MaximumLikelihoodwithBias-CorrectedCalibrationisHard-To-BeatatLabelShiftAdaptationAmrM.Alexandari1AnshulKundaje12Avantishrikumar1Abstract1.IntroductionLabelshiftreferstothephenomenonwheretheImagine...
BeyondSignalPropagation:isFeatureDiversityNecessaryinDeepNeuralNetworkInitialization?YanivBlumenfeld1DarGilboa2DanielSoudry1Abstractadetailedanalysisoftheirdynamicsandgeneralizationproperties.Unlik...
Recht–Re´NoncommutativeArithmetic-GeometricMeanConjectureisFalseZehuaLai1Lek-HengLim1Abstractinmachinelearningcomputations,stochasticvariantsofgradientdescent(Bottou,2010;Johnson&Zhang,2013;Stoch...
UnreproducibleResearchisReproducibleXavierBouthillier1Ce´sarLaurent1PascalVincent123Abstractvatedbypastevidenceoflackofscientificrigour,researcherbiases,andfraud(Eisner,2018).Theapparentcontradict...
WhatistheEffectofImportanceWeightinginDeepLearning?JonathonByrd1ZacharyC.Lipton1AbstractEq[f(x)],importancesamplingproducesanunbiasedesti-matebyweightingeachsamplexaccordingtothelikelihoodImportanc...
isGeneratorConditioningCausallyRelatedtoGANPerformance?AugustusOdena1JacobBuckman1CatherineOlsson1TomB.Brown1ChristopherOlah1ColinRaffel1IanGoodfellow1Abstract&Bottou,2017).Themostnotableofthesepat...
GreedisStillGood:MaximizingMonotoneSubmodular+Supermodular(BP)FunctionsWenruoBai1JeffreyA.Bilmes12Abstractoptimalchoiceateachstageinthehopeoffindingagoodglobalsolution.Itisoneofthesimplest,mostwide...
WhyisPosteriorSamplingBetterthanOptimismforReinforcementLearning?IanOsband12BenjaminVanRoy1Abstractmateoffuturevalueandselectstheactionwiththegreatestestimate.Ifaselectedactionisnotnear-optimal,the...
ImprovingViterbiisHard:BetterRuntimesImplyFasterCliqueAlgorithmsArtursBackurs1ChristosTzamos1AbstractO(Tn2)timeforanyHMMwithnstatesandanobserva-tionsequenceoflengthT.ThisalgorithmisknownastheThecla...