011623078025877476,
0.020428072661161423,
-0.07136981934309006,
-0.032195642590522766,
-0.030308915302157402,
-0.0019712375942617655,
-0.07039613276720047,
-0.03471418842673302,
-0.00900429580360651,
-0.01921364665031433,
-0.010461166501045227,
0.006497781723737717,
0.05392010509967804,
0.033554792404174805,
0.124313585460186,
0.011302701197564602,
-0.007229361683130264,
0.04561997950077057,
-0.005028917919844389,
-0.03723404183983803,
0.061416223645210266,
0.0018991193501278758,
0.02189711481332779,
-0.04227335751056671,
-5.1631303676913376e-8,
-0.06452983617782593,
0.08135315775871277,
-0.05606158822774887,
0.02071256935596466,
0.038882698863744736,
-0.0707307979464531,
-0.0013475243467837572,
0.06038990989327431,
-0.03680150955915451,
0.010910075157880783,
0.0916314348578453,
-0.08320537209510803,
-0.036367785185575485,
-0.03158517926931381,
0.03171978145837784,
0.09404770284891129,
-0.005328693427145481,
0.07615087181329727,
0.008470837958157063,
0.021229710429906845,
0.026174627244472504,
0.11551014333963394,
0.011228686198592186,
-0.06288682669401169,
0.03676397725939751,
-0.05593821406364441,
0.0011365441605448723,
0.059779971837997437,
0.044007621705532074,
0.01955990679562092,
-0.049238137900829315,
0.024653064087033272,
0.0032094321213662624,
-0.029819564893841743,
0.029554447159171104,
0.11681678891181946,
0.04687987267971039,
-0.03702230751514435,
-0.007525392808020115,
0.03746432065963745,
0.04494136571884155,
0.06490618735551834,
-0.06335660070180893,
0.0620119534432888,
-0.09604205191135406,
-0.007427735719829798,
0.05605389550328255,
-0.1074567586183548,
-0.018017476424574852,
-0.04836218059062958,
0.01212760154157877,
-0.021300500258803368,
-0.024709448218345642,
0.06100320443511009,
0.03339506313204765,
-0.02646925300359726,
-0.008768171072006226,
-0.10997969657182693,
-0.03353031352162361,
0.10183354467153549,
-0.005698532331734896,
0.050221607089042664,
-0.06625737249851227,
-0.016089944168925285
] |
TLDR: Token-Level Detective Reward Model for Large Vision Language Models | https://openreview.net/forum?id=Zy2XgaGpDw | [
"Deqing Fu",
"Tong Xiao",
"Rui Wang",
"Wang Zhu",
"Pengchuan Zhang",
"Guan Pang",
"Robin Jia",
"Lawrence Chen"
] | Poster | Although reward models have been successful in improving multimodal large language models, the reward models themselves remain brutal and contain minimal information. Notably, existing reward models only mimic human annotations by assigning only one feedback to any text, no matter how long the text is. In the realm of multimodal language models, where models are required to process both images and texts, a naive reward model may learn implicit biases toward texts and become less grounded in images. In this paper, we propose a **T**oken-**L**evel **D**etective **R**eward Model (**TLDR**) to provide fine-grained annotations to each text token. We first introduce a perturbation-based method to generate synthetic hard negatives and their token-level labels to train TLDR models. Then we show the rich usefulness of TLDR models both in assisting off-the-shelf models to self-correct their generations, and in serving as a hallucination evaluation tool. We show that TLDR automatically trains a token-level likelihood optimization, and can improve the base model's performance significantly. Finally, we show that TLDR models can significantly speed up human annotation by 3 times to acquire a broader range of high-quality vision language data. | vision language model, multimodal, reward model | null | 1,023 | 2410.04734 | [
-0.07696700841188431,
-0.06742695719003677,
0.047083254903554916,
-0.0019378961296752095,
0.11163695156574249,
0.0056794872507452965,
-0.013334310613572598,
0.024169106036424637,
0.0859108567237854,
-0.007093783933669329,
0.017808062955737114,
-0.1006973385810852,
0.09412267804145813,
0.096910759806633,
0.02365637570619583,
-0.004585741087794304,
0.05019032955169678,
0.09669847041368484,
-0.11061814427375793,
-0.061187271028757095,
0.004931099247187376,
0.03958364576101303,
0.07569956034421921,
-0.05676417797803879,
0.01092027872800827,
0.0050821807235479355,
0.0058163064531981945,
-0.05397850275039673,
-0.00452651921659708,
-0.06119142100214958,
-0.003958048764616251,
-0.023182794451713562,
0.024755582213401794,
0.058365873992443085,
-0.0544608049094677,
0.03457336127758026,
-0.09926649928092957,
-0.010719476267695427,
0.02214672602713108,
-0.05826742202043533,
-0.06965747475624084,
-0.03568694368004799,
-0.004480542615056038,
-0.001123500056564808,
0.11356266587972641,
-0.01082612480968237,
0.0065957242622971535,
-0.010785617865622044,
0.012224274687469006,
0.005694154649972916,
-0.04408605769276619,
-0.046869274228811264,
-0.006029755342751741,
-0.0035512906033545732,
-0.07124491780996323,
-0.016762053593993187,
0.08599312603473663,
-0.016824275255203247,
0.0345207117497921,
-0.020762039348483086,
-0.0015822791028767824,
-0.07880894094705582,
-0.041782427579164505,
-0.04142045974731445,
-0.007283774670213461,
-0.06372849643230438,
0.017855532467365265,
0.0008431095629930496,
-0.023209242150187492,
0.04159526154398918,
0.012167696841061115,
0.03654922917485237,
0.015840796753764153,
0.010316929779946804,
-0.026225892826914787,
0.05547407269477844,
0.020519325509667397,
-0.048056893050670624,
0.08283144980669022,
-0.11966439336538315,
-0.02211998775601387,
0.004464150872081518,
0.0811200886964798,
-0.0029334984719753265,
0.06969191133975983,
-0.019593918696045876,
-0.01244007796049118,
0.04496540129184723,
0.04283517971634865,
0.0292977225035429,
-0.047218866646289825,
-0.08106399327516556,
0.018275195732712746,
-0.018957819789648056,
0.020965253934264183,
0.023926913738250732,
-0.05124322324991226,
-0.049777597188949585,
-0.07045318931341171,
0.10016437619924545,
0.042115818709135056,
0.06956660747528076,
-0.023507313802838326,
-0.07221076637506485,
0.07253054529428482,
0.02035096101462841,
0.061589788645505905,
0.0009960051393136382,
0.09376607835292816,
-0.12858745455741882,
-0.004269367549568415,
0.006736979819834232,
0.011179502122104168,
0.03737344965338707,
0.05840165913105011,
-0.037250813096761703,
0.03831014409661293,
-0.0010396024445071816,
0.032212574034929276,
0.044392358511686325,
0.04844942316412926,
-0.044584937393665314,
-0.037291426211595535,
0.01153081376105547,
-0.006628414615988731,
-0.027482185512781143,
-0.05053039640188217,
6.998445207574649e-34,
0.03666180744767189,
0.09557436406612396,
0.003519721096381545,
0.032510917633771896,
0.03807467594742775,
-0.027015823870897293,
-0.05114982649683952,
0.05116468667984009,
-0.06763956695795059,
-0.03065929189324379,
-0.009126103483140469,
-0.009104667231440544,
-0.08344616740942001,
0.14261890947818756,
-0.005632042419165373,
-0.0025338628329336643,
-0.05182947963476181,
-0.02234778180718422,
0.0052567049860954285,
0.04275684803724289,
0.0011052978225052357,
-0.006078145932406187,
-0.0009308598819188774,
-0.021061690524220467,
0.01963183283805847,
0.014046335592865944,
0.1291245073080063,
-0.10382674634456635,
-0.06321913748979568,
0.008543227799236774,
-0.10277073830366135,
-0.04334385320544243,
0.08632136881351471,
0.008289379999041557,
0.05080723389983177,
-0.001808130880817771,
-0.0043734097853302956,
0.017764993011951447,
0.01451902836561203,
-0.03704778477549553,
-0.023730576038360596,
0.032518062740564346,
-0.049928728491067886,
-0.024528050795197487,
-0.0758395865559578,
0.009321996942162514,
0.018708448857069016,
0.031281281262636185,
-0.08306106925010681,
-0.018376318737864494,
0.07417785376310349,
0.008246318437159061,
-0.016831325367093086,
-0.040951840579509735,
-0.03864753618836403,
0.02251061424612999,
-0.02052886411547661,
0.017305390909314156,
0.024318944662809372,
-0.054049935191869736,
0.04446176439523697,
0.057504210621118546,
0.05026286467909813,
0.04195692762732506,
-0.05566296726465225,
0.019914602860808372,
0.0004767106147482991,
0.019566599279642105,
0.029624510556459427,
0.042215410619974136,
-0.008383922278881073,
0.05748717859387398,
-0.0417439341545105,
-0.07082944363355637,
0.007223252672702074,
0.028370937332510948,
0.09139657765626907,
-0.07623279094696045,
0.04761192575097084,
0.019713807851076126,
-0.020159615203738213,
-0.07508604973554611,
0.024817269295454025,
-0.0900060161948204,
-0.02964855171740055,
-0.05382481589913368,
0.025074519217014313,
-0.11078731715679169,
0.01869134232401848,
-0.007181690074503422,
-0.000536155654117465,
-0.03627126291394234,
-0.04495241120457649,
0.026060566306114197,
0.01846688613295555,
-2.3600882217813896e-34,
0.004852319601923227,
0.0033701974898576736,
-0.052917663007974625,
-0.011367525905370712,
-0.07431231439113617,
-0.07755948603153229,
0.005388043355196714,
0.045192647725343704,
0.02347257174551487,
-0.03438197821378708,
-0.059397414326667786,
0.00605795718729496,
0.010914294049143791,
-0.014041935093700886,
0.0503925122320652,
-0.058057550340890884,
0.0482330247759819,
0.05137896165251732,
0.003721985500305891,
0.06209970638155937,
-0.017676612362265587,
0.10996583849191666,
-0.16957581043243408,
0.06041458249092102,
-0.027092665433883667,
0.12977366149425507,
-0.04316508024930954,
0.016405915841460228,
-0.06973264366388321,
-0.0005641505122184753,
-0.03259410336613655,
0.008523158729076385,
0.023611919954419136,
-0.014384862035512924,
-0.0530267097055912,
0.04478434845805168,
0.061860036104917526,
-0.035373084247112274,
-0.05993904545903206,
0.090784452855587,
0.050484005361795425,
0.015149991028010845,
-0.06635312736034393,
0.0037815356627106667,
-0.0640251636505127,
-0.11223933845758438,
-0.05550002679228783,
-0.0022678973618894815,
0.043243151158094406,
0.042103711515665054,
-0.02692217007279396,
-0.016377460211515427,
-0.06232709437608719,
0.025322670117020607,
-0.10977593064308167,
-0.10174371302127838,
-0.04958216845989227,
-0.027027808129787445,
0.07297133654356003,
0.030378011986613274,
-0.08974876999855042,
-0.01471040304750204,
0.01074579544365406,
-0.016990412026643753,
-0.051578402519226074,
-0.07502918690443039,
-0.0178880225867033,
0.01040040422230959,
0.002888744231313467,
0.013868477195501328,
0.07220958918333054,
0.018382569774985313,
-0.002692030742764473,
0.04719346761703491,
0.02161075361073017,
0.011277380399405956,
0.03572764992713928,
-0.010783855803310871,
-0.004683636594563723,
-0.09854232519865036,
-0.005556254647672176,
-0.04241863638162613,
0.04160566255450249,
0.08692429214715958,
0.03610600158572197,
0.06811263412237167,
0.026188090443611145,
0.10920803993940353,
-0.00504684541374445,
0.007692201528698206,
-0.02087252028286457,
-0.002808487042784691,
0.05647910758852959,
0.09274592250585556,
-0.0009494525729678571,
-3.822131233732762e-8,
-0.05073818564414978,
-0.0002482641430106014,
-0.00994084496051073,
0.01084329467266798,
-0.0007957416237331927,
-0.007992339320480824,
-0.023128489032387733,
0.04230966046452522,
-0.007279235403984785,
-0.03468669205904007,
0.07849972695112228,
0.05234676972031593,
-0.05281839892268181,
-0.1308586746454239,
-0.056133054196834564,
0.09263034909963608,
0.029042638838291168,
0.06981644034385681,
0.009557279758155346,
-0.010751500725746155,
0.06597592681646347,
0.055283863097429276,
0.035380348563194275,
-0.03704154118895531,
-0.0036806424614042044,
-0.03968453407287598,
-0.007489528972655535,
0.047366783022880554,
0.052606191486120224,
-0.030840707942843437,
0.026828547939658165,
0.028119031339883804,
-0.03133181482553482,
-0.001642350573092699,
0.08727400004863739,
0.052405793219804764,
0.03200700134038925,
-0.03174152597784996,
-0.055784791707992554,
0.0030902784783393145,
0.04663073644042015,
0.057849276810884476,
-0.038735199719667435,
-0.03008344955742359,
0.01917589083313942,
0.015323208644986153,
0.01628091372549534,
-0.1370626538991928,
0.053536493331193924,
-0.03185418248176575,
0.009208313189446926,
-0.026572009548544884,
-0.03908361494541168,
0.13059096038341522,
-0.00846191681921482,
0.010666154325008392,
0.043328121304512024,
0.05496322363615036,
-0.009626796469092369,
0.12655089795589447,
0.059994518756866455,
0.015370985493063927,
-0.0023794230073690414,
-0.040793467313051224
] |
Statistical Advantages of Perturbing Cosine Router in Mixture of Experts | https://openreview.net/forum?id=faDMOmnsjx | [
"Huy Nguyen",
"Pedram Akbarian",
"Huyen Trang Pham",
"Thien Trang Nguyen Vu",
"Shujian Zhang",
"Nhat Ho"
] | Poster | The cosine router in Mixture of Experts (MoE) has recently emerged as an attractive alternative to the conventional linear router. Indeed, the cosine router demonstrates favorable performance in image and language tasks and exhibits better ability to mitigate the representation collapse issue, which often leads to parameter redundancy and limited representation potentials. Despite its empirical success, a comprehensive analysis of the cosine router in MoE has been lacking. Considering the least square estimation of the cosine routing MoE, we demonstrate that due to the intrinsic interaction of the model parameters in the cosine router via some partial differential equations, regardless of the structures of the experts, the estimation rates of experts and model parameters can be as slow as $\mathcal{O}(1/\log^{\tau}(n))$ where $\tau > 0$ is some constant and $n$ is the sample size. Surprisingly, these pessimistic non-polynomial convergence rates can be circumvented by the widely used technique in practice to stabilize the cosine router --- simply adding noises to the $\ell^2$-norms in the cosine router, which we refer to as *perturbed cosine router*. Under the strongly identifiable settings of the expert functions, we prove that the estimation rates for both the experts and model parameters under the perturbed cosine routing MoE are significantly improved to polynomial rates. Finally, we conduct extensive simulation studies in both synthetic and real data settings to empirically validate our theoretical results. | mixture of experts, cosine router, perturbation | null | 1,016 | 2405.14131 | [
-0.023494696244597435,
-0.07781911641359329,
0.0067355213686823845,
0.050528571009635925,
0.00390861788764596,
-0.04810498282313347,
-0.07310588657855988,
0.02608010359108448,
0.03834531083703041,
-0.058867719024419785,
-0.00890043843537569,
0.027458220720291138,
0.00041553156916052103,
-0.00891140941530466,
-0.04861176386475563,
-0.01805756241083145,
0.06232037395238876,
0.033205028623342514,
-0.05332519859075546,
-0.08842411637306213,
-0.01972978562116623,
-0.02752475067973137,
-0.01548804435878992,
-0.031060855835676193,
0.046963680535554886,
-0.00724441185593605,
0.05130748078227043,
-0.003125778865069151,
0.020393775776028633,
-0.06473193317651749,
0.031278882175683975,
-0.019380804151296616,
-0.04777979105710983,
0.008828122168779373,
-0.08410832285881042,
0.00866013765335083,
0.06537029892206192,
-0.016005778685212135,
0.08089426159858704,
0.034586843103170395,
-0.029766814783215523,
0.11271493881940842,
0.026469629257917404,
0.036792952567338943,
0.06750883162021637,
-0.06900013238191605,
0.050350621342659,
-0.03338588774204254,
-0.011986671015620232,
-0.06311921775341034,
-0.07535067200660706,
-0.024404210969805717,
0.004615046549588442,
-0.06189608573913574,
-0.0365876778960228,
0.011161839589476585,
-0.009396507404744625,
0.05109235644340515,
0.03178724646568298,
0.04763513803482056,
-0.0396110936999321,
-0.08781701326370239,
-0.034247878938913345,
0.035134125500917435,
-0.014070656150579453,
0.005562628153711557,
0.14749839901924133,
-0.028777798637747765,
-0.0014240776654332876,
0.06730164587497711,
-0.0974959284067154,
0.0815415233373642,
-0.08652745187282562,
0.04314332827925682,
0.044408950954675674,
-0.023495279252529144,
0.019564298912882805,
-0.006095572374761105,
0.03661184385418892,
-0.015361984260380268,
0.06007459759712219,
0.021601729094982147,
-0.027410732582211494,
-0.07639914005994797,
0.11693760007619858,
-0.032483696937561035,
-0.08114749193191528,
0.011025137268006802,
0.1041274219751358,
-0.05399487912654877,
-0.13112890720367432,
-0.06292037665843964,
-0.04332228749990463,
0.008105764165520668,
0.007617758587002754,
0.04585043713450432,
0.06615446507930756,
0.0009523683693259954,
-0.02872495725750923,
0.08838193863630295,
-0.037123095244169235,
-0.06701341271400452,
0.038705550134181976,
0.03125040978193283,
-0.018396133556962013,
0.000023444363250746392,
0.1021953821182251,
0.06273433566093445,
0.07123176753520966,
0.048683930188417435,
-0.05092011019587517,
0.04079676419496536,
-0.005599981639534235,
0.011950459331274033,
-0.018068501725792885,
-0.04527125880122185,
0.05058121681213379,
-0.024950355291366577,
0.06947123259305954,
0.034619662910699844,
-0.007281031459569931,
-0.03271344304084778,
0.030540524050593376,
0.04342007264494896,
0.030278338119387627,
0.004199635703116655,
0.015584570355713367,
3.333333167579437e-33,
-0.004700840916484594,
0.17886382341384888,
0.017648523673415184,
-0.01968778856098652,
-0.017688823863863945,
0.0066155316308140755,
0.0059430645778775215,
-0.01876154914498329,
0.052876830101013184,
-0.006673197727650404,
-0.049807414412498474,
0.04497933015227318,
-0.07410365343093872,
-0.0017889884766191244,
0.0578201524913311,
0.002854566788300872,
-0.037479303777217865,
0.050258830189704895,
-0.04236593842506409,
-0.010381064377725124,
0.08887898921966553,
-0.04719411954283714,
-0.024674925953149796,
0.012031597085297108,
-0.020709795877337456,
-0.019137661904096603,
0.07259519398212433,
-0.041441090404987335,
-0.03164845332503319,
-0.0068296706303954124,
-0.015809200704097748,
0.012248455546796322,
0.02294704131782055,
0.0006806404562667012,
0.07253225892782211,
-0.04130437970161438,
-0.01809961162507534,
-0.038298096507787704,
-0.003696083789691329,
0.011196008883416653,
-0.025780875235795975,
0.04043455421924591,
0.018959540873765945,
0.007720576133579016,
-0.09890077263116837,
-0.09581483155488968,
0.019152462482452393,
0.04784945398569107,
-0.10819287598133087,
-0.03039165586233139,
-0.05781424045562744,
0.02951556071639061,
-0.08763294667005539,
0.08538033068180084,
0.07420150190591812,
-0.0455983430147171,
0.006320583634078503,
-0.04213929921388626,
0.024101847782731056,
0.035734884440898895,
0.0017235043924301863,
0.024403974413871765,
0.05224334076046944,
0.002317097969353199,
0.057693686336278915,
0.04066959023475647,
-0.06899845600128174,
0.05950724333524704,
0.07553093880414963,
-0.002489957958459854,
-0.007922641932964325,
-0.0018916751723736525,
-0.03254833072423935,
-0.021460069343447685,
0.0592743419110775,
-0.04535982757806778,
-0.03679881989955902,
-0.051468029618263245,
-0.0026145405136048794,
0.04392200708389282,
-0.051691874861717224,
-0.02278372459113598,
0.026462405920028687,
-0.03276713192462921,
-0.033965080976486206,
-0.014455277472734451,
0.029875323176383972,
-0.07046101242303848,
-0.019315162673592567,
-0.06461989134550095,
-0.08907151967287064,
0.04391639307141304,
0.00020340393530204892,
0.014631315134465694,
-0.013100461103022099,
-3.225132218664107e-33,
-0.06401732563972473,
0.048307325690984726,
0.05455983802676201,
0.07524508982896805,
-0.13213013112545013,
-0.042601604014635086,
0.06955686211585999,
-0.006884003058075905,
0.07343059033155441,
-0.0928940698504448,
-0.02051113173365593,
-0.029081793501973152,
0.05084872618317604,
-0.010233743116259575,
0.11433181166648865,
-0.026894787326455116,
-0.05681968852877617,
0.03745326027274132,
0.10158443450927734,
0.007559305522590876,
0.035206522792577744,
0.007902976125478745,
-0.08871372044086456,
-0.030560005456209183,
-0.017763713374733925,
0.03421337530016899,
-0.03818172961473465,
0.003856165334582329,
-0.016081321984529495,
0.025577759370207787,
-0.024909308180212975,
-0.003930701874196529,
0.0025454757269471884,
-0.035532396286726,
-0.0453057736158371,
0.15440069139003754,
0.0015119134914129972,
0.07150086760520935,
0.026887238025665283,
0.006953337695449591,
0.08288043737411499,
0.02930445224046707,
-0.00913169514387846,
-0.01529773324728012,
0.019274383783340454,
-0.0880843847990036,
0.012172437272965908,
-0.08712451905012131,
-0.03392884135246277,
0.03549271449446678,
-0.011735731735825539,
0.0489024892449379,
-0.0954563170671463,
0.017607921734452248,
-0.06099662929773331,
0.016726983711123466,
-0.06313768029212952,
-0.028625722974538803,
0.08946727216243744,
-0.02053176797926426,
-0.008293741382658482,
0.005590187385678291,
-0.030158426612615585,
0.08107288181781769,
0.031104089692234993,
-0.010654653422534466,
-0.029127435758709908,
-0.01587962545454502,
0.05828689783811569,
0.04775027185678482,
-0.05476033687591553,
0.015930835157632828,
0.06062539666891098,
0.08265942335128784,
0.02297073043882847,
-0.04046871140599251,
0.01379375346004963,
-0.045521706342697144,
-0.07548894733190536,
-0.03434564918279648,
-0.04947223141789436,
0.029850255697965622,
0.023522133007645607,
-0.09621555358171463,
0.027986809611320496,
0.05692321062088013,
-0.03372008726000786,
0.020456504076719284,
-0.00915675237774849,
-0.028537049889564514,
0.027662625536322594,
0.01617586240172386,
0.05820995941758156,
0.00483661238104105,
-0.04197292774915695,
-5.4413387573504224e-8,
-0.071635402739048,
0.06185417249798775,
0.01867128163576126,
-0.04192004352807999,
0.14266136288642883,
0.026692725718021393,
-0.035258617252111435,
-0.07036451250314713,
-0.03856878727674484,
-0.03077911026775837,
0.09537749737501144,
-0.011050569824874401,
-0.10123129189014435,
0.012601442635059357,
-0.003497574245557189,
0.07993344962596893,
-0.018969301134347916,
0.016558023169636726,
-0.0069564953446388245,
-0.01803007163107395,
-0.009050180204212666,
0.01745324768126011,
-0.009906765073537827,
-0.009938366711139679,
0.059188392013311386,
-0.010345318354666233,
-0.04953940957784653,
0.109217070043087,
-0.011424015276134014,
0.030955610796809196,
-0.08357394486665726,
0.04337342828512192,
0.024491354823112488,
0.037136100232601166,
-0.0008318801410496235,
0.09944716095924377,
0.035619232803583145,
0.016495762392878532,
-0.06320444494485855,
0.012346659786999226,
-0.030335035175085068,
0.0957568883895874,
-0.024586381390690804,
0.04246895760297775,
0.1606721132993698,
0.02159734070301056,
0.03648664802312851,
-0.09067323058843613,
0.015735983848571777,
0.017998619005084038,
-0.012838982976973057,
0.014587956480681896,
-0.07476803660392761,
0.04725094884634018,
0.09882353246212006,
-0.030289411544799805,
-0.0037836283445358276,
-0.10438404232263565,
-0.028580334037542343,
0.037261318415403366,
-0.014456916600465775,
0.0056000761687755585,
-0.048832569271326065,
-0.019331946969032288
] |
Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios | https://openreview.net/forum?id=PVHoELf5UN | [
"Li Huaqiu",
"HuXiaowan",
"Haoqian Wang"
] | Poster | Real-world low-light images often suffer from complex degradations such as local overexposure, low brightness, noise, and uneven illumination. Supervised methods tend to overfit to specific scenarios, while unsupervised methods, though better at generalization, struggle to model these degradations due to the lack of reference images. To address this issue, we propose an interpretable, zero-reference joint denoising and low-light enhancement framework tailored for real-world scenarios. Our method derives a training strategy based on paired sub-images with varying illumination and noise levels, grounded in physical imaging principles and retinex theory. Additionally, we leverage the Discrete Cosine Transform (DCT) to perform frequency domain decomposition in the sRGB space, and introduce an implicit-guided hybrid representation strategy that effectively separates intricate compounded degradations. In the backbone network design, we develop retinal decomposition network guided by implicit degradation representation mechanisms. Extensive experiments demonstrate the superiority of our method. Code will be available at https://github.com/huaqlili/unsupervised-light-enhance-ICLR2025. | unsupervised learning, low-light image enhancement, image denoising | An interpretable joint method for image denoising and enhancement | 1,007 | 2503.14535 | [
-0.013852724805474281,
-0.04315556585788727,
0.09338304400444031,
0.06106260046362877,
0.08502227067947388,
-0.061296265572309494,
-0.04332762584090233,
-0.05993847921490669,
0.008790919557213783,
-0.051827047020196915,
-0.04389535263180733,
0.014100305736064911,
0.010913929902017117,
0.039440009742975235,
-0.054563555866479874,
0.01670907810330391,
0.04798325523734093,
0.10280583053827286,
-0.08878602832555771,
-0.04852275922894478,
0.07463128864765167,
-0.08785116672515869,
-0.059503260999917984,
-0.03139382228255272,
0.04340602084994316,
-0.0095622343942523,
0.037767525762319565,
-0.049275193363428116,
0.01490459218621254,
-0.017128968611359596,
0.02754916436970234,
0.024205032736063004,
-0.0636681541800499,
-0.03211243823170662,
-0.025329269468784332,
0.01581159420311451,
0.1087016761302948,
-0.012215045280754566,
0.0023692594841122627,
0.017195189371705055,
0.029336415231227875,
0.10249678045511246,
-0.011777638457715511,
0.000807746488135308,
0.08608096837997437,
-0.007395756896585226,
0.0028768666088581085,
-0.09204144775867462,
-0.012555785477161407,
-0.007409247569739819,
0.09047577530145645,
0.05867500603199005,
-0.05985427647829056,
0.08876857161521912,
0.032828234136104584,
0.0664629340171814,
-0.018917053937911987,
-0.033649008721113205,
0.028458483517169952,
0.0019616130739450455,
0.035539258271455765,
-0.021848689764738083,
0.04078708216547966,
-0.019256925210356712,
0.015730660408735275,
0.02407325990498066,
0.03280378133058548,
-0.009797167032957077,
0.0026236625853925943,
0.04116441309452057,
-0.06750229746103287,
0.01950901746749878,
-0.01746680773794651,
-0.02022353932261467,
-0.02020801417529583,
0.09930290281772614,
0.034462820738554,
-0.020555410534143448,
0.02070196159183979,
-0.05497973784804344,
0.07827373594045639,
0.03611038252711296,
0.05200550705194473,
-0.022710781544446945,
0.1330358386039734,
0.031248657032847404,
-0.0356755368411541,
0.04577067866921425,
-0.021316302940249443,
0.023142503574490547,
-0.1285235732793808,
0.01470633503049612,
-0.05881894379854202,
-0.002290330594405532,
-0.03656579926609993,
0.00004541884845821187,
0.059510014951229095,
-0.023959815502166748,
-0.0765022411942482,
0.06565291434526443,
-0.021892601624131203,
-0.009283858351409435,
-0.002950250869616866,
0.009206774644553661,
0.020242637023329735,
-0.02878209762275219,
0.06696229428052902,
0.13283413648605347,
0.008007088676095009,
0.0014581183204427361,
0.02304454706609249,
0.009686100296676159,
-0.04614797979593277,
-0.029987242072820663,
0.02899579331278801,
0.02579338476061821,
0.06197114288806915,
0.024484340101480484,
0.04383813217282295,
-0.07958956807851791,
-0.023797523230314255,
-0.032545797526836395,
-0.07117193937301636,
-0.024800533428788185,
0.09153371304273605,
-0.033970076590776443,
-0.027968039736151695,
1.4135528953569248e-33,
0.002824387513101101,
0.04436257481575012,
0.008357922546565533,
-0.006423473358154297,
0.10391249507665634,
-0.08927745372056961,
-0.0436415895819664,
-0.013549461960792542,
-0.0348818376660347,
0.046281278133392334,
0.05582747980952263,
0.0225547943264246,
-0.04175352305173874,
0.04334048926830292,
0.04996742680668831,
-0.014889072626829147,
0.010628307238221169,
0.012043928727507591,
0.011302072554826736,
-0.06439836323261261,
-0.03800934553146362,
0.049190402030944824,
-0.015665067359805107,
0.005964204203337431,
0.005527134984731674,
-0.014777879230678082,
0.024838505312800407,
0.011317027732729912,
-0.00010220434342045337,
-0.020050300285220146,
-0.09811782836914062,
0.0717344731092453,
0.03322233259677887,
0.002537079155445099,
0.0030653050635010004,
-0.03968071565032005,
0.04359225183725357,
-0.012773089110851288,
-0.018504247069358826,
0.02786741405725479,
0.0009489418007433414,
0.09178665280342102,
0.011672144755721092,
-0.10237540304660797,
-0.0417877659201622,
0.004287464078515768,
0.023145167157053947,
0.1189059242606163,
-0.07120927423238754,
-0.0022920281626284122,
0.001994891557842493,
-0.02845887467265129,
-0.041154101490974426,
-0.03229358419775963,
-0.11037946492433548,
-0.009342857636511326,
0.03551582992076874,
-0.03304426744580269,
0.05190197750926018,
0.011585051193833351,
0.04908009618520737,
-0.05279308557510376,
-0.009997601620852947,
0.06704515963792801,
0.03721506521105766,
-0.009846743196249008,
0.007093043066561222,
0.013454996980726719,
-0.021645335480570793,
-0.06143159791827202,
-0.05045408383011818,
0.08763886988162994,
0.024614039808511734,
-0.03819216787815094,
0.06708277016878128,
0.03974839299917221,
0.05579555407166481,
0.011629633605480194,
0.058074288070201874,
0.05352175608277321,
-0.059944670647382736,
0.034462060779333115,
-0.019213389605283737,
-0.05524567887187004,
-0.062461577355861664,
0.05398626998066902,
-0.025854580104351044,
0.003572339192032814,
-0.1156553402543068,
-0.12290018796920776,
-0.054426804184913635,
-0.027099743485450745,
0.0194998886436224,
-0.0565602108836174,
0.0005189235671423376,
3.7701116035144742e-34,
0.024277284741401672,
0.09882479161024094,
0.0049686916172504425,
-0.03185120224952698,
-0.013873114250600338,
-0.032986488193273544,
0.031859539449214935,
0.01793479733169079,
0.06151403486728668,
-0.06264525651931763,
0.049367114901542664,
-0.04823426529765129,
-0.04649221897125244,
-0.009811317548155785,
-0.011121373623609543,
-0.026676559820771217,
-0.10041312873363495,
-0.061494454741477966,
-0.07507453113794327,
0.004287160933017731,
0.02033638022840023,
0.0888332724571228,
-0.022860104218125343,
0.047687653452157974,
-0.039731450378894806,
0.1560593843460083,
0.0060858032666146755,
0.07019061595201492,
0.0012187117245048285,
-0.04484475404024124,
-0.03371378779411316,
0.019158562645316124,
-0.012481885962188244,
-0.04090467095375061,
-0.02659854292869568,
0.08607994019985199,
-0.01772439293563366,
-0.022679708898067474,
-0.017975151538848877,
0.015082101337611675,
0.06095559149980545,
0.02055087685585022,
-0.09729407727718353,
0.004898007959127426,
0.04126930981874466,
0.0009839697740972042,
-0.03531552851200104,
-0.014202818274497986,
0.041301023215055466,
0.0300325695425272,
-0.05352332442998886,
0.012836973182857037,
-0.12727755308151245,
0.042949777096509933,
-0.05406263470649719,
-0.09572959691286087,
-0.008243154734373093,
0.05762164294719696,
0.14626818895339966,
0.0042063831351697445,
0.006103656720370054,
0.004042003769427538,
-0.06892688572406769,
0.04091458022594452,
-0.06212550401687622,
0.023556891828775406,
0.03494547680020332,
0.03527940809726715,
0.014660440385341644,
0.023184897378087044,
-0.010991320945322514,
0.004939578473567963,
0.015588866546750069,
-0.023945126682519913,
0.040946390479803085,
-0.10878793895244598,
-0.010128196328878403,
0.007967907004058361,
-0.02560792677104473,
0.0007818007143214345,
0.018455535173416138,
-0.06498955935239792,
-0.02577834203839302,
0.09177331626415253,
0.10546256601810455,
0.04235037416219711,
-0.07636506855487823,
-0.015375453047454357,
0.04503298178315163,
-0.026686683297157288,
-0.07503832131624222,
-0.012552794069051743,
-0.017522089183330536,
0.02064887061715126,
-0.027343565598130226,
-4.3443741049031814e-8,
0.003557520220056176,
0.03429262712597847,
-0.03658134117722511,
-0.05080539733171463,
0.004969020374119282,
-0.11300812661647797,
-0.024241287261247635,
0.06205227971076965,
0.03562072664499283,
-0.012660100124776363,
0.06696146726608276,
-0.1201721578836441,
-0.06723704934120178,
-0.02631896175444126,
-0.0863824412226677,
0.0907260850071907,
0.027186650782823563,
0.024232646450400352,
0.01262222696095705,
-0.00063440011581406,
-0.13572336733341217,
-0.02312968485057354,
0.018253011628985405,
0.0310819149017334,
0.10799332708120346,
0.003400996560230851,
0.010256443172693253,
0.04787937551736832,
0.0923016294836998,
0.021547192707657814,
0.06443478912115097,
0.031077083200216293,
0.11907704174518585,
0.02324051596224308,
-0.03613751009106636,
0.016045507043600082,
0.059126656502485275,
0.08470945805311203,
-0.08273860067129135,
0.025529460981488228,
-0.13504858314990997,
0.020915640518069267,
-0.055464331060647964,
0.018763411790132523,
0.022951550781726837,
-0.04596995934844017,
0.052331872284412384,
-0.003265865845605731,
-0.02193196304142475,
0.06355442106723785,
0.020387180149555206,
0.016962556168437004,
0.022990550845861435,
-0.025287818163633347,
-0.01422903686761856,
-0.062301889061927795,
0.05244816094636917,
0.05847739055752754,
0.008915954269468784,
0.040005385875701904,
0.022341424599289894,
-0.046130113303661346,
-0.017701556906104088,
0.024698888882994652
] |
Learning Clustering-based Prototypes for Compositional Zero-Shot Learning | https://openreview.net/forum?id=eE2PXlNydB | [
"Hongyu Qu",
"Jianan Wei",
"Xiangbo Shu",
"Wenguan Wang"
] | Poster | Learning primitive (i.e., attribute and object) concepts from seen compositions is the primary challenge of Compositional Zero-Shot Learning (CZSL). Existing CZSL solutions typically rely on oversimplified data assumptions, e.g., modeling each primitive with a single centroid primitive presentation, ignoring the natural diversities of the attribute (resp. object) when coupled with different objects (resp. attribute). In this work, we develop ClusPro, a robust clustering-based prototype mining framework for CZSL that defines the conceptual boundaries of primitives through a set of diversified prototypes. Specifically, ClusPro conducts within-primitive clustering on the embedding space for automatically discovering and dynamically updating prototypes. To learn high-quality embeddings for discriminative prototype construction, ClusPro repaints a well-structured and independent primitive embedding space, ensuring intra-primitive separation and inter-primitive decorrelation through prototype-based contrastive learning and decorrelation learning. Moreover, ClusPro effectively performs prototype clustering in a non-parametric fashion without the introduction of additional learnable parameters or computational budget during testing. Experiments on three benchmarks demonstrate ClusPro outperforms various top-leading CZSL solutions under both closed-world and open-world settings. Our code is available at CLUSPRO. | Compositional Zero-Shot Learning, Prototype Learning, Representation Disentanglement | We propose a clustering-based prototype mining framework for compositional zero-shot learning, which defines conceptual boundaries of primitives through a set of diversified prototypes, and automatically discovers these prototypes via clustering. | 1,003 | 2502.06501 | [
0.020580964162945747,
-0.04873469099402428,
0.0293880645185709,
0.015627602115273476,
0.05456643924117088,
0.012366714887320995,
-0.036840617656707764,
-0.06442723423242569,
0.018380528315901756,
-0.02168998494744301,
-0.0357561856508255,
-0.08005788177251816,
0.07640424370765686,
0.05832524225115776,
-0.038979414850473404,
0.025760173797607422,
0.0905708596110344,
0.057700611650943756,
-0.07746570557355881,
-0.02279667556285858,
-0.0702955350279808,
-0.03850020095705986,
0.08214617520570755,
0.012978020124137402,
0.009779469110071659,
0.05325261130928993,
0.03381495550274849,
-0.014174536801874638,
0.10308513045310974,
-0.02034701779484749,
0.01923508383333683,
0.05125570669770241,
-0.06041186675429344,
0.026961276307702065,
-0.00386737659573555,
0.09438653290271759,
0.06634252518415451,
0.0978388637304306,
-0.08069664984941483,
-0.008538912050426006,
-0.003463377244770527,
0.08957932889461517,
-0.020739130675792694,
0.020895345136523247,
0.03978202864527702,
0.10058906674385071,
-0.047443874180316925,
-0.06757448613643646,
-0.04066505655646324,
0.0348576195538044,
-0.05029337480664253,
0.002223142422735691,
-0.10746879875659943,
0.030616121366620064,
0.02688652090728283,
0.043130163103342056,
0.01226876676082611,
-0.0824677124619484,
0.07309439778327942,
-0.05820925161242485,
-0.029852768406271935,
-0.12523142993450165,
0.08275040984153748,
-0.0009706109995022416,
-0.04024062678217888,
-0.05725962668657303,
-0.02707224152982235,
0.06745856255292892,
0.005551192909479141,
-0.02910405769944191,
0.03765540197491646,
0.1065916195511818,
-0.02754446305334568,
0.08661459386348724,
-0.002554698847234249,
0.10538914799690247,
0.05775696411728859,
-0.07981748878955841,
-0.006525622680783272,
-0.11398161947727203,
-0.010391447693109512,
0.020766567438840866,
-0.02761591225862503,
-0.0028762631118297577,
0.10717548429965973,
0.0880357101559639,
0.025031542405486107,
0.0523843914270401,
-0.01879848726093769,
0.011609809473156929,
-0.07951996475458145,
0.019210198894143105,
0.009798985905945301,
-0.04172274470329285,
0.020418239757418633,
-0.010606538504362106,
0.04816339164972305,
-0.06273221224546432,
-0.0005858418298885226,
0.08936560899019241,
-0.03865998610854149,
0.055017534643411636,
0.07542120665311813,
-0.1078382059931755,
-0.013360017910599709,
0.030432607978582382,
-0.053291164338588715,
-0.029405314475297928,
0.016620151698589325,
-0.07700175791978836,
-0.0020708355586975813,
0.04106345772743225,
-0.054979171603918076,
-0.01654367707669735,
-0.0319397933781147,
-0.006130007095634937,
0.027311114594340324,
-0.07258151471614838,
0.054167576134204865,
0.0008847079589031637,
0.0830661728978157,
0.012973878532648087,
-0.036176133900880814,
0.060025401413440704,
0.026355966925621033,
-0.06768063455820084,
-0.06671089679002762,
1.4800045552816183e-34,
0.09277529269456863,
-0.0002996128750964999,
0.06370376795530319,
0.06793740391731262,
0.01825774647295475,
-0.024338046088814735,
-0.0038618124090135098,
0.11190059781074524,
-0.051742617040872574,
0.005256629083305597,
-0.0055580441839993,
0.038258470594882965,
-0.038351260125637054,
0.07883669435977936,
0.07015380263328552,
-0.02402643859386444,
-0.01348133571445942,
-0.0177096426486969,
-0.04593265801668167,
-0.07125895470380783,
-0.04674926772713661,
-0.013582990504801273,
-0.020009974017739296,
0.010575050488114357,
-0.04189985245466232,
0.0299361664801836,
0.06526326388120651,
-0.06399505585432053,
-0.049328576773405075,
0.006807388737797737,
-0.07344191521406174,
0.034776944667100906,
-0.0010260158451274037,
0.04009993001818657,
0.04909394308924675,
-0.07738999277353287,
-0.008512300439178944,
-0.06457265466451645,
0.06256592273712158,
-0.03256463259458542,
0.012928733602166176,
-0.027439072728157043,
-0.056931667029857635,
-0.056148696690797806,
-0.02995002642273903,
-0.05841910094022751,
-0.008001521229743958,
0.01015450805425644,
-0.042680639773607254,
-0.062179964035749435,
0.052838150411844254,
0.02534773014485836,
0.008874724619090557,
-0.06525047868490219,
-0.08908827602863312,
0.036472681909799576,
-0.028803512454032898,
-0.057066477835178375,
0.02449173666536808,
-0.041957657784223557,
0.0015013619558885694,
0.1165902316570282,
-0.0841975286602974,
0.0438084676861763,
-0.061284083873033524,
0.07553013414144516,
0.03114888072013855,
0.061706457287073135,
0.0948929414153099,
0.0202421136200428,
0.022301003336906433,
0.07856310904026031,
-0.06610569357872009,
-0.060203127562999725,
0.058817822486162186,
0.012442019768059254,
0.0360896997153759,
-0.04461418837308884,
0.00891169998794794,
0.05631615221500397,
-0.068470798432827,
-0.010028711520135403,
-0.01550941914319992,
0.0319041945040226,
-0.05641624704003334,
-0.08072762936353683,
0.061083532869815826,
-0.02620868757367134,
-0.05466095358133316,
-0.03213571012020111,
-0.008788587525486946,
-0.03873879835009575,
-0.10424462705850601,
-0.034135978668928146,
0.022644013166427612,
-1.7921245244590888e-33,
-0.011908960528671741,
0.04179908335208893,
-0.03688138723373413,
0.008217555470764637,
0.029827216640114784,
-0.03554109111428261,
-0.05892549827694893,
-0.0053466130048036575,
-0.017857424914836884,
-0.04515717551112175,
-0.0019075721502304077,
-0.04373162239789963,
-0.07036145031452179,
-0.07849019020795822,
0.047520801424980164,
0.035173218697309494,
-0.09303873777389526,
0.056430768221616745,
0.015936829149723053,
0.05760557949542999,
-0.013505086302757263,
0.03617691248655319,
-0.05437682196497917,
0.028461866080760956,
-0.007064028177410364,
0.07288369536399841,
-0.02635146677494049,
-0.008728264831006527,
-0.005933547858148813,
0.08959084004163742,
0.006200755946338177,
-0.027176814153790474,
-0.005234668031334877,
-0.028394905850291252,
-0.0786503329873085,
-0.05748444423079491,
0.0035146570298820734,
-0.044172946363687515,
0.03515835106372833,
-0.029002558439970016,
-0.018576763570308685,
-0.006390796974301338,
-0.12301483750343323,
0.04269376024603844,
0.007587973494082689,
-0.09097672253847122,
0.01471275556832552,
0.01165871974080801,
0.0483354851603508,
0.039300546050071716,
-0.08576487004756927,
-0.005795488599687815,
-0.055645592510700226,
-0.05100170895457268,
-0.017598096281290054,
-0.0110288942232728,
0.026365874335169792,
-0.02183537557721138,
0.07972601056098938,
0.033266134560108185,
0.034912481904029846,
-0.05878552794456482,
0.02234688214957714,
0.008302121423184872,
0.017360569909214973,
-0.045779138803482056,
-0.027457524091005325,
-0.007914319634437561,
-0.05455056577920914,
-0.007532950025051832,
0.02733040787279606,
0.07027554512023926,
0.004918825346976519,
-0.0764034315943718,
-0.018925949931144714,
-0.011523560620844364,
0.040014345198869705,
0.04590599238872528,
0.009701190516352654,
-0.018469564616680145,
0.032557688653469086,
-0.04371361434459686,
0.026151955127716064,
0.11445624381303787,
0.09940639138221741,
0.07074619084596634,
-0.023967834189534187,
0.07138574868440628,
0.03821428865194321,
0.0005788233829662204,
-0.053278714418411255,
0.04268473759293556,
0.003961277659982443,
0.07039550691843033,
-0.041685912758111954,
-3.8778736666245095e-8,
-0.11403980106115341,
-0.0029321983456611633,
-0.01634632796049118,
-0.02971571311354637,
0.006488737650215626,
-0.002944463863968849,
-0.06530652195215225,
0.0690917894244194,
-0.06336475163698196,
-0.022255782037973404,
0.05032387375831604,
0.003976230509579182,
-0.0026165475137531757,
-0.09370533376932144,
-0.0185919851064682,
0.08216825127601624,
-0.0014215090777724981,
-0.009170702658593655,
-0.01816440559923649,
0.07756301760673523,
0.04047539085149765,
0.019716812297701836,
-0.009856445714831352,
0.06670907139778137,
-0.024873489513993263,
-0.14564548432826996,
-0.05272465944290161,
0.016383418813347816,
0.11095014214515686,
-0.015151730738580227,
-0.014715075492858887,
0.0014540002448484302,
0.05688737705349922,
0.01763208769261837,
0.056464843451976776,
0.07797501981258392,
-0.004987480118870735,
-0.009985414333641529,
-0.08752590417861938,
-0.0481748953461647,
0.03679338097572327,
0.06260165572166443,
-0.0034250321332365274,
-0.014681575819849968,
0.035089097917079926,
-0.002230505459010601,
0.054768599569797516,
-0.04942934215068817,
0.008989077061414719,
0.04541061818599701,
-0.04231312498450279,
0.029123730957508087,
-0.07600173354148865,
0.060157570987939835,
-0.0011404183460399508,
0.0361342616379261,
-0.036928270012140274,
-0.03817407414317131,
0.06472127139568329,
-0.026140930131077766,
0.019930843263864517,
0.01949312351644039,
-0.015009435825049877,
0.03930312395095825
] |
SAGEPhos: Sage Bio-Coupled and Augmented Fusion for Phosphorylation Site Detection | https://openreview.net/forum?id=hLwcNSFhC2 | [
"Jingjie Zhang",
"Hanqun CAO",
"Zijun Gao",
"Xiaorui Wang",
"Chunbin Gu"
] | Poster | Phosphorylation site prediction based on kinase-substrate interaction plays a vital role in understanding cellular signaling pathways and disease mechanisms. Computational methods for this task can be categorized into kinase-family-focused and individual kinase-targeted approaches. Individual kinase-targeted methods have gained prominence for their ability to explore a broader protein space and provide more precise target information for kinase inhibitors. However, most existing individual kinase-based approaches focus solely on sequence inputs, neglecting crucial structural information. To address this limitation, we introduce SAGEPhos (Structure-aware kinAse-substrate bio-coupled and bio-auGmented nEtwork for Phosphorylation site prediction), a novel framework that modifies the semantic space of main protein inputs using auxiliary inputs at two distinct modality levels. At the inter-modality level, SAGEPhos introduces a Bio-Coupled Modal Fusion method, distilling essential kinase sequence information to refine task-oriented local substrate feature space, creating a shared semantic space that captures crucial kinase-substrate interaction patterns. Within the substrate's intra-modality domain, it focuses on Bio-Augmented Fusion, emphasizing 2D local sequence information while selectively incorporating 3D spatial information from predicted structures to complement the sequence space. Moreover, to address the lack of structural information in current datasets, we contribute a new, refined phosphorylation site prediction dataset, which incorporates crucial structural elements and will serve as a new benchmark for the field. Experimental results demonstrate that SAGEPhos significantly outperforms baseline methods, notably achieving almost 10\% and 12\% improvements in prediction accuracy and AUC-ROC, respectively. We further demonstrate our algorithm's robustness and generalization through stable results across varied data partitions and significant improvements in zero-shot scenarios. These results underscore the effectiveness of constructing a larger and more precise protein space in advancing the state-of-the-art in phosphorylation site prediction. We release the SAGEPhos models and code at https://github.com/ZhangJJ26/SAGEPhos. | Deep Learning, Bioinformatics, Phosphorylation prediction | null | 1,000 | 2502.07384 | [
-0.05492904409766197,
-0.08947340399026871,
-0.02646934986114502,
-0.001461067353375256,
0.09477647393941879,
-0.058569036424160004,
0.004349383059889078,
-0.0011981864226981997,
0.0343199223279953,
-0.01704755425453186,
-0.027236247435212135,
-0.023388059809803963,
0.032184701412916183,
0.06711207330226898,
0.028020493686199188,
0.06834279745817184,
-0.049293626099824905,
0.01966673508286476,
-0.013739688321948051,
-0.004134359769523144,
0.06227091699838638,
-0.0056213513016700745,
-0.02107818052172661,
0.04568145424127579,
0.06461413949728012,
-0.005777652375400066,
0.04308456555008888,
0.056498441845178604,
-0.05844050273299217,
-0.034487444907426834,
0.0842907577753067,
0.02223363146185875,
0.0020520661491900682,
-0.006861946079879999,
-0.08043424040079117,
0.0752214640378952,
-0.10840136557817459,
0.013458998873829842,
-0.033540450036525726,
-0.10801694542169571,
0.06811395287513733,
-0.036326296627521515,
0.024504728615283966,
0.025392360985279083,
0.020414847880601883,
0.05190873146057129,
-0.04516056552529335,
-0.04330049082636833,
0.13131830096244812,
0.006541764829307795,
-0.0451735183596611,
-0.06517279148101807,
0.009431922808289528,
0.10546740889549255,
-0.024837080389261246,
0.11326619237661362,
-0.030915748327970505,
0.040496475994586945,
0.002423836849629879,
-0.008853497914969921,
0.07075602561235428,
-0.03417449817061424,
0.011387702077627182,
0.03434154763817787,
0.019637905061244965,
0.037747789174318314,
0.03360416740179062,
0.04023126885294914,
-0.078818179666996,
0.01882253773510456,
-0.061740826815366745,
0.049876660108566284,
-0.04966907203197479,
0.057688433676958084,
0.024803638458251953,
0.08982997387647629,
-0.034461673349142075,
0.054244689643383026,
-0.002627295209094882,
-0.05953574553132057,
0.031778380274772644,
0.08141430467367172,
0.061085086315870285,
0.04271093010902405,
0.03563564270734787,
0.05141766369342804,
-0.042454514652490616,
0.012435480020940304,
0.045420803129673004,
-0.035691358149051666,
0.0004518166824709624,
-0.062322948127985,
0.03531213104724884,
-0.052343327552080154,
-0.00046267747529782355,
0.03292060270905495,
0.013182993046939373,
-0.026816237717866898,
-0.002119690179824829,
0.04359862580895424,
-0.0041535706259310246,
-0.00014911482867319137,
-0.0028589111752808094,
-0.04952702671289444,
0.056015655398368835,
0.009881758131086826,
-0.0032507465220987797,
0.008014928549528122,
0.01754995808005333,
-0.04814677685499191,
0.0008012804319150746,
0.02573586069047451,
0.10877509415149689,
0.0007095041801221669,
0.034048330038785934,
0.056302204728126526,
0.08403618633747101,
-0.03998362645506859,
0.04436316341161728,
0.031144076958298683,
-0.01438118051737547,
-0.04694105684757233,
-0.025296662002801895,
0.025890085846185684,
-0.03502368554472923,
0.04856891557574272,
-0.0907960906624794,
1.4421096020647329e-33,
0.03192955255508423,
0.03836648166179657,
0.0623910166323185,
0.006835582200437784,
0.01918916590511799,
-0.051906317472457886,
-0.03874514251947403,
-0.04762255400419235,
-0.034644171595573425,
-0.010435638949275017,
0.018151763826608658,
-0.04214900732040405,
0.03919705003499985,
0.11099742352962494,
-0.038252171128988266,
-0.02418171800673008,
0.038688596338033676,
0.08424153178930283,
-0.010548945516347885,
-0.03111271560192108,
-0.06817184388637543,
0.0020478758960962296,
-0.07725002616643906,
-0.060514990240335464,
-0.028084849938750267,
0.05757281556725502,
0.04380117356777191,
0.04115211218595505,
-0.03816911205649376,
-0.0003314547357149422,
-0.12081438302993774,
0.0591668076813221,
-0.04069531336426735,
0.014974786899983883,
0.053939834237098694,
-0.07606370747089386,
-0.03856523707509041,
-0.08118658512830734,
0.022257713600993156,
0.008329898118972778,
0.0021501732990145683,
-0.011654137633740902,
-0.07304887473583221,
0.03303886950016022,
-0.04672461003065109,
-0.030311381444334984,
-0.04926386475563049,
0.04845300316810608,
-0.013680780306458473,
-0.04027203470468521,
-0.03378038853406906,
0.0010258827824145555,
0.06417567282915115,
-0.009238001890480518,
-0.05075882375240326,
-0.03635178506374359,
-0.10285531729459763,
-0.07827762514352798,
0.09147978574037552,
-0.0476614385843277,
-0.02278551086783409,
-0.025783084332942963,
-0.08522014319896698,
0.09957855194807053,
-0.045081716030836105,
0.01595144532620907,
-0.056577395647764206,
0.025054918602108955,
0.05205630511045456,
0.10920009762048721,
0.0015740935923531651,
0.07239816337823868,
0.024211402982473373,
-0.029744839295744896,
-0.01726074330508709,
-0.031351473182439804,
0.020810062065720558,
-0.019427435472607613,
-0.05360497534275055,
0.09660468250513077,
-0.02843407168984413,
0.04531583935022354,
-0.049322355538606644,
0.047020282596349716,
-0.09164619445800781,
-0.10909082740545273,
0.07781625539064407,
-0.07514338195323944,
-0.15634563565254211,
-0.0689251571893692,
-0.020919516682624817,
-0.018211795017123222,
0.024329567328095436,
-0.03223071247339249,
-0.026138652116060257,
-2.2311911336163986e-33,
0.09679650515317917,
0.05701908469200134,
0.07928106188774109,
-0.08181724697351456,
-0.010471248999238014,
0.019549110904335976,
0.03606082499027252,
-0.08264066278934479,
0.10044508427381516,
0.008364740759134293,
0.0059616463258862495,
-0.006570666097104549,
0.0018220096826553345,
-0.06762087345123291,
0.039748188108205795,
0.021827081218361855,
-0.10233955830335617,
0.09060005843639374,
0.037818823009729385,
0.03095836564898491,
0.02312922663986683,
0.003930623177438974,
-0.07108470052480698,
0.0729367658495903,
-0.002071960363537073,
-0.014483307488262653,
-0.042833734303712845,
0.061151646077632904,
-0.048909835517406464,
0.02448078989982605,
-0.036110855638980865,
0.017953503876924515,
-0.009227295406162739,
0.002951365662738681,
-0.010566323064267635,
0.062194760888814926,
-0.05766427516937256,
-0.05472499877214432,
-0.007559351157397032,
-0.04873790591955185,
0.05497505143284798,
-0.021593721583485603,
-0.11898579448461533,
-0.04083842411637306,
-0.011977359652519226,
0.03562955930829048,
-0.050992172211408615,
0.06535419821739197,
0.013416362926363945,
-0.04337189719080925,
0.024222824722528458,
0.03213723376393318,
-0.11113578081130981,
0.05654311552643776,
0.009085113182663918,
-0.0015655519673600793,
0.037884682416915894,
0.036984026432037354,
0.06044633686542511,
-0.06010129675269127,
-0.0913623720407486,
-0.06659118086099625,
0.03609994426369667,
0.016895458102226257,
-0.006377100013196468,
0.01801171898841858,
0.04902505502104759,
0.017798839136958122,
-0.009394840337336063,
0.04365991801023483,
-0.0025895193684846163,
0.08366310596466064,
0.052418678998947144,
0.07724834978580475,
-0.010491928085684776,
0.004948142450302839,
0.012784835882484913,
-0.06191811338067055,
-0.07050911337137222,
-0.001728723174892366,
0.05474930256605148,
-0.04714381322264671,
-0.032379377633333206,
-0.001056318636983633,
0.0035834817681461573,
0.09259715676307678,
-0.007917073555290699,
0.07968667149543762,
-0.014508485794067383,
-0.04931430518627167,
-0.02388904243707657,
-0.0213899128139019,
-0.07978818565607071,
0.1285611093044281,
0.042172592133283615,
-3.697430983606864e-8,
0.05726635456085205,
0.023729052394628525,
-0.0449984110891819,
-0.03409815579652786,
-0.03211357071995735,
0.0069212219677865505,
-0.012611311860382557,
0.05085064098238945,
-0.014241868630051613,
-0.01466175727546215,
0.007235650904476643,
-0.02877604402601719,
0.045308034867048264,
-0.05027836188673973,
0.03704871982336044,
0.04432016238570213,
-0.001094656065106392,
0.038850218057632446,
-0.03666262328624725,
0.015466558746993542,
0.04161236062645912,
0.006928058806806803,
0.010337233543395996,
-0.0026864223182201385,
0.13982811570167542,
-0.09266888350248337,
-0.04549294710159302,
0.12019854784011841,
0.028140930458903313,
-0.024942783638834953,
0.06023920699954033,
-0.04121732711791992,
-0.034513890743255615,
0.03858033940196037,
0.01766236126422882,
0.117057666182518,
-0.1262560486793518,
-0.004420848097652197,
0.012653966434299946,
0.04177245870232582,
-0.08782144635915756,
0.0014493305934593081,
-0.049128253012895584,
-0.026931360363960266,
-0.028280029073357582,
-0.00843295268714428,
0.04754016175866127,
0.017074353992938995,
-0.03303678706288338,
-0.04367279261350632,
-0.00680492352694273,
0.05354263260960579,
0.037055548280477524,
-0.024252498522400856,
0.011688784696161747,
0.0018633630825206637,
-0.040549710392951965,
-0.052437569946050644,
-0.001170129980891943,
-0.07385670393705368,
0.07666482031345367,
-0.028450146317481995,
-0.02489972487092018,
-0.008674980141222477
] |
Depth Any Video with Scalable Synthetic Data | https://openreview.net/forum?id=gWqFbnKsqR | [
"Honghui Yang",
"Di Huang",
"Wei Yin",
"Chunhua Shen",
"Haifeng Liu",
"Xiaofei He",
"Binbin Lin",
"Wanli Ouyang",
"Tong He"
] | Poster | Video depth estimation has long been hindered by the scarcity of consistent and scalable ground truth data, leading to inconsistent and unreliable results. In this paper, we introduce Depth Any Video, a model that tackles the challenge through two key innovations. First, we develop a scalable synthetic data pipeline, capturing real-time video depth data from diverse virtual environments, yielding 40,000 video clips of 5-second duration, each with precise depth annotations. Second, we leverage the powerful priors of generative video diffusion models to handle real-world videos effectively, integrating advanced techniques such as rotary position encoding and flow matching to further enhance flexibility and efficiency. Unlike previous models, which are limited to fixed-length video sequences, our approach introduces a novel mixed-duration training strategy that handles videos of varying lengths and performs robustly across different frame rates—even on single frames. At inference, we propose a depth interpolation method that enables our model to infer high-resolution video depth across sequences of up to 150 frames. Our model outperforms all previous generative depth models in terms of spatial accuracy and temporal consistency. The code and model weights are open-sourced. | Video Depth Estimation, Synthetic Game Data | null | 999 | 2410.10815 | [
-0.07338441163301468,
-0.14448073506355286,
0.052613236010074615,
-0.03420432657003403,
0.09650279581546783,
0.006437238771468401,
-0.12739546597003937,
-0.04766032099723816,
0.04919394105672836,
-0.1051497608423233,
-0.05062482878565788,
-0.08662369847297668,
-0.09173461049795151,
0.037626709789037704,
-0.03327731415629387,
-0.02359345555305481,
0.0692598968744278,
0.13185828924179077,
-0.08967979997396469,
-0.02263679727911949,
0.05340519919991493,
-0.029560130089521408,
0.019801227375864983,
0.004468497820198536,
0.06113545224070549,
0.030158240348100662,
0.0426044762134552,
-0.057737164199352264,
0.010643109679222107,
-0.05081096291542053,
0.04597749933600426,
0.006997890770435333,
-0.0024511548690497875,
0.011842128820717335,
-0.025615112856030464,
0.00019527772383298725,
-0.025062117725610733,
-0.005528980400413275,
-0.038430072367191315,
-0.023267772048711777,
0.04087945446372032,
0.0381363146007061,
0.05630730837583542,
0.02154165878891945,
-0.017224933952093124,
0.06966647505760193,
0.04942246153950691,
-0.11533623188734055,
0.05156015604734421,
0.041339170187711716,
-0.06365470588207245,
0.012175358831882477,
-0.0020151014905422926,
0.08164787292480469,
-0.0467521958053112,
0.024515444412827492,
0.04342233017086983,
-0.03273040056228638,
0.036991287022829056,
0.02919023670256138,
0.03524794057011604,
-0.07734699547290802,
-0.004166347440332174,
0.00323436944745481,
-0.07974036782979965,
-0.07603717595338821,
0.02659549005329609,
0.024969518184661865,
-0.042799994349479675,
0.004581065848469734,
-0.05577582120895386,
0.06619744002819061,
-0.04702959209680557,
0.041801486164331436,
-0.020707597956061363,
-0.021762240678071976,
0.0592392273247242,
-0.05282176285982132,
0.028299475088715553,
-0.082972452044487,
0.08342698961496353,
-0.020298892632126808,
0.07360828667879105,
0.03830142691731453,
0.04200940579175949,
0.013208827003836632,
0.03397669643163681,
0.12736350297927856,
-0.056975118815898895,
0.043767381459474564,
-0.09954226016998291,
0.008766413666307926,
-0.02940310351550579,
-0.027084896340966225,
0.06760504841804504,
-0.03761427104473114,
-0.009124458767473698,
-0.08372487127780914,
0.022883106023073196,
0.05394820496439934,
0.000967209751252085,
0.027149375528097153,
-0.048472583293914795,
-0.03160111978650093,
0.07885344326496124,
0.04401100426912308,
0.0447421632707119,
0.08779537677764893,
0.006316404789686203,
0.02036045677959919,
0.05985260754823685,
-0.015153823420405388,
0.055017463862895966,
-0.033624302595853806,
0.047290727496147156,
0.04574999958276749,
-0.03680923581123352,
-0.02736605331301689,
-0.056624636054039,
0.04134655371308327,
-0.003946254961192608,
-0.04159644618630409,
-0.04742128401994705,
0.011355006136000156,
0.014309599064290524,
-0.04599645733833313,
0.04809261113405228,
-1.6989026107744744e-34,
-0.013231500051915646,
0.02269524149596691,
0.056164368987083435,
0.03656412661075592,
0.04863850399851799,
0.013592693023383617,
-0.00003238629506085999,
0.028082111850380898,
-0.03970026597380638,
0.011801891028881073,
-0.0238384660333395,
-0.0745827704668045,
-0.08528303354978561,
-0.00035581568954512477,
-0.044240742921829224,
0.022596165537834167,
-0.041403427720069885,
0.037832263857126236,
0.018564319238066673,
0.07833549380302429,
0.06630074977874756,
-0.05230104923248291,
-0.014559482224285603,
0.015762560069561005,
0.005891518667340279,
-0.027154475450515747,
0.05967889353632927,
-0.021131359040737152,
0.05512525886297226,
-0.02970876917243004,
-0.10729527473449707,
-0.01082687359303236,
0.014268658123910427,
0.02616260200738907,
0.06457803398370743,
-0.0024278457276523113,
-0.013075682334601879,
-0.015498865395784378,
-0.042104125022888184,
0.020817168056964874,
0.012078565545380116,
0.058056510984897614,
-0.018876999616622925,
0.001887324615381658,
-0.10329324752092361,
-0.09161680191755295,
0.04570123180747032,
0.058384887874126434,
-0.12802563607692719,
0.12063603103160858,
0.07539962977170944,
0.03464416414499283,
0.009292597882449627,
-0.12003536522388458,
-0.07467111200094223,
-0.026342405006289482,
0.09128968417644501,
0.03458300977945328,
0.005300352815538645,
0.024879390373826027,
-0.00005299090116750449,
0.012173143215477467,
-0.0649118572473526,
0.04029136523604393,
-0.04807659238576889,
-0.020475907251238823,
0.013617852702736855,
0.027352668344974518,
0.032385874539613724,
0.03342548757791519,
-0.04133240133523941,
0.009421992115676403,
0.0035949326120316982,
-0.0666450634598732,
0.01816803775727749,
0.009956448338925838,
-0.019405245780944824,
-0.04203781485557556,
0.013564031571149826,
0.06971357762813568,
-0.0222565159201622,
-0.003460867563262582,
0.0680253729224205,
-0.010534628294408321,
0.016721853986382484,
-0.028029609471559525,
-0.07522089034318924,
-0.03406337648630142,
-0.031232498586177826,
-0.035387713462114334,
-0.051189202815294266,
-0.09034692496061325,
-0.015578052029013634,
-0.05053672939538956,
0.06712070107460022,
1.6048438461311203e-33,
0.07391705363988876,
0.07727571576833725,
-0.02150612138211727,
0.07692869752645493,
0.04029176011681557,
-0.050632476806640625,
-0.010666737332940102,
0.10261830687522888,
0.06470632553100586,
-0.07287631183862686,
-0.03179755061864853,
0.011892936192452908,
0.04888676479458809,
0.005011920351535082,
0.02463373728096485,
-0.04239363968372345,
0.03769925981760025,
-0.057166919112205505,
-0.08595164120197296,
-0.002281932160258293,
0.03656906634569168,
-0.007358852308243513,
-0.04486491158604622,
-0.01206213142722845,
-0.07852558046579361,
0.09823393821716309,
-0.0290108323097229,
0.07082568109035492,
0.031518302857875824,
-0.009234701283276081,
0.07236337661743164,
-0.05224775895476341,
0.02622062712907791,
-0.07692814618349075,
-0.09573834389448166,
0.05378220975399017,
0.09500783681869507,
0.032994817942380905,
-0.08527009189128876,
0.06565069407224655,
0.034086596220731735,
-0.055106330662965775,
-0.025877410545945168,
0.03681739792227745,
-0.04315412417054176,
0.01140019204467535,
-0.033142756670713425,
-0.01969308964908123,
0.001216038130223751,
0.014961439184844494,
-0.07082390785217285,
0.09874007105827332,
-0.07125455141067505,
0.03026026487350464,
0.05727085843682289,
-0.10912550240755081,
-0.13877201080322266,
0.017026783898472786,
0.06748057156801224,
-0.03576648607850075,
0.020766891539096832,
-0.028589017689228058,
-0.08672894537448883,
-0.026354225352406502,
0.032648343592882156,
0.00016258608957286924,
-0.05653071776032448,
-0.00316900247707963,
-0.010317381471395493,
0.0325460284948349,
0.040588751435279846,
0.006918899714946747,
0.0072661167941987514,
0.026605214923620224,
0.028879940509796143,
-0.01975821703672409,
0.040293898433446884,
-0.000819861248601228,
0.07048239558935165,
0.04429621994495392,
0.007225843146443367,
-0.05805005878210068,
0.059353187680244446,
0.03657418489456177,
0.10383982956409454,
0.09450549632310867,
0.03942010924220085,
0.03449522331357002,
0.008010054007172585,
0.052635397762060165,
-0.0313388891518116,
0.04953775554895401,
-0.048370108008384705,
-0.013133314438164234,
-0.03317897766828537,
-3.322342578826465e-8,
-0.029682278633117676,
0.049405597150325775,
0.008466093800961971,
-0.020653577521443367,
-0.05577997490763664,
-0.04425814747810364,
0.09225825220346451,
0.08964909613132477,
0.05836721137166023,
0.010528258979320526,
0.05443226173520088,
-0.0031551134306937456,
0.06144207715988159,
-0.019570020958781242,
0.025182189419865608,
0.04553497955203056,
-0.020276464521884918,
-0.03938913345336914,
-0.01597086526453495,
0.13374054431915283,
-0.04298942908644676,
0.04810922220349312,
0.013189303688704967,
-0.05830639600753784,
0.02542002499103546,
-0.00922603439539671,
0.011815827339887619,
0.043825339525938034,
0.10120854526758194,
-0.02697605825960636,
-0.005913106258958578,
0.01271754875779152,
-0.03143823891878128,
-0.10704470425844193,
-0.039717692881822586,
-0.0011853871401399374,
0.055904142558574677,
0.024322407320141792,
0.0035298101138323545,
-0.06372896581888199,
-0.04744277894496918,
0.006936730816960335,
-0.01380279753357172,
0.0142535250633955,
-0.002120583551004529,
0.048425763845443726,
0.04446738213300705,
-0.042417917400598526,
-0.01916603371500969,
-0.017067978158593178,
0.01968252845108509,
-0.004083322826772928,
0.0002598344872239977,
0.07156901806592941,
0.09014606475830078,
0.020468637347221375,
0.029007310047745705,
-0.026480520144104958,
0.014990551397204399,
0.05100369453430176,
-0.06177378445863724,
-0.03301285579800606,
0.005035506095737219,
0.006224348209798336
] |
SPA: 3D Spatial-Awareness Enables Effective Embodied Representation | https://openreview.net/forum?id=6TLdqAZgzn | [
"Haoyi Zhu",
"Honghui Yang",
"Yating Wang",
"Jiange Yang",
"Limin Wang",
"Tong He"
] | Poster | In this paper, we introduce SPA, a novel representation learning framework that emphasizes the importance of 3D spatial awareness in embodied AI. Our approach leverages differentiable neural rendering on multi-view images to endow a vanilla Vision Transformer (ViT) with intrinsic spatial understanding. We present the most comprehensive evaluation of embodied representation learning to date, covering 268 tasks across 8 simulators with diverse policies in both single-task and language-conditioned multi-task scenarios. The results are compelling: SPA consistently outperforms more than 10 state-of-the-art representation methods, including those specifically designed for embodied AI, vision-centric tasks, and multi-modal applications, while using less training data. Furthermore, we conduct a series of real-world experiments to confirm its effectiveness in practical scenarios. These results highlight the critical role of 3D spatial awareness for embodied representation learning. Our strongest model takes more than 6000 GPU hours to train and we are committed to open-sourcing all code and model weights to foster future research in embodied representation learning. | embodied AI, representation learning, 3D spatial awareness, multi-view image, robot manipulation, neural rendering | We introduce SPA, a novel framework that enhances 3D spatial awareness in embodied AI representation learning, outperforming existing models across 268 tasks and 8 simulators. | 996 | 2410.08208 | [
0.01134521421045065,
-0.1438388079404831,
0.05079447850584984,
-0.006484257522970438,
0.037145331501960754,
0.028485620394349098,
-0.0035249851644039154,
-0.06461354345083237,
0.02358521707355976,
-0.01669369451701641,
-0.06064886599779129,
-0.04396917298436165,
-0.01164581160992384,
0.09359937161207199,
0.042193200439214706,
0.007339006755501032,
0.13736878335475922,
0.07716193795204163,
-0.09675856679677963,
0.04012240469455719,
0.07317627221345901,
-0.05636805295944214,
0.03778607025742531,
0.0033358752261847258,
0.08125953376293182,
0.009205145761370659,
0.04090491309762001,
-0.1111476719379425,
0.0006825134623795748,
-0.04479155316948891,
0.026358960196375847,
-0.04922487959265709,
-0.04982183873653412,
-0.02878120355308056,
-0.03170644864439964,
0.05281520262360573,
-0.023938177153468132,
0.0038280589506030083,
-0.038799479603767395,
-0.09590526670217514,
-0.04687454551458359,
0.011916324496269226,
-0.0016366071067750454,
0.046871814876794815,
0.08233282715082169,
0.032834261655807495,
0.02367825247347355,
-0.1072879433631897,
0.045029543340206146,
-0.04998011887073517,
-0.05787188187241554,
-0.12246193736791611,
0.04467848688364029,
-0.04170827940106392,
-0.03852678835391998,
0.03465067595243454,
0.036932624876499176,
-0.029974699020385742,
-0.022308489307761192,
-0.01088622398674488,
0.025608722120523453,
-0.05003087595105171,
0.05737294256687164,
0.009970391169190407,
0.014677217230200768,
-0.09757545590400696,
0.026858583092689514,
-0.007573742885142565,
-0.0234981719404459,
-0.06445017457008362,
-0.003401367459446192,
0.018928904086351395,
-0.004381394479423761,
-0.0889124646782875,
0.08266513049602509,
0.02483295276761055,
0.004029812291264534,
-0.07981691509485245,
0.04341008886694908,
-0.08894066512584686,
0.06795717775821686,
0.07132278382778168,
-0.010354903526604176,
0.039100632071495056,
0.06446465104818344,
0.028171520680189133,
-0.014723192900419235,
0.08947127312421799,
-0.045795489102602005,
0.048808518797159195,
-0.019808780401945114,
-0.041758958250284195,
0.036842141300439835,
-0.03221892938017845,
0.003746018512174487,
-0.03944997861981392,
-0.012193326838314533,
-0.08830320090055466,
-0.04021542891860008,
0.0478685200214386,
-0.03343764692544937,
-0.013169683516025543,
0.08011520653963089,
0.03131905570626259,
0.023850327357649803,
-0.003005116945132613,
0.023371446877717972,
0.03987663611769676,
-0.04501035064458847,
-0.01622307114303112,
0.04338601231575012,
-0.027102384716272354,
-0.0041793338023126125,
0.08007127791643143,
-0.009151792153716087,
0.017308035865426064,
0.019471103325486183,
-0.0009336655493825674,
0.08545547723770142,
0.034109145402908325,
-0.021165452897548676,
-0.009374622255563736,
0.017789417877793312,
0.023011278361082077,
0.024124186486005783,
-0.04738567769527435,
-0.0296697448939085,
-7.846748971040207e-34,
0.027466870844364166,
0.027019880712032318,
0.05298982188105583,
0.033613815903663635,
0.04897566884756088,
-0.04033812880516052,
0.025042511522769928,
0.028054242953658104,
-0.006344756111502647,
0.014081154018640518,
-0.029152818024158478,
0.007944627664983273,
0.014725952409207821,
0.17899364233016968,
-0.016727901995182037,
-0.027201248332858086,
-0.0564718022942543,
0.05546189472079277,
-0.02653946727514267,
0.027125578373670578,
0.029765494167804718,
-0.01264185644686222,
-0.03470441326498985,
-0.006700439378619194,
-0.07399780303239822,
0.004265415482223034,
0.003666964825242758,
-0.04893943667411804,
-0.05122571066021919,
-0.012363520450890064,
-0.09411628544330597,
0.07033167034387589,
-0.013616911135613918,
-0.0004559444496408105,
0.04563178867101669,
-0.041279375553131104,
-0.0037580819334834814,
-0.054674793034791946,
0.016353469341993332,
0.03089829906821251,
0.009797364473342896,
0.04381122440099716,
-0.0004378591547720134,
-0.0547836497426033,
-0.05724543333053589,
0.017264079302549362,
0.021706560626626015,
0.0732320174574852,
-0.06015953794121742,
-0.006691962014883757,
0.03586481511592865,
0.0553751178085804,
-0.04803978651762009,
-0.050564881414175034,
-0.004064044449478388,
-0.06579773873090744,
0.011821597814559937,
0.03406276926398277,
0.08049272000789642,
-0.04172086343169212,
-0.02177290804684162,
0.015812458470463753,
0.015280188992619514,
0.07326525449752808,
0.005407711490988731,
0.017380518838763237,
0.02620822750031948,
0.09418938308954239,
0.04517655819654465,
0.0439794696867466,
-0.06803455203771591,
0.03330580145120621,
-0.021623395383358,
-0.05118080973625183,
0.01448359526693821,
-0.05176343768835068,
0.07358915358781815,
-0.10269201546907425,
-0.013355337083339691,
0.05973658710718155,
0.015270480886101723,
0.004683276638388634,
0.011156372725963593,
0.004981433041393757,
-0.008357110433280468,
-0.0738009661436081,
0.03625403717160225,
-0.08518722653388977,
-0.03075219690799713,
-0.010636214166879654,
-0.07357726991176605,
-0.11705891788005829,
-0.03542402759194374,
-0.037067823112010956,
-0.04052960127592087,
6.254529934236918e-34,
0.02754264697432518,
0.013274593278765678,
-0.036965176463127136,
0.002121944446116686,
-0.05400853976607323,
-0.07552395015954971,
0.013646475039422512,
0.03498541936278343,
-0.03369477763772011,
-0.016513867303729057,
-0.053214456886053085,
0.02740277536213398,
0.044328473508358,
0.0062638442032039165,
0.04486904293298721,
0.0015666401013731956,
0.000787644530646503,
-0.012311938218772411,
-0.06174318864941597,
0.01449368055909872,
0.02443500980734825,
0.06546664237976074,
-0.03933635726571083,
0.02477247081696987,
-0.0886053740978241,
0.0851091668009758,
-0.014529484324157238,
0.025967277586460114,
-0.026442011818289757,
0.06806878745555878,
0.022407911717891693,
-0.02368403598666191,
0.09769370406866074,
0.00951384101063013,
-0.06890495866537094,
0.014496623538434505,
0.07644940167665482,
0.0008301582420244813,
-0.04897039383649826,
0.023363368585705757,
0.04524703323841095,
-0.07264168560504913,
-0.030677765607833862,
0.04464751109480858,
-0.03081529773771763,
0.024289218708872795,
-0.05522981658577919,
-0.04611662030220032,
-0.04344417154788971,
0.09091229736804962,
-0.00921965204179287,
0.042050886899232864,
-0.04114894941449165,
-0.060494184494018555,
-0.024544185027480125,
-0.15506158769130707,
-0.008458430878818035,
-0.018496910110116005,
0.08932165801525116,
-0.0037826078478246927,
-0.013143775053322315,
-0.0038254959508776665,
-0.0035350972320884466,
-0.018025685101747513,
-0.008783150464296341,
0.019551819190382957,
-0.08498009294271469,
-0.010672066360712051,
0.05671502649784088,
-0.0019005778012797236,
0.014837097376585007,
0.008279281668365002,
-0.010301892645657063,
0.013752874918282032,
0.006874704733490944,
0.02205456979572773,
0.08762581646442413,
-0.0003913148830179125,
0.015158778987824917,
0.01753339357674122,
-0.0003055762208532542,
-0.07250091433525085,
0.06379694491624832,
0.06522280722856522,
0.138881117105484,
0.15783773362636566,
-0.045983001589775085,
0.03809560462832451,
0.0029869009740650654,
0.040347859263420105,
-0.029413491487503052,
0.1328549087047577,
-0.02247488498687744,
0.04115254059433937,
0.019780009984970093,
-3.6838564199115353e-8,
-0.12806087732315063,
0.03805268183350563,
0.09946722537279129,
-0.011580215767025948,
-0.05107446759939194,
0.03278202563524246,
0.03692343458533287,
0.01998208835721016,
-0.08438186347484589,
-0.0037327036261558533,
0.0388435423374176,
-0.07963661104440689,
0.017313048243522644,
-0.012272913008928299,
0.0966222733259201,
0.05346868559718132,
0.0014778943732380867,
0.03966505825519562,
-0.03280823305249214,
0.0426906943321228,
-0.012128633446991444,
0.04577537253499031,
0.00933685339987278,
-0.07693210244178772,
0.013520597480237484,
-0.037225350737571716,
-0.08880732953548431,
0.0875479057431221,
0.03131532296538353,
-0.03592010959982872,
0.056652847677469254,
0.027266299352049828,
0.04622798040509224,
-0.03924562409520149,
-0.04752356931567192,
0.03319716081023216,
0.03686303272843361,
-0.032185494899749756,
-0.02426309883594513,
-0.04450962319970131,
-0.014690117910504341,
0.06847517937421799,
0.001335826120339334,
-0.012102804146707058,
0.038440339267253876,
0.020898137241601944,
-0.02099878527224064,
-0.1306326538324356,
-0.019179057329893112,
0.02254503220319748,
-0.08702737092971802,
0.013446475379168987,
-0.043517597019672394,
0.15414999425411224,
0.09139705449342728,
0.00562939140945673,
-0.02469663694500923,
-0.0060885390266776085,
0.018297750502824783,
0.10214866697788239,
0.01729939691722393,
-0.002649900270625949,
-0.10161377489566803,
-0.008807510137557983
] |
Incremental Causal Effect for Time to Treatment Initialization | https://openreview.net/forum?id=0mtz0pet1z | [
"Andrew Ying",
"Zhichen Zhao",
"Ronghui Xu"
] | Poster | We consider time to treatment initialization. This can commonly occur in preventive medicine, such as disease screening and vaccination; it can also occur with non-fatal health conditions such as HIV infection without the onset of AIDS. While traditional causal inference focused on ‘when to treat’ and its effects, including their possible dependence on subject characteristics, we consider the incremental causal effect when the intensity of time to treatment initialization is intervened upon. We provide identification of the incremental causal effect without the commonly required positivity assumption, as well as an estimation framework using inverse probability weighting. We illustrate our approach via simulation, and apply it to a rheumatoid arthritis study to evaluate the incremental effect of time to start methotrexate on joint pain. | Causal Inference, Positivity, Incremental intervention, Incremental Causal Effect, Inverse probability weighting | We propose a novel causal estimand for studies with continuous time to initialize treatment. | 995 | 2409.13097 | [
-0.05372725799679756,
-0.037842754274606705,
0.07652057707309723,
0.08207663148641586,
0.027060866355895996,
0.00009930512896971777,
0.004895064979791641,
0.09290696680545807,
0.10782690346240997,
0.00021946871129330248,
-0.041442256420850754,
0.0858536884188652,
0.05724802985787392,
0.06242584064602852,
-0.050094060599803925,
-0.0020986213348805904,
0.12393031269311905,
-0.0220450758934021,
-0.017951011657714844,
0.05165255442261696,
-0.029740918427705765,
-0.03162567317485809,
-0.025281107053160667,
-0.06371401250362396,
-0.03108285926282406,
-0.05290240794420242,
0.007790316361933947,
0.03875480964779854,
0.030425168573856354,
0.05800841003656387,
0.03988886997103691,
0.06614456325769424,
-0.0036944712046533823,
-0.032731667160987854,
-0.0915352925658226,
0.02425212226808071,
-0.0983114242553711,
-0.022652197629213333,
-0.051968831568956375,
0.025417402386665344,
0.012573245912790298,
0.007998154498636723,
-0.0030368026345968246,
0.020031942054629326,
0.1022837683558464,
-0.031508177518844604,
0.009248637594282627,
0.0725395679473877,
-0.020722437649965286,
0.07903438806533813,
0.029715614393353462,
0.024542026221752167,
0.02199736051261425,
-0.014494157396256924,
0.04239548742771149,
0.004025585949420929,
0.01929359883069992,
-0.0722646713256836,
-0.06310444325208664,
0.022489190101623535,
-0.03143411502242088,
-0.0006948218797333539,
-0.035972416400909424,
0.000731668493244797,
0.06482455879449844,
0.018987048417329788,
0.03452247753739357,
0.019169189035892487,
0.012547088786959648,
0.13352197408676147,
-0.07259664684534073,
0.06191202625632286,
-0.06137612834572792,
0.043983545154333115,
-0.012943735346198082,
-0.05030925199389458,
-0.024650247767567635,
0.06425857543945312,
0.06042546406388283,
-0.07737284898757935,
0.03102598339319229,
0.032665908336639404,
0.015105711296200752,
-0.0105277793481946,
0.028968434780836105,
-0.004968824330717325,
0.03145027160644531,
0.10038627684116364,
0.017842356115579605,
0.024275347590446472,
-0.001746390713378787,
0.054026439785957336,
-0.05202505737543106,
0.007009138353168964,
0.017366833984851837,
0.02844577096402645,
-0.05254276841878891,
0.019241653382778168,
-0.01619853265583515,
0.03315987065434456,
-0.08708349615335464,
-0.04049032926559448,
0.040586329996585846,
0.042833730578422546,
-0.02083701267838478,
-0.08879978209733963,
-0.0037801158614456654,
0.002934441901743412,
-0.04670378193259239,
0.01723654940724373,
0.053228601813316345,
-0.02820778824388981,
0.08925969153642654,
-0.012709615752100945,
-0.031285036355257034,
0.022882845252752304,
0.060383427888154984,
0.00018558897136244923,
0.013596368953585625,
-0.07072827219963074,
0.10110988467931747,
-0.09120350331068039,
0.006962853018194437,
-0.08435007929801941,
-0.006317047867923975,
0.01877431944012642,
0.06388768553733826,
2.1465685809976484e-33,
0.036562733352184296,
-0.051598433405160904,
0.015597989782691002,
0.01972346566617489,
-0.009818215854465961,
-0.033944424241781235,
-0.023802120238542557,
-0.04015447199344635,
0.029462268576025963,
0.00437941774725914,
0.02180490456521511,
0.024401221424341202,
0.006639271043241024,
0.030583512037992477,
-0.042902469635009766,
0.05665222555398941,
-0.005258571356534958,
0.13721545040607452,
-0.02401505410671234,
-0.014835803769528866,
0.0153978755697608,
-0.11381010711193085,
-0.054582562297582626,
0.015741905197501183,
-0.07261130958795547,
0.02203664928674698,
-0.020556028932332993,
-0.003874445566907525,
-0.023451795801520348,
-0.04202666133642197,
-0.0017287768423557281,
0.035847727209329605,
-0.008570145815610886,
-0.03107384778559208,
-0.042125117033720016,
-0.033619191497564316,
0.02565997838973999,
-0.022410733625292778,
-0.0385131910443306,
-0.02640131674706936,
0.06544415652751923,
0.025447819381952286,
0.05029924958944321,
-0.045084260404109955,
-0.016815321519970894,
-0.07316282391548157,
-0.11469551175832748,
-0.03126366063952446,
-0.03483365848660469,
0.032330118119716644,
0.028631987050175667,
0.010973632335662842,
0.08331456780433655,
-0.08604440838098526,
-0.10680967569351196,
0.03348514065146446,
-0.0704316720366478,
0.046249013394117355,
-0.04326095059514046,
0.047697991132736206,
-0.0061065638437867165,
-0.018752897158265114,
-0.0178285650908947,
0.004262040369212627,
-0.012190230190753937,
0.11323600262403488,
0.013469576835632324,
-0.11390365660190582,
0.03662744536995888,
0.048874933272600174,
0.04208047315478325,
0.013566408306360245,
-0.034033771604299545,
0.07014599442481995,
0.02724745124578476,
0.010617355816066265,
0.011802145279943943,
0.015431031584739685,
0.024960054084658623,
-0.005102305673062801,
-0.016924215480685234,
-0.07079071551561356,
-0.050335466861724854,
0.06080299988389015,
0.030013300478458405,
-0.017971213907003403,
0.018177278339862823,
-0.09192418307065964,
-0.1516125649213791,
-0.10634594410657883,
0.0053704651072621346,
0.016798807308077812,
0.026005592197179794,
0.06061364337801933,
0.07539574801921844,
-3.629295829151611e-33,
-0.011350126005709171,
0.0480777844786644,
0.033526722341775894,
-0.03858513385057449,
0.05940351262688637,
0.007359858136624098,
0.013219700194895267,
-0.00002685244908207096,
0.09029959887266159,
-0.04189680889248848,
0.03905041515827179,
-0.0790516659617424,
-0.015228131785988808,
0.04786930978298187,
-0.025333158671855927,
0.03197210282087326,
0.021109627559781075,
-0.0023861790541559458,
-0.05363883450627327,
0.07937055826187134,
0.05806725472211838,
0.024154534563422203,
-0.007447717245668173,
-0.040029022842645645,
-0.011727889068424702,
0.004335091914981604,
0.05238264426589012,
-0.008095061406493187,
-0.05569710582494736,
-0.0336320661008358,
-0.09176062047481537,
-0.0398622564971447,
-0.014633676037192345,
0.015282192267477512,
-0.10501854866743088,
0.10230317711830139,
0.04368466138839722,
-0.054634906351566315,
-0.005090671591460705,
-0.028450369834899902,
0.11768350005149841,
-0.026955071836709976,
-0.11750668287277222,
-0.03563405200839043,
0.031783170998096466,
-0.07179908454418182,
-0.023968910798430443,
0.061126548796892166,
0.09067007899284363,
-0.04372293874621391,
-0.012713465839624405,
0.07381551712751389,
0.07666687667369843,
0.04401891306042671,
-0.016637751832604408,
-0.10573729127645493,
0.059522151947021484,
-0.16323736310005188,
0.003603168996050954,
0.03570837154984474,
-0.11336368322372437,
0.036210861057043076,
0.055548638105392456,
0.0016837305156514049,
-0.03327355161309242,
0.03064936399459839,
0.003290001302957535,
-0.03956269472837448,
-0.0012749260058626533,
-0.057468216866254807,
-0.04527103155851364,
0.043923262506723404,
-0.01660001091659069,
-0.05695490166544914,
0.0066658672876656055,
-0.06384889036417007,
0.023465026170015335,
-0.007827677763998508,
-0.05524902790784836,
-0.06139523163437843,
-0.05548415705561638,
-0.06510099768638611,
0.006006889045238495,
-0.0036541554145514965,
-0.09357325732707977,
0.04277591407299042,
0.012086305767297745,
-0.025122541934251785,
0.06494911015033722,
0.07777750492095947,
-0.07681214809417725,
0.007655189838260412,
0.01470864750444889,
0.054529570043087006,
-0.09507914632558823,
-4.3014953376996345e-8,
0.061051685363054276,
0.027184007689356804,
0.013110932894051075,
0.039172906428575516,
0.06380154192447662,
0.05505779758095741,
-0.07137081027030945,
-0.03796868398785591,
-0.019767843186855316,
-0.03481002151966095,
0.04265706241130829,
0.0942850187420845,
0.0800447016954422,
-0.004403256345540285,
-0.08460046350955963,
0.08306633681058884,
0.016108088195323944,
-0.001025839359499514,
-0.06196603551506996,
-0.0533917061984539,
-0.0004073869495186955,
0.013023249804973602,
-0.009603003971278667,
-0.06938420981168747,
0.09357412159442902,
-0.023209156468510628,
-0.01609746739268303,
-0.012521035969257355,
0.04440025985240936,
0.0094232689589262,
-0.04040158540010452,
0.06209241971373558,
-0.06237410008907318,
0.07193920016288757,
-0.010957973077893257,
-0.03643960505723953,
0.05292828381061554,
-0.005654003005474806,
-0.04507219418883324,
-0.08191873133182526,
0.004004382528364658,
0.09164579957723618,
-0.029782917350530624,
0.05497479438781738,
0.026556167751550674,
-0.013141846284270287,
-0.12402085214853287,
-0.03009062632918358,
0.005780064966529608,
-0.043693941086530685,
-0.008100457489490509,
0.012688695453107357,
0.0044922735542058945,
-0.059001583606004715,
-0.006212335545569658,
0.054907023906707764,
0.02686138078570366,
-0.03740578889846802,
0.011207563802599907,
-0.02642912045121193,
0.00697630550712347,
-0.020349910482764244,
0.0017600608989596367,
0.008915767073631287
] |
Leveraging Sub-Optimal Data for Human-in-the-Loop Reinforcement Learning | https://openreview.net/forum?id=DSyHRkpI7v | [
"Calarina Muslimani",
"Matthew E. Taylor"
] | Poster | To create useful reinforcement learning (RL) agents, step zero is to design a suitable reward function that captures the nuances of the task. However, reward engineering can be a difficult and time-consuming process.
Instead, human-in-the-loop RL methods hold the promise of learning reward functions from human feedback. Despite recent successes, many of the human-in-the-loop RL methods still require numerous human interactions to learn successful reward functions.
To improve the feedback efficiency of human-in-the-loop RL methods (i.e., require less human interaction), this paper introduces Sub-optimal Data Pre-training, SDP, an approach that leverages reward-free, sub-optimal data to improve scalar- and preference-based RL algorithms. In SDP, we start by pseudo-labeling all low-quality data with the minimum environment reward. Through this process, we obtain reward labels
to pre-train our reward model without requiring human labeling or preferences.
This pre-training phase provides the reward model a head start in learning, enabling it to recognize that low-quality transitions should be assigned low rewards. Through extensive experiments with both simulated and human teachers, we find that SDP can at least meet, but often significantly improve, state of the art human-in-the-loop RL performance across a variety of simulated robotic tasks. | Reinforcement Learning, Human-in-the-loop, Preference learning, Learning from scalar feedback | null | 990 | 2405.00746 | [
-0.052819568663835526,
-0.01990457810461521,
0.017121272161602974,
0.02957414649426937,
-0.023956049233675003,
0.018825963139533997,
0.0024749443400651217,
0.03344695642590523,
-0.01666409522294998,
-0.0021834804210811853,
-0.035778701305389404,
-0.08410106599330902,
0.02719639055430889,
0.06651099026203156,
0.0717935562133789,
0.07904386520385742,
0.11057225614786148,
-0.003763393498957157,
-0.018915340304374695,
-0.04909488931298256,
0.0024769147858023643,
-0.0007423998322337866,
0.01997152715921402,
0.004983136430382729,
-0.009599474258720875,
-0.021364597603678703,
0.023243572562932968,
-0.024646982550621033,
0.058846380561590195,
0.023761123418807983,
0.0642770305275917,
-0.05401702597737312,
0.10296186059713364,
-0.06972845643758774,
-0.018029922619462013,
0.05789288505911827,
-0.09168454259634018,
-0.03233249858021736,
-0.045508693903684616,
0.02339339815080166,
-0.02678374946117401,
-0.02242359146475792,
-0.01199498400092125,
0.07060343772172928,
0.0019979956559836864,
0.013756109401583672,
-0.03469369187951088,
-0.0767354890704155,
0.026166196912527084,
0.00944492593407631,
-0.007109867874532938,
0.053158506751060486,
0.015448623336851597,
0.012620928697288036,
-0.00009326582949142903,
0.06873553991317749,
0.029291126877069473,
0.009318103082478046,
0.0183878093957901,
-0.07252602279186249,
-0.012998390942811966,
-0.15186133980751038,
-0.009004914201796055,
0.021251117810606956,
-0.07340651750564575,
-0.0627349242568016,
-0.014572235755622387,
0.04148465022444725,
0.039688922464847565,
0.010650109499692917,
-0.013064905069768429,
0.07015693932771683,
0.029148900881409645,
-0.012524068355560303,
-0.009029959328472614,
-0.015811290591955185,
-0.023380834609270096,
-0.020962893962860107,
0.0384877473115921,
-0.01496030017733574,
-0.0445002019405365,
0.011783607304096222,
-0.023669425398111343,
0.0363604910671711,
0.0720207467675209,
0.02188671939074993,
0.06398133933544159,
0.04608071222901344,
0.029187316074967384,
0.00305794645100832,
-0.056915849447250366,
0.05492047965526581,
-0.03800848871469498,
-0.01947552151978016,
-0.08554055541753769,
0.12123820185661316,
-0.0355241484940052,
-0.08478958159685135,
-0.0754120945930481,
0.11014039069414139,
0.017315072938799858,
0.07502362877130508,
-0.08390118926763535,
0.017012054100632668,
0.0012026390759274364,
-0.06745775789022446,
0.02963024377822876,
0.049823515117168427,
0.01268051564693451,
-0.07670244574546814,
0.025994734838604927,
0.013259055092930794,
0.012273565866053104,
0.00997824314981699,
-0.021739626303315163,
-0.0006951378891244531,
0.021835781633853912,
-0.04167194291949272,
-0.009417001157999039,
0.05440457910299301,
0.0112397950142622,
-0.052404675632715225,
0.024604223668575287,
0.047937557101249695,
0.02978520281612873,
-0.06506529450416565,
-0.05586980655789375,
-1.4542784481541398e-33,
0.03922168165445328,
0.027489639818668365,
0.07621841877698898,
-0.03604760393500328,
0.015515950508415699,
-0.033782973885536194,
0.03189094737172127,
-0.019266575574874878,
-0.0418141707777977,
0.005200403276830912,
0.06533247977495193,
0.011687698774039745,
-0.1019611656665802,
0.08402758091688156,
0.049861446022987366,
-0.041305843740701675,
-0.02424391731619835,
0.012967238202691078,
0.0032685385085642338,
-0.07898207753896713,
0.013159209862351418,
-0.03544161096215248,
0.067116878926754,
-0.00024477194529026747,
0.04505397751927376,
0.032871127128601074,
0.044060926884412766,
-0.01732855848968029,
0.010083094239234924,
-0.03692937269806862,
-0.007388512138277292,
-0.007284474093466997,
0.09336791932582855,
0.0009474660619162023,
-0.06435209512710571,
0.024688048288226128,
0.004207867197692394,
-0.0017319063190370798,
-0.014651893638074398,
0.06369636207818985,
0.054480571299791336,
0.04557115212082863,
0.023667393252253532,
0.02036006562411785,
-0.05171794816851616,
-0.0911649689078331,
0.07911774516105652,
0.009765395894646645,
-0.12517577409744263,
-0.000019582303139031865,
0.027334686368703842,
0.00922916829586029,
-0.014626850374042988,
-0.05234231427311897,
-0.06614045053720474,
0.026380490511655807,
0.051305823028087616,
0.047450099140405655,
-0.025252612307667732,
0.04459872096776962,
-0.02391819842159748,
-0.0010221408447250724,
0.008797429502010345,
0.08474596589803696,
-0.019698243588209152,
0.06990061700344086,
0.012581509537994862,
0.00459256349131465,
0.00921428669244051,
-0.01983095332980156,
-0.008401492610573769,
0.08990701287984848,
0.01219012401998043,
-0.040827427059412,
0.1346183568239212,
0.0381111316382885,
0.035377003252506256,
-0.14417192339897156,
0.05876631662249565,
-0.04276656359434128,
0.01422630250453949,
-0.035071227699518204,
-0.03179362416267395,
-0.011105531826615334,
0.02321256697177887,
-0.055360548198223114,
0.004739462863653898,
-0.09962239116430283,
0.02377518266439438,
-0.0006925856578163803,
-0.1217871829867363,
-0.011574896983802319,
-0.023635828867554665,
0.04623303934931755,
0.01651589572429657,
1.9474089812416438e-33,
0.058198608458042145,
-0.05831359699368477,
-0.022642958909273148,
0.028446992859244347,
0.01092074066400528,
0.032997164875268936,
0.021471943706274033,
-0.047702934592962265,
0.013549352064728737,
0.02023710310459137,
-0.04036884754896164,
-0.019265322014689445,
0.04827805235981941,
0.019891509786248207,
0.017624596133828163,
-0.018900496885180473,
-0.01801101118326187,
0.021703427657485008,
-0.011610749177634716,
0.002773020649328828,
-0.0058865537866950035,
0.1900375336408615,
-0.10990648716688156,
0.0640619620680809,
-0.016290487721562386,
0.021262802183628082,
0.04794996231794357,
0.06489083915948868,
0.0289307851344347,
0.04024116322398186,
-0.06244676187634468,
-0.04829196631908417,
0.012733071111142635,
-0.07453146576881409,
-0.04743651673197746,
0.06925325840711594,
0.08033239841461182,
0.008252830244600773,
-0.06997881829738617,
0.07349418848752975,
0.03915909305214882,
-0.07989327609539032,
-0.08606915920972824,
-0.061103131622076035,
-0.06789599359035492,
-0.06867828965187073,
-0.09523262828588486,
0.007319817319512367,
-0.060384124517440796,
-0.01129742618650198,
-0.07318674027919769,
0.015638137236237526,
-0.11154253035783768,
-0.015475127846002579,
-0.04961903765797615,
-0.038950253278017044,
-0.09817735850811005,
-0.01990582048892975,
0.12639930844306946,
-0.017720689997076988,
-0.07199998199939728,
0.03923342004418373,
-0.005016133654862642,
0.0648449957370758,
-0.03154611214995384,
-0.03672048822045326,
0.038305968046188354,
0.029098549857735634,
0.0042568594217300415,
-0.010509181767702103,
0.02610589750111103,
0.04926864802837372,
0.050240159034729004,
-0.06752203404903412,
0.05090111866593361,
-0.028185324743390083,
-0.03561415895819664,
-0.002905159490182996,
0.0649384930729866,
-0.04857797175645828,
0.008509506471455097,
0.07973971962928772,
0.015624910593032837,
0.056086763739585876,
0.048176780343055725,
0.07441917806863785,
-0.027253732085227966,
0.059162843972444534,
0.002568142954260111,
-0.018338065594434738,
-0.04159319028258324,
0.0364467091858387,
0.04902264103293419,
0.11048366874456406,
-0.05615025386214256,
-4.3342488709186e-8,
-0.06815830618143082,
0.0211242213845253,
0.10728434473276138,
0.06967025250196457,
0.0493786595761776,
0.0036952097434550524,
-0.012061277404427528,
0.05610204488039017,
0.01362469419836998,
-0.011960450559854507,
0.023549845442175865,
0.056443531066179276,
-0.017812665551900864,
-0.10222893208265305,
-0.04515834525227547,
0.07771293818950653,
0.0484481081366539,
-0.026606343686580658,
0.040064241737127304,
-0.04312124103307724,
0.025132672861218452,
0.0318867564201355,
0.009150560013949871,
-0.10565899312496185,
0.0546712726354599,
-0.08077676594257355,
-0.028895702213048935,
0.09506746381521225,
0.10193297266960144,
-0.03746509179472923,
0.006340434774756432,
-0.020438579842448235,
0.0650157630443573,
0.05186119303107262,
0.10795603692531586,
0.029284711927175522,
0.05418054014444351,
-0.011503922753036022,
-0.03872581198811531,
-0.04731951653957367,
-0.00836967583745718,
0.10134304314851761,
-0.03146616369485855,
-0.07157305628061295,
-0.04940696433186531,
0.02299441769719124,
-0.07195677608251572,
-0.08585536479949951,
0.0648791491985321,
-0.06240484118461609,
-0.004127610940486193,
-0.07797878235578537,
-0.060694921761751175,
0.028562067076563835,
0.055648788809776306,
-0.010816416703164577,
0.012401724234223366,
-0.026923589408397675,
0.03136616200208664,
0.05162883922457695,
0.008155381307005882,
0.015555857680737972,
-0.04063641279935837,
-0.004732051398605108
] |
Subsets and Splits