WhatAreBayesianNeuralNetworkPosteriorsReallyLike?PavelIzmailov1SharadVikram2MatthewD.Hoffman2AndrewGordonWilson1Abstractforneuralnetworkspromisesimprovedpredictions,reli-ableuncertaintyestimates,an...
TwoHeadsAreBetterThanOne:Hypergraph-EnhancedGraphReasoningforVisualEventRatiocinationWenboZheng12LanYan23ChaoGou4Fei-YueWang2Abstract1.IntroductionEvenwithastillimage,humanscanratiocinateInaclassro...
NotAllMemoriesAreCreatedEqual:LearningtoForgetbyExpiringSainbayarSukhbaatar1DaJu1SpencerPoff1StephenRoller1ArthurSzlam1JasonWeston1AngelaFan12AbstractSukhbaatAretal.,2019a).However,acriticalcompone...
LinearTransformersAreSecretlyFastWeightProgrammersImanolSchlag∗1KazukiIrie∗1Ju¨rgenSchmidhuber1Abstractfieldnetwork(Ramsaueretal.,2021;Krotov&Hopfield,2016;Demircigiletal.,2017).Itextendsaformof...
EvaluatingRobustnessofPredictiveUncertaintyEstimation:AreDirichlet-basedModelsReliable?Anna-KathrinKopetzki1BertrandCharpentier1DanielZügner1SandhyaGiri1StephanGünnemann1AbstractFigure1.Visualiza...
CumulantsofHawkesProcessesAreRobusttoObservationNoiseWilliamTrouleau1JalalEtesami2MatthiasGrossglauser1NegarKiyavash2PatrickThiran1AbstracttheyAreusedtomodelthestochastictimeevolutionoflimitorderbo...
TransformersAreRNNs:FastAutoregressiveTransformerswithLinearAttentionAngelosKatharopoulos12ApoorvVyas12NikolaosPappas3Franc¸oisFleuret12Abstractbytheglobalreceptivefieldofself-attention,whichpro-c...
NeuralNetworksAreConvexRegularizers:ExactPolynomial-timeConvexOptimizationFormulationsforTwo-layerNetworksMertPilanci1TolgaErgen1Abstractconnectionstokernelmethods,andshowedthatrandomlyinitializedg...
InterpretationsAreUseful:PenalizingExplanationstoAlignNeuralNetworkswithPriorKnowledgeLauraRieger1ChandanSingh2W.JamesMurdoch3BinYu23Abstractimproveamodel.ForanexplanationofadeeplearningmodeltoPred...
TheOddsAreOdd:AStatisticalTestforDetectingAdversarialExamplesKevinRoth1YannicKilcher1ThomasHofmann1Abstract2017;Trame`retal.,2017).ThisappArentvulnerabilityisworrisomeasdeepnetsstarttoproliferatein...
OptimalityImpliesKernelSumClassifiersAreStatisticallyEfficientRaphaelA.Meyer1JeanHonorio1AbstractweconsiderbinaryclassificationwithKernelSupportVec-torMachines(SVM),whichArecomputedbysolvingaWeprop...
AreGenerativeClassifiersMoreRobusttoAdversarialAttacks?YingzhenLi1JohnBradshaw23YashSharma4AbstractAdversarialtraining,whichaugmentsthetrainingdatawithadversariallyperturbedinputs,hasshownmoderates...
AdversarialExamplesAreaNaturalConsequenceofTestErrorinNoiseNicolasFord12JustinGilmer1NicholasCarlini1EkinD.Cubuk1Abstract(Rosenfeldetal.,2018).Atthesametime,theyArealsosensitivetosmall,worst-casepe...
WhenSamplesAreStrategicallySelectedHanruiZhang1YuCheng1VincentConitzer1Abstract1988;Friedman,1993)),thereisstilltheconcernthattheagentsendsonlyabiasedcollectionofsamples.IfwedoInstandardclassificat...
TighterVariationalBoundsAreNotNecessarilyBetterTomRainforth1AdamR.Kosiorek12TuanAnhLe2ChrisJ.Maddison1MaximilianIgl2FrankWood3YeeWhyeTeh1Abstract&Kamp,1988;Hinton&Zemel,1994;Gregoretal.,2016;Chenet...
SpuriousLocalMinimaAreCommoninTwo-LayerReLUNeuralNetworksItaySafran1OhadShamir1Abstractlearning,andtensordecomposition,donothavespuriouslocalminimaundersuitableassumptions,inwhichcaselo-Weconsidert...
NotAllSamplesAreCreatedEqual:DeepLearningwithImportanceSamplingAngelosKatharopoulos12Franc¸oisFleuret12Abstractmodel.Tothisend,weproposeanovelimportancesamplingschemethatacceleratesthetrainingofan...
Strongly-TypedAgentsAreGuaranteedtoInteractSafelyDavidBalduzzi1Abstractbeadaptedtoanotherwithoutadverseeffects?Theprob-lemsfallundermechanismdesign,abranchofgametheoryAsartificialagentsproliferate,...
HowCloseAretheEigenvectorsoftheSampleandActualCovarianceMatrices?AndreasLoukas1Abstractbetterwhenonlyasubsetofthespectrumisofinterest.Concretely,ourobjectiveistocharacterizehowmanysam-Howmanysample...