2010-2016精選論文2015_batch_normalization_accelerating_deep_network_training_by_reducing_internal_covariate_shift_第1頁(yè)
已閱讀1頁(yè)，還剩8頁(yè)未讀， 繼續(xù)免費(fèi)閱讀

下載本文檔

版權(quán)說明：本文檔由用戶提供并上傳，收益歸屬內(nèi)容提供方，若內(nèi)容存在侵權(quán)，請(qǐng)進(jìn)行舉報(bào)或認(rèn)領(lǐng)

文檔簡(jiǎn)介

1、BatchNmalization:AcceleratingDeepwkTrainingbyReducingInternalCovariateShiftSergeyIoffeSIOFFE@ChristianSzegedySZEGEDY@Google1600AmphitheatrePkwyMountainViewCA94043AbstractTrainingDeepNeuralwksiscomplicatedbythefactthatthe

2、distributionofeachlayer’sinputschangesduringtrainingastheparametersofthepreviouslayerschange.Thisslowsdownthetrainingbyrequiringlowerlearningratescarefulparameterinitializationmakesitnotiouslyhardtotrainmodelswithsaturat

3、ingnonlinearities.Werefertothisphenomenonasinternalcovariateshiftaddresstheproblembynmalizinglayerinputs.Ourmethoddrawsitsstrengthfrommakingnmalizationapartofthemodelarchitectureperfmingthenmalizationfeachtrainingminibat

4、ch.BatchNmalizationallowsustousemuchhigherlearningratesbelesscarefulaboutinitializationinsomecaseseliminatestheneedfout.AppliedtoastateoftheartimageclassificationmodelBatchNmalizationachievesthesameaccuracywith14timesfew

5、ertrainingstepsbeatstheiginalmodelbyasignificantmargin.UsinganensembleofbatchnmalizedwksweimproveuponthebestpublishedresultonImageclassification:reaching4.82%top5testerrexceedingtheaccuracyofhumanraters.1.IntroductionDee

6、plearninghasdramaticallyadvancedthestateoftheartinvisionspeechmanyotherareas.Stochasticgradientdescent(SGD)hasprovedtobeaneffectivewayoftrainingdeepwksSGDvariantssuchasmomentum(Sutskeveretal.2013)Adagrad(Duchietal.2011)h

7、avebeenusedtoachievestateoftheartperfmance.SGDoptimizestheparametersΘofthewksoastoProceedingsofthe32ndInternationalConferenceonMachineLearningLilleFrance2015.JMLR:Wigningthelattertermwouldleadtotheexplosiondescribedabove

8、.WithinthisframewkwhiteningthelayerinputsisexpensiveasitrequirescomputingthecovariancematrixCov[x]=Ex∈X[xxT]?E[x]E[x]TitsinversesquareroottoproducethewhitenedactivationsCov[x]?12(x?E[x])aswellasthederivativesofthesetrans

9、fmsfbackpropagation.Thismotivatesustoseekanalternativethatperfmsinputnmalizationinawaythatisdifferentiabledoesnotrequiretheanalysisoftheentiretrainingsetaftereveryparameterupdate.Someofthepreviousapproaches(e.g.(Lyuinthe

10、jointcaseregularizationwouldberequiredsincetheminibatchsizeislikelytobesmallerthanthenumberofactivationsbeingwhitenedresultinginsingularcovariancematrices.ConsideraminibatchBofsizem.Sincethenmalizationisappliedtoeachacti

11、vationindependentlyletusfocusonaparticularactivationx(k)omitkfclarity.WehavemvaluesofthisactivationintheminibatchB=x1...m.Letthenmalizedvaluesbe?x1...mtheirlineartransfmationsbey1...m.WerefertothetransfmBNγβ:x1...m→y1...

12、mastheBatchNmalizingTransfm.WepresenttheBNTransfminAlgithm1.Inthealgithm?isaconstantaddedtotheminibatchvariancefnumericalstability.TheBNtransfmcanbeaddedtoawktomanipulateanyactivation.Inthenotationy=BNγβ(x)weindicatethat

溫馨提示

  • 1. 本站所有資源如無(wú)特殊說明，都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請(qǐng)下載最新的WinRAR軟件解壓。
  • 2. 本站的文檔不包含任何第三方提供的附件圖紙等，如果需要附件，請(qǐng)聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
  • 3. 本站RAR壓縮包中若帶圖紙，網(wǎng)頁(yè)內(nèi)容里面會(huì)有圖紙預(yù)覽，若沒有圖紙預(yù)覽就沒有圖紙。
  • 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
  • 5. 眾賞文庫(kù)僅提供信息存儲(chǔ)空間，僅對(duì)用戶上傳內(nèi)容的表現(xiàn)方式做保護(hù)處理，對(duì)用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯，并不能對(duì)任何下載內(nèi)容負(fù)責(zé)。
  • 6. 下載文件中如有侵權(quán)或不適當(dāng)內(nèi)容，請(qǐng)與我們聯(lián)系，我們立即糾正。
  • 7. 本站不保證下載資源的準(zhǔn)確性、安全性和完整性, 同時(shí)也不承擔(dān)用戶因使用這些下載資源對(duì)自己和他人造成任何形式的傷害或損失。

評(píng)論

0/150

提交評(píng)論