DeepReDuce:ReLUReductionforFastPrivateInferenceNandanKumarJha1ZahraGhodsi1SiddharthGarg1BrandonReagen1AbstractTest(top-1)accuracy(%)80DeepReDuceTherecentriseofprivacyconcernshasledre-75SAFENetsearc...
RobustOne-BitRecoveryviaReLUGenerativeNetworks:Near-OptimalStatisticalRateandGlobalLandscapeAnalysisShuangQiu1XiaohanWei2ZhuoranYang3Abstract1.IntroductionWestudytherobustone-bitcompressedsens-Quan...
Reverse-EngineeringDeepReLUNetworksDavidRolnick1KonradP.Ko¨rding1Abstractconstructadversarialattacks.TheoutputofaneuralnetworkdependsonitsarTherelatedquestionofreverse-engineeringbiologicalneu...
OptimizationTheoryforReLUNeuralNetworksTrainedwithNormalizationLayersYonatanDukler1QuanquanGu2GuidoMontúfar134AbstractofadditionalnormalizationmethodsfollowedBN,notablyincludingLayerNormalization(...
ConstructiveUniversalHigh-DimensionalDistributionGenerationthroughDeepReLUNetworksDmytroPerekrestenko1StephanMu¨ller1HelmutBo¨lcskei12Abstractquence.Here,theinputlayersizeisdeterminedbythedimensi...
TheMultilinearStructureofReLUNetworksThomasLaurent1JamesH.vonBrecht2AbstractΩ2Ω3WestudythelosssurfaceofneuralnetworksΩ1Ω4equippedwithahingelosscriterionandReLUorleakyReLUnonlinearities.Anysuchn...