Tiltingtheplayingfield:DynamicallossfunctionsformachinelearningMiguelRuiz-Garcia12GeZhang1SamuelS.Schoenholz3AndreaJ.Liu1Abstracttionscale(Glorot&Bengio,2010;Xiaoetal.,2018)orlearningrateschedule(H...
SignaturedDeepFictitiousPlayforMeanFieldGameswithCommonNoiseMingMin1RuimengHu12Abstractrealisticmodels,theproblemusuallylackstractabilityandneedsnumericalmethods.WithalargenumberofplayersExistingde...
AMean-fieldAnalysisofDeepResNetandBeyond:TowardsProvableOptimizationViaOverparameterizationFromDepthYipingLu1ChaoMa2YulongLu3JianfengLu3LexingYing4Abstract1.IntroductionTrainingdeepneuralnetworkswi...
MeanFieldMulti-AgentReinforcementLearningYaodongYang1RuiLuo1MinneLi1MingZhou2WeinanZhang2JunWang1AbstractInstead,accountingfortheextrainformationfromconjec-turingthepoliciesofotheragentsisbeneficia...
DynamicalIsometryandaMeanFieldTheoryofRNNs:GatingEnablesSignalPropagationinRecurrentNeuralNetworksMinminChen1JeffreyPennington2SamuelS.Schoenholz2Abstracttion(Gravesetal.,2013)andrecommendationsyst...
DynamicalIsometryandaMeanFieldTheoryofCNNs:HowtoTrain10,000-LayerVanillaConvolutionalNeuralNetworksLechaoXiao12YasamanBahri12JaschaSohl-Dickstein1SamuelS.Schoenholz1JeffreyPennington1AbstractFigure...