遷移學(xué)習(xí)算法研究-莊福振New

遷移學(xué)習(xí)算法研究莊福振中國科學(xué)院計(jì)算技術(shù)研究所2016年4月18日TrainingDataClassifierUnseenData(…,long,T)good!Whatif…2傳統(tǒng)監(jiān)督機(jī)器學(xué)習(xí)(1/2)2023/2/1[fromProf.QiangYang]傳統(tǒng)監(jiān)督機(jī)器學(xué)習(xí)(2/2)32023/2/1傳統(tǒng)監(jiān)督學(xué)習(xí)同源、獨(dú)立同分布兩個(gè)基本假設(shè)標(biāo)注足夠多的訓(xùn)練樣本在實(shí)際應(yīng)用中通常不能滿足！訓(xùn)練集測(cè)試集分類器訓(xùn)練集測(cè)試集分類器遷移學(xué)習(xí)42023/2/1實(shí)際應(yīng)用學(xué)習(xí)場(chǎng)景HP新聞Lenovo新聞不同源、分布不一致人工標(biāo)記訓(xùn)練樣本，費(fèi)時(shí)耗力遷移學(xué)習(xí)運(yùn)用已有的知識(shí)對(duì)不同但相關(guān)領(lǐng)域問題進(jìn)行求解的一種新的機(jī)器學(xué)習(xí)方法放寬了傳統(tǒng)機(jī)器學(xué)習(xí)的兩個(gè)基本假設(shè)遷移學(xué)習(xí)場(chǎng)景(1/4)52023/2/1遷移學(xué)習(xí)場(chǎng)景無處不在遷移知識(shí)遷移知識(shí)圖像分類HP新聞Lenovo新聞新聞網(wǎng)頁分類異構(gòu)特征空間6Theappleisthepomaceousfruitoftheappletree,speciesMalusdomesticaintherosefamilyRosaceae...BananaisthecommonnameforatypeoffruitandalsotheherbaceousplantsofthegenusMusawhichproducethiscommonlyeatenfruit...Training:TextFuture:ImagesApplesBananas遷移學(xué)習(xí)場(chǎng)景(2/4)2023/2/1[fromProf.QiangYang]XinJin,FuzhenZhuang,SinnoJialinPan,ChangyingDu,PingLuo,QingHe:HeterogeneousMulti-taskSemanticFeatureLearningforClassification.CIKM2015:1847-1850.TestTestTrainingTrainingClassifierClassifier72.65%DVDElectronicsElectronics84.60%ElectronicsDrop!遷移學(xué)習(xí)場(chǎng)景(3/4)72023/2/1[fromProf.QiangYang]8DVDElectronicsBookKitchenClothesVideogameFruitHotelTeaImpractical!遷移學(xué)習(xí)場(chǎng)景(4/4)2023/2/1[fromProf.QiangYang]OutlineConceptLearningforTransferLearningConceptLearningbasedonNon-negativeMatrixTri-factorizationforTransferLearningConceptLearningbasedonProbabilisticLatentSemanticAnalysisforTransferLearningTransferLearningusingAuto-encodersTransferLearningfromMultipleSourceswithAutoencoderRegularizationSupervisedRepresentationLearning:TransferLearningwithDeepAuto-encoders92023/2/1ConceptLearningbasedonNon-negativeMatrixTri-factorizationforTransferLearningConceptLearningforTransferLearning102023/2/1IntroductionManytraditionallearningtechniquesworkwellonlyundertheassumption:Trainingandtestdatafollowthesamedistribution

Training(labeled)ClassifierTest(unlabeled)FromdifferentcompaniesEnterpriseNewsClassification:includingtheclasses“ProductAnnouncement”,“Businessscandal”,“Acquisition”,……Productannouncement:HP'sjust-releasedLaserJetProP1100printerandtheLaserJetProM1130andM1210multifunctionprinters,price…performance

...AnnouncementforLenovoThinkPad

ThinkCentre–price$150offLenovoK300desktopusingcouponcode...LenovoThinkPad

ThinkCentre–price$200offLenovoIdeaPadU450plaptopusing....theirperformanceHPnewsLenovonewsDifferentdistributionFail!11ConceptLearningforTransferLearning2023/2/1Motivation(1/3)ExampleAnalysis

Productannouncement:HP'sjust-releasedLaserJetProP1100printerandtheLaserJetProM1130andM1210multifunctionprinters,price…performance

...AnnouncementforLenovoThinkPad

ThinkCentre–price$150offLenovoK300desktopusingcouponcode...LenovoThinkPad

ThinkCentre–price$200offLenovoIdeaPadU450plaptopusing....theirperformanceHPnewsLenovonewsProductwordconceptLaserJet,printer,price,performanceThinkPad,ThinkCentre,price,performanceRelatedProductannouncementdocumentclass:12Sharesomecommonwords:announcement,price,performance…indicateConceptLearningforTransferLearning2023/2/1Motivation(2/3)ExampleAnalysis:

HPLaserJet,printer,price,performanceetal.LenovoThinkpad,Thinkcentre,price,performanceetal.Thewordsexpressingthesamewordconceptaredomain-dependent

13ProductProductannouncementwordconceptindicatesTheassociationbetweenwordconceptsanddocumentclassesisdomain-independent

ConceptLearningforTransferLearning2023/2/1Motivation(3/3)14Furtherobservations:Differentdomainsmayusesamekeywordstoexpressthesameconcept(denotedasidenticalconcept)Differentdomainsmayalsousedifferentkeywordstoexpressthesameconcept(denotedasalikeconcept)Differentdomainsmayalsohavetheirowndistinctconcepts(denotedasdistinctconcept)TheidenticalandalikeconceptsareusedasthesharedconceptsforknowledgetransferWetrytomodelthesethreekindsofconceptssimultaneouslyfortransferlearningtextclassificationConceptLearningforTransferLearning2023/2/1PreliminaryKnowledgeBasicformulaofmatrixtri-factorization:wheretheinputXistheword-documentco-occurrencematrix

denotesconceptinformation,mayvaryindifferentdomainsFdenotesthedocumentclassificationinformation

indeedistheassociationbetweenwordconceptsanddocumentclasses,mayretainstablecrossdomainsGS15ConceptLearningforTransferLearning2023/2/1Previousmethod-MTrickinSDM2010(1/2)SketchmapofMTrick

SourcedomainXs

FsGsFtGtTargetdomainXtSKnowledgeTransfer16ConceptLearningforTransferLearning2023/2/1Consideringthealikeconcepts MTrick(2/2)OptimizationproblemforMTrickG0isthesupervisioninformationtheassociationSissharedasbridgetotransferknowledge17ConceptLearningforTransferLearningDualTransferLearning(Longetal.,SDM2012),consideringidenticalandalikeconcepts2023/2/1TriplexTransferLearning(TriTL)(1/5)Furtherdividethewordconceptsintothreekinds:

18F1,identicalconcepts;F2,alikeconcepts;F3,distinctconceptsInput:ssourcedomainXr(1≤r≤s)withlabelinformation,ttargetdomainXr(s+1≤r≤s+t)WeproposeTriplexTransferLearningframeworkbasedonmatrixtri-factorization(TriTLforshort)

2023/2/1ConceptLearningforTransferLearningF1,S1andS2

aresharedasthebridgeforknowledgetransferacrossdomainsThesupervisioninformationisintegratedbyGr(1≤r≤s)insourcedomainsTriTL(2/5)OptimizationProblem

192023/2/1ConceptLearningforTransferLearningTriTL(3/5)Wedevelopanalternativelyiterativealgorithmtoderivethesolutionandtheoreticallyanalyzeitsconvergence 202023/2/1ConceptLearningforTransferLearningTriTL(4/5)Classificationontargetdomains When1≤r≤s,Grcontainsthelabelinformation,soweremainitunchangedduringtheiterations

whenxibelongstoclassj,thenGr(i,j)=1,elseGr(i,j)=0Aftertheiteration,weobtaintheoutputGr(s+1≤r≤s+t),thenwecanperformclassificationaccordingtoGr212023/2/1ConceptLearningforTransferLearningTriTL(5/5)AnalysisofAlgorithmConvergence Accordingtothemethodologyofconvergenceanalysisinthetwoworks[Leeetal.,NIPS’01]and[Dingetal.,KDD’06],thefollowingtheoremholds.Theorem(Convergence):Aftereachroundofcalculatingtheiterativeformulas,theobjectivefunctionintheoptimizationproblemwillconvergemonotonically.222023/2/1ConceptLearningforTransferLearning232023/2/1rec.autosrec.motorcyclesrec.baseballrec.hockeysci.cryptsic.electronicssci.medsci.spacecomp.graphicscomp.sys.ibm.pc.hardwarecomp.sys.mac.hardwarecomp.windows.xtalk.politics.misctalk.politics.gunstalk.politics.mideasttalk.religion.miscrecscicomptalkDataPreparation(1/3)20Newsgroups Fourtopcategories,eachtopcategorycontainsfoursub-categories SentimentClassification,fourdomains:books,dvd,electronics,kitchenRandomlyselecttwodomainsassources,andtherestastargets,then6problemscanbeconstructed

ConceptLearningforTransferLearning242023/2/1rec.autosrec.motorcyclesrec.baseballrec.hockeysci.cryptsic.electronicssci.medsci.spacerec+sci-baseballcrypySourcedomainautosspaceTargetdomainFortheclassificationproblemwithonesourcedomainandonetargetdomain,wecanconstruct144()

problemsDataPreparation(2/3)Constructclassificationtasks(TraditionalTL)ConceptLearningforTransferLearning252023/2/1Constructnewtransferlearningproblemsrec.autosrec.motorcyclesrec.baseballrec.hockeysci.cryptsic.electronicssci.medsci.spacerec+sci-baseballcrypyautosspacecomp.graphicscomp.sys.ibm.pc.hardwarecomp.sys.mac.hardwarecomp.windows.xtalk.politics.misctalk.politics.gunstalk.politics.mideasttalk.religion.misccomptalkautosgraphicsMoredistinctconceptsmayexist!DataPreparation(3/3)SourcedomainTargetdomainConceptLearningforTransferLearning262023/2/1ComparedAlgorithmsConceptLearningforTransferLearningTraditionallearningAlgorithmsSupervisedLearning:LogisticRegression(LR)[Davidetal.,00]SupportVectorMachine(SVM)[Joachims,ICML’99]Semi-supervisedLearning:TSVM[Joachims,ICML’99]TransferlearningMethods:CoCC[Daietal.,KDD’07],DTL[Longetal.,SDM’12]Classificationaccuracyisusedastheevaluationmeasure272023/2/1ExperimentalResults(1/3)ConceptLearningforTransferLearningSorttheproblemswiththeaccuracyofLRDegreeoftransferdifficultyeasierGenerally,thelowerofaccuracyofLRcanindicatethehardertotransfer,whilethehigheronesindicatetheeasiertotransferharder282023/2/1ExperimentalResults(2/3)ConceptLearningforTransferLearningComparisonsamongTriTL,DTL,MTrick,CoCC,TSVM,SVMandLRondatasetrecvs.sci(144problems)TriTLcanperformwelleventheaccuracyofLRislowerthan65%292023/2/1ExperimentalResults(3/3)ConceptLearningforTransferLearningResultsonnewtransferlearningproblems,weonlyselecttheproblems,whoseaccuraciesofLRarebetween(50%,55%](Onlyslightlybetterthanrandomclassification,thustheymightbemuchmoredifficult).Weobtain65problemsTriTLalsooutperformsallthebaselinesConclusionsExplicitlydefinethreekindsofwordconcepts,i.e.,identicalconcept,alikeconceptanddistinctconceptProposeageneraltransferlearningframeworkbasedonnonnegativematrixtri-factorization,whichsimultaneouslymodelthethreekindsofconcepts(TriTL)Extensiveexperimentsshowtheeffectivenessoftheproposedapproach,especiallywhenthedistinctconceptsmayexist302023/2/1ConceptLearningforTransferLearningConceptLearningbasedonProbabilisticLatentSemanticAnalysisforTransferLearningConceptLearningforTransferLearning312023/2/1322023/2/1MotivationConceptLearningforTransferLearningProductannouncement:HP'sjust-releasedLaserJetProP1100printerandtheLaserJetProM1130andM1210multifunctionprinters,price…performance

...AnnouncementforLenovoThinkPad

ThinkCentre–price$150offLenovoK300desktopusingcouponcode...LenovoThinkPad

ThinkCentre–price$200offLenovoIdeaPadU450plaptopusing....theirperformanceHPnewsLenovonewsProductwordconceptLaserJet,printer,price,performanceThinkPad,ThinkCentre,price,performanceRelatedProductannouncementdocumentclass:Sharesomecommonwords:announcement,price,performance…indicateRetrospecttheexample

332023/2/1SomenotationsddocumentydocumentclasszwordconceptSomedefinitionse.g.,p(price|Product),p(LaserJet|Product,)wwordrdomaine.g,p(Product|Productannouncement)PreliminaryKnowledge(1/3)ConceptLearningforTransferLearning342023/2/1ConceptLearningforTransferLearningPreliminaryKnowledge(2/3)ProductLaserJet,printer,announcement,price,ThinkPad,ThinkCentre,announcement,priceProductannouncementp(w|z,r1)p(w|z,r2)p(z|y)p(w|z,r1)≠p(w|z,r2)E.g.,p(LaserJet|Product,HP)≠p(LaserJet|Product,Lenovo)p(z|y,r1)=p(z|y,r2)E.g.,p(Product|Productannoucement,HP)=p(Product|Productannoucement,Lenovo)Alikeconcept352023/2/1DualPLSA

(D-PLSA)Jointprobabilityoverallvariablesp(w,d)=p(w|z)p(z|y)p(d|y)p(y)GivendatadomainX,theproblemofmaximumloglikelihoodislogp(X;θ)=logΣz

p(Z,X;θ)

θ

includesalltheparametersp(w|z),p(z|y),p(d|y),p(y).Z

denotesallthelatentvariablesPreliminaryKnowledge(3/3)TheproposedtransferlearningalgorithmbasedonD-PLSA,denotedasHIDCConceptLearningforTransferLearning362023/2/1Identicalconceptp(w|za)p(za|y)AlikeconceptTheextensionandintensionaredomainindependentp(w|zb,r)p(zb|y)HIDC(1/3)Theextensionisdomaindependent,whiletheintensionisdomainindependentConceptLearningforTransferLearning372023/2/1Distinctconceptp(w|zc,r)p(zc|y,r)ThejointprobabilitiesofthesethreegraphicalmodelsHIDC(2/3)TheextensionandintensionarebothdomaindependentConceptLearningforTransferLearning382023/2/1Givens+t

datadomainsX={X1,…,Xs,Xs+1,…,Xs+t},withoutlossofgenerality,thefirstsdomainsaresourcedomains,andthelefttdomainsaretargetdomainsConsiderthethreekindsofconcepts:TheLog

likelihoodfunctionislogp(X;θ)=logΣz

p(Z,X;θ)

θ

includesallparametersp(w|za),p(w|zb,r),p(w|zc,r),p(za|y),p(zb|y),p(zc|y,r),p(d|y,r),p(y|r),p(r).HIDC(3/3)ConceptLearningforTransferLearning392023/2/1UsetheEMalgorithmtoderivethesolutionsEStep:ModelSolution(1/4)ConceptLearningforTransferLearning402023/2/1M

Step：ModelSolution(2/4)ConceptLearningforTransferLearning412023/2/1Semi-supervisedEMalgorithm:whenrisfromsourcedomains,thelabeledinformationp(d|y,r)isknownandp(y|r)

canbeinferedp(d|y,r)=1/ny,r,ifdbelongsyindomainr,ny,risthenumberofdocumentsinclassyindomainr,else

p(d|y,c)=0p(y|r)=ny,r/nr

,nr

isthenumberofdocumentsindomainr

whenrisfromsourcedomains,p(d|y,r)andp(y|r)keepunchangedduringtheiterations,whichsupervisetheoptimizingprocessModelSolution(3/4)ConceptLearningforTransferLearning422023/2/1ClassificationfortargetdomainsAfterweobtainthefinalsolutionsofp(w|za),p(w|zb,r),p(w|zc,r),p(za|y),p(zb|y),p(zc|y,r),p(d|y,r),p(y|r),p(r)Wecancomputetheconditionalprobabilities:

ThenthefinalpredictionisDuringtheiterations,alldomainssharep(w|za),p(za|y),p(zb|y),

whichactasthebridgeforknowledgetransferModelSolution(4/4)ConceptLearningforTransferLearning432023/2/1BaselinesComparedAlgorithmsSupervisedLearning:LogisticRegression(LG)[Davidetal.,00]SupportVectorMachine(SVM)[Joachims,ICML’99]Semi-supervisedLearning:TSVM[Joachims,ICML’99]TransferLearning:CoCC[Daietal.,KDD’07]CD-PLSA[Zhuangetal.,CIKM’10]DTL[Longetal.,SDM’12]OurMethodsHIDCMeasure:classificationaccuracyConceptLearningforTransferLearning442023/2/1Resultsonnewtransferlearningproblems,weselecttheproblems,whoseaccuraciesofLRarehigherthan50%,then334problemsareobtainedExperimentalResults(1/5)ConceptLearningforTransferLearning452023/2/1Resultsonnewtransferlearningproblems,weselecttheproblems,whoseaccuraciesofLRarehigherthan50%,then334problemsareobtainedExperimentalResults(2/5)ConceptLearningforTransferLearning462023/2/1ExperimentalResults(3/5)ConceptLearningforTransferLearning472023/2/1Sourcedomain:S

(rec.autos,

sci.space)，Targetdomain:T(rec.sport.hockey,talk.politics.mideast)STSTDistinctconceptSTAlikeconceptExperimentalResults(4/5)ConceptLearningforTransferLearning482023/2/1ExperimentalResults(5/5)ConceptLearningforTransferLearningIndeed,theproposedprobabilisticmethodHIDCisalsobetterthanTriTLThismayduetothereasonthatthereismoreclearerprobabilisticexplanationofHIDCp1(z,y)=p2(z,y)orp1(z|y)=p2(z|y)whichisbetter?p(z|y)p(y)492023/2/1[1]FuzhenZhuang,PingLuo,HuiXiong,QingHe,YuhongXiong,ZhongzhiShi:ExploitingAssociationsbetweenWordClustersandDocumentClassesforCross-DomainTextCategorization.SDM2010,pp.13-24.[2]FuzhenZhuang,PingLuo,ZhiyongShen,QingHe,YuhongXiong,ZhongzhiShi,HuiXiong:CollaborativeDual-PLSA:miningdistinctionandcommonalityacrossmultipledomainsfortextclassification.CIKM2010,pp.359-368.[3]FuzhenZhuang,PingLuo,ZhiyongShen,QingHe,YuhongXiong,ZhongzhiShi,HuiXiong:MiningDistinctionandCommonalityacrossMultipleDomainsUsingGenerativeModelforTextClassification.IEEETrans.Knowl.DataEng.24(11):2025-2039(2012).[3]FuzhenZhuang,PingLuo,ChangyingDu,QingHe,ZhongzhiShi:Triplextransferlearning:exploitingbothsharedanddistinctconceptsfortextclassification.WSDM2013,pp.425-434.[4]FuzhenZhuang,PingLuo,PeifengYin,QingHe,ZhongzhiShi.:ConceptLearningforCross-domainTextClassification:aGeneralProbabilisticFramework.IJCAI2013,pp.1960-1966.ReferencesConceptLearningforTransferLearningOutlineConceptLearningforTransferLearningConceptLearningbasedonNon-negativeMatrixTri-factorizationforTransferLearningConceptLearningbasedonProbabilisticLatentSemanticAnalysisforTransferLearningTransferLearningusingAuto-encodersTransferLearningfromMultipleSourceswithAutoencoderRegularizationSupervisedRepresentationLearning:TransferLearningwithDeepAuto-encoders502023/2/1TransferLearningfromMultipleSourceswithAutoencoderRegularization512023/2/1TransferLearningUsingAuto-encoders52Motivation(1/2)TransferlearningbasedonoriginalfeaturespacemayfailtoachievehighperformanceonTargetdomaindataWeconsidertheautoencodertechniquetocollaborativelyfindanewrepresentationofbothsourceandtargetdomaindataElectronicsVideoGames

Compact;easytooperate;verygoodpicture,excited

aboutthequality;lookssharp!Averygood

game!Itisactionpacked

andfullofexcitement.Iamverymuchhooked

onthisgame.522023/2/1TransferLearningUsingAuto-encodersPreviousmethodsoftentransferfromonesourcedomaintoonetargetdomainWeconsidertheconsensusregularizedframeworkforlearningfrommultiplesourcedomainsDVDBookKitchenElectronicsWeproposeatransferlearningframeworkofconsensusregularizationautoencoderstolearnfrommultiplesourcesMotivation(2/2)532023/2/1TransferLearningUsingAuto-encodersAutoencoderNeuralNetworkMinimizingthereconstructionerrortoderivethesolution:whereh,garenonlinearactivationfunction,e.g.,Sigmoidfunction,forencodinganddecoding542023/2/1TransferLearningUsingAuto-encodersConsensusMeasure-(1/3)Example:three-classclassificationproblem,threeclassifierspredictinstancesf1f2f3f1f2f3x1111x2333x3222x4231x5313x6123ConstraintSource1:D1Source2:D2Source3:D3552023/2/1TransferLearningUsingAuto-encodersConsensusMeasure-(2/3)Example:three-classclassificationproblem,predictiononinstancexMinimalentropy,MaximalConsensusMaximalentropy,MinimalConsensusEntropybasedConsensusMeasure(Luoetal.,CIKM’08)θiistheparametervectorofclassifieri,Cistheclasslabelset562023/2/1TransferLearningUsingAuto-encodersConsensusMeasure-(3/3)Forsimplicity,theconsensusmeasureforbinaryclassificationcanberewrittenasInthiswork,weimposetheconsensusregularizationtoautoencoders,andtrytoimprovethelearningperformancefrommultiplesourcedomainssincetheireffectsonmakingthepredictionconsensusaresimilar.572023/2/1TransferLearningUsingAuto-encodersSomeNotations

SourcedomainsGivenrsourcedomains:,i.e.,

,.

ThefirstcorrespondingdatamatrixisTargetdomainThecorrespondingdatamatrixis

Thegoalistotrainaclassifier

ftomakeprecisepredictionson.582023/2/1TransferLearningUsingAuto-encodersFrameworkofCRAThedatafromallsourceandtargetdomainssharethesameencodinganddecodingweightsTheclassifierstrainedfromthenewrepresentationareregularizedtopredictthesameresultsontargetdomaindata592023/2/1TransferLearningUsingAuto-encodersOptimizationProblemofCRATheoptimizationproblem:ReconstructionError602023/2/1TransferLearningUsingAuto-encodersOptimizationProblemofCRATheoptimizationproblem:ConsensusRegularization612023/2/1TransferLearningUsingAuto-encodersOptimizationProblemofCRATheoptimizationproblem:ThetotallossofsourceclassifiersoverthecorrespondingsourcedomaindatawiththehiddenrepresentationWeighdecayterm622023/2/1TransferLearningUsingAuto-encodersTheSolutionofCRAWeusethegradientdescentmethodtoderivethesolutionofallparameters?isthelearningrate.ThetimecomplexityisO(rnmk)Theoutput:theencodinganddecodingparameters,andsourceclassifierswithlatentrepresentation.632023/2/1TransferLearningUsingAuto-encodersTargetClassifierConstructionTwoScheme:Trainthesourceclassifiersbasedonandcombinethemas,whereCombineallthesourcedomaindataasZSandtrainaunifiedclassifierusinganysupervisedlearningalgorithms,e.g.,SVM,LogisticRegression(LR).ThetwoaccuraciesaredenotedasCRAvandCRAu,respectively642023/2/1TransferLearningUsingAuto-encodersDataSets-(1/2)ImageData(fromLuoetal.,CIKM08)(Someexamples)AB

A1A2A3A4B1B2B3B4Threesources:A1B1A2B2A3B3Targetdomain:A4B4Totally,96()3-sourcevs1-targetdomain(3vs1)probleminstancescanbeconstructedfortheexperimentalevaluation652023/2/1TransferLearningUsingAuto-encodersDataSets-(2/2)SentimentClassification(fromBlitzeretal.,ACL07)Four3-sourcevs1-targetdomainclassificationproblemsareconstructedDVDBookKitchenElectronicsTheaccuracyontargetdomaindataisusedastheevaluationmeasureBothSVMandLRareusedtotrainclassifiersonthenewrepresentation662023/2/1TransferLearningUsingAuto-encodersAllComparedAlgorithmsBaselinesSupervisedlearningonoriginalfeatures:SVM

[Joachims,ICML’99],LogisticRegression(LR)[Davidetal.,00]Embeddingmethodbasedonautoencoders(EAER)[Yuetal.,ECML’13]MarginalizedStackedDenoisingAutoencoders

(mSDA)[Chenetal.,ICML’12]TransferComponentAnalysis(TCA)[Panetal.,TNN’11]Transferlearningfrommultiplesources(CCR3)(Luoetal.,CIKM’08)Ourmethod:CRAvandCRAuForthemethodswhichcannothandlemultiplesources,wetraintheclassifiersfromeachsourcedomainandmergeddataofallsources(r+1accuracies).Finally,maximal,meanandminimalvaluesarereported.672023/2/1TransferLearningUsingAuto-encoders68ExperimentalResults-(1/2)TransferLearningwithMultipleSourcesviaConsensusRegularizationAutoencodersFuzhenZhuang,XiaohuCheng,SinnoJialinPan,WenchaoYu,QingHe,andZhongzhiShiResultson96imageclassificationproblems69ExperimentalResults-(2/2)TransferLearningwithMultipleSourcesviaConsensusRegularizationAutoencodersFuzhenZhuang,XiaohuCheng,SinnoJialinPan,WenchaoYu,QingHe,andZhongzhiShiResultson4sentimentclassificationproblemsConclusionsThewellknownrepresentationlearningtechniqueautoencoderisconsidered,andweformalizetheautoencodersandconsensusregularizationintoaunifiedoptimizationframeworkExtensivecomparisonexperimentsonimageandsentimentdataareconductedtoshowtheeffectivenessoftheproposealgorithm702023/2/1TransferLearningUsingAuto-encodersSupervisedRepresentationLearning:TransferLearningwithDeepAuto-encoders712023/2/1TransferLearningUsingAuto-encodersAutoencoderisanunsupervisedfeaturelearningalgorithm,whichcannoteffectivelymakeuseofthelabelinformationLimitationofBasicAutoencoderContributionofThisWorkWeextendAutoencodertomulti-layerstructure,andincorporatethelabelasonelayerMotivation722023/2/1TransferLearningUsingAuto-encoders源領(lǐng)域和目標(biāo)領(lǐng)域共享編碼和解碼權(quán)重利用KL距離對(duì)隱層空間進(jìn)行約束利用多類回歸模型對(duì)類標(biāo)層進(jìn)行約束FrameworkofTLDA(1/5)732023/2/1TransferLearningUsingAuto-encoders目標(biāo)是最小化重構(gòu)誤差：DeepAutoencoderFrameworkofTLDA(2/5)742023/2/1TransferLearningUsingAuto-encodersKL距離KL距離衡量的是兩個(gè)概率分布的差異情況，計(jì)算公式如下：以上KL距離并不滿足傳