2010-2016精選論文2015-1503.04069_第1頁(yè)
已閱讀1頁(yè)，還剩17頁(yè)未讀， 繼續(xù)免費(fèi)閱讀

下載本文檔

版權(quán)說(shuō)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、LSTM:ASearchSpaceOdysseyKlausGreffRupeshKumarSrivastavaJanKoutn?kBasR.SteunebrinkJurgenSchhubereSwissAILabIDSIAIstitutoDalleMollediStudisull’IntelligenzaArtificialeUniversit`adellaSvizzeraitaliana(USI)Scuolauniversitaria

2、professionaledellaSvizzeraitaliana(SUPSI)Galleria26928MannoLuganoSwitzerlAbstractSeveralvariantsoftheLongShtTermMemy(LSTM)architecturefrecurrentneuralwkshavebeenproposedsinceitsinceptionin1995.Inrecentyearsthesewkshavebe

3、comethestateoftheartmodelsfavarietyofmachinelearningproblems.ThishasledtoarenewedinterestinunderstingtheroleutilityofvariouscomputationalcomponentsoftypicalLSTMvariants.Inthispaperwepresentthefirstlargescaleanalysisofeig

4、htLSTMvariantsonthreerepresentativetasks:speechrecognitionhwritingrecognitionpolyphonicmusicmodeling.ThehyperparametersofallLSTMvariantsfeachtaskwereoptimizedseparatelyusingromsearchtheirimptancewasassessedusingthepowerf

5、ulfANOVAframewk.Intotalwesummarizetheresultsof5400experimentalruns(≈15yearsofCPUtime)whichmakesourstudythelargestofitskindonLSTMwks.OurresultsshowthatnoneofthevariantscanimproveuponthestardLSTMarchitecturesignificantlyde

6、monstratethefgetgatetheoutputactivationfunctiontobeitsmostcriticalcomponents.Wefurtherobservethatthestudiedhyperparametersarevirtuallyindependentderiveguidelinesftheirefficientadjustment.1.IntroductionRecurrentneuralwksw

7、ithLongShtTermMemy(whichwewillconciselyrefertoasLSTMs)haveemergedasaneffectivescalablemodelfseverallearningproblemsrelatedtosequentialdata.Earliermethodsfattackingtheseproblemswereusuallyhdesignedwkaroundstodealwiththese

8、quentialnatureofdatasuchaslanguageaudiosignals.SinceLSTMsareeffectiveatcapturinglongtermtempaldependencieswithoutsufferingfromtheoptimizationhurdlesthatplaguesimplerecurrentwks(SRNs)(Hochreiter1991Bengioetal.1994)theyhav

9、ebeenusedtoadvancethestateoftheartfmanydifficultproblems.Thisincludeshwritingrecognition(Gravesetal.2009Phametal.2013Doetschetal.2014)generation(Gravesetal.2013)languagemodeling(Zarembaetal.2014)translation(Luongetal.201

10、4)acousticmodelingofspeech(Saketal.2014)speechsynthesis(Fanetal.2014)proteinsecondarystructureprediction(Snderby1997).HoweverLSTMsarenowappliedtomanylearningproblemswhichdiffersignificantlyinscalenaturefromtheproblemstha

11、ttheseimprovementswereinitiallytestedon.AsystematicstudyoftheutilityofvariouscomputationalcomponentswhichcompriseLSTMs(seeFigure1)wasmissing.ThispaperfillsthatgapsystematicallyaddressestheopenquestionofimprovingtheLSTMar

12、chitecture.WeevaluatethemostpopularLSTMarchitecture(vanillaLSTMSection2)eightdifferentvariantsthereofonthreebenchmarkproblems:acousticmodelinghwritarXiv:1503.04069v1[cs.NE]13Mar2015LSTM:ASearchSpaceOdysseynectionswasd.Th

13、usthatstudydidnotusetheexactgradientftraining.Anotherfeatureofthatversionwastheuseoffullgaterecurrencewhichmeansthatallthegatesreceivedrecurrentinputsfromallgatesattheprevioustimestepinadditiontotherecurrentinputsfromthe

14、blockoutputs.Thisfeaturedidnotappearinanyofthelaterpapers.3.2.FgetGateThefirstpapertosuggestamodificationoftheLSTMarchitectureintroducedthefgetgate(Gersetal.1999)enablingtheLSTMtoresetitsownstate.Thisallowedlearningofcon

15、tinualtaskssuchasembeddedRebergrammar.3.3.PeepholeConnectionsGers&Schhuber(2000)arguedthatindertolearnprecisetimingsthecellneedstocontrolthegates.Sofarthiswasonlypossiblethroughanopenoutputgate.Peepholeconnections(connec

16、tionsfromthecelltothegatesblueinFigure1)wereaddedtothearchitectureindertomakeprecisetimingseasiertolearn.AdditionallytheoutputactivationfunctionwasomittedastherewasnoevidencethatitwasessentialfsolvingtheproblemsthatLSTMh

17、adbeentestedonsofar.3.4.FullGradientThefinalmodificationtowardsthevanillaLSTMwasdonebyGraves&Schhuber(2005).Thisstudypresentedthefullbackpropagationthroughtime(BPTT)trainingfLSTMwkswiththearchitecturedescribedinSection2p

18、resentedresultsontheTIMITbenchmark.UsingfullBPTThadtheaddedadvantagethatLSTMgradientscouldbecheckedusingfinitedifferencesmakingpracticalimplementationsmereliable.3.5.OtherVariantsSinceitsintroductionthevanillaLSTMhasbeen

19、themostcommonlyusedarchitecturebutothervariantshavebeensuggestedtoo.BefetheintroductionoffullBPTTtrainingGersetal.(2002)utilizedatrainingmethodbasedonExtendedKalmanFilteringwhichenabledtheLSTMtobetrainedonsomepathologica

20、lcasesatthecostofhighcomputationalcomplexity.Schhuberetal.(2007)proposedusingahybridevolutionbasedmethodinsteadofBPTTftrainingbutretainedthevanillaLSTMarchitecture.Bayeretal.(2009)evolveddifferentLSTMblockarchitecturesth

21、atmaximizefitnessoncontextsensitivegrammars.Saketal.(2014)introducedalinearprojectionlayerthatprojectstheoutputoftheLSTMlayerdownbeferecurrentfwardconnectionsindertoreducetheamountofparametersfLSTMwkswithmanyblocks.Byint

22、roducingatrainablescalingparameterftheslopeofthegateactivationfunctionsDoetschetal.(2014)wereabletoimprovetheperfmanceofLSTMonanofflinehwritingrecognitiondataset.InwhattheycallDynamicCtexMemyOtteetal.(2014)improvedconver

23、gencespeedofLSTMbyaddingrecurrentconnectionsbetweenthegatesofasingleblock(butnotbetweentheblocks).Choetal.(2014)proposedasimplificationoftheLSTMarchitecturecalledGatedRecurrentUnit(GRU).Theyusedneitherpeepholeconnections

24、noutputactivationfunctionscoupledtheinputthefgetgateintoanupdategate.Finallytheiroutputgate(calledresetgate)onlygatestherecurrentconnectionstotheblockinput(Wz).Chungetal.(2014)perfmedaninitialcomparisonbetweenGRULSTMrept

25、edmixedresults.4.EvaluationSetupThefocusofourstudyistocomparedifferentLSTMvariantsnottoachievestateoftheartresults.Therefeourexperimentsaredesignedtokeepthesetupsimplethecomparisonsfair.ThevanillaLSTMisusedasabaselineeva

26、luatedtogetherwitheightofitsvariants.Eachvariantaddsremovesmodifiesthebaselineinexactlyoneaspectwhichallowstoisolatetheireffect.Threedifferentdatasetsfromdifferentdomainsareusedtoaccountfcrossdomainvariations.Sincehyperp

27、arameterspaceislargeimpossibletotraversecompletelyromsearchwasusedindertoobtainthebestperfminghyperparameters(Bergstra&Bengio2012)feverycombinationofvariantdataset.Thereafterallanalysesfocusedonthe10%bestperfmingtrialsfe

28、achvariantdataset(Section5.1)makingtheresultsrepresentativefthecaseofreasonablehyperparametertuningeffts.Romsearchwasalsochosenftheaddedbenefitofprovidingenoughdatafanalyzingthegeneraleffectofvarioushyperparametersonthep

29、erfmanceofeachLSTMvariant(Section5.2).4.1.DatasetsEachdatasetissplitintothreeparts:atrainingsetavalidationsetwhichisusedfearlystoppingfoptimizingthehyperparametersatestsetfthefinalevaluation.Detailsofpreprocessingfeachda

30、tasetareprovidedinthesupplementarymaterial.4.1.1.TIMITTheTIMITSpeechcpus(Garofoloetal.1993)islargeenoughtobeareasonableacousticmodelingbenchmarkfspeechrecognitionyetitissmallenoughtokeepalargestudysuchasoursmanageable.Ou

溫馨提示

  • 1. 本站所有資源如無(wú)特殊說(shuō)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