《迁移学习算法研究-庄福振New学习资料》由会员分享，可在线阅读，更多相关《迁移学习算法研究-庄福振New学习资料（87页珍藏版）》请在金锄头文库上搜索。
1、迁移学习算法研究 庄福振中国科学院计算技术研究所2016年4月18日 TrainingData Whatif 2 传统监督机器学习 1 2 2020 4 24 fromProf QiangYang 迁移学习 4 2020 4 24 实际应用学习场景 迁移学习 运用已有的知识对不同但相关领域问题进行求解的一种新的机器学习方法放宽了传统机器学习的两个基本假设 迁移学习场景 1 4 5 2020 4 24 迁移学习场景无处不在 异构特征空间 6 Theappleisthepomaceousfruitoftheappletree speciesMalusdomesticaintherosefamily
2、Rosaceae BananaisthecommonnameforatypeoffruitandalsotheherbaceousplantsofthegenusMusawhichproducethiscommonlyeatenfruit Training Text Future Images Apples Bananas 迁移学习场景 2 4 2020 4 24 fromProf QiangYang XinJin FuzhenZhuang SinnoJialinPan ChangyingDu PingLuo QingHe HeterogeneousMulti taskSemanticFeat
3、ureLearningforClassification CIKM2015 1847 1850 Test Test Training Training Classifier Classifier 72 65 DVD Electronics Electronics 84 60 Electronics Drop 迁移学习场景 3 4 7 2020 4 24 fromProf QiangYang 8 DVD Electronics Book Kitchen Clothes Videogame Fruit Hotel Tea Impractical 迁移学习场景 4 4 2020 4 24 fromP
4、rof QiangYang Outline ConceptLearningforTransferLearningConceptLearningbasedonNon negativeMatrixTri factorizationforTransferLearningConceptLearningbasedonProbabilisticLatentSemanticAnalysisforTransferLearningTransferLearningusingAuto encodersTransferLearningfromMultipleSourceswithAutoencoderRegulari
5、zationSupervisedRepresentationLearning TransferLearningwithDeepAuto encoders 9 2020 4 24 ConceptLearningbasedonNon negativeMatrixTri factorizationforTransferLearning ConceptLearningforTransferLearning 10 2020 4 24 Introduction Manytraditionallearningtechniquesworkwellonlyundertheassumption Traininga
6、ndtestdatafollowthesamedistribution Training labeled Classifier Test unlabeled EnterpriseNewsClassification includingtheclasses ProductAnnouncement Businessscandal Acquisition Productannouncement HP sjust releasedLaserJetProP1100printerandtheLaserJetProM1130andM1210multifunctionprinters price perfor
7、mance AnnouncementforLenovoThinkPadThinkCentre price 150offLenovoK300desktopusingcouponcode LenovoThinkPadThinkCentre price 200offLenovoIdeaPadU450plaptopusing theirperformance HPnews Lenovonews Fail 11 ConceptLearningforTransferLearning 2020 4 24 Motivation 1 3 ExampleAnalysis Productannouncement H
8、P sjust releasedLaserJetProP1100printerandtheLaserJetProM1130andM1210multifunctionprinters price performance AnnouncementforLenovoThinkPadThinkCentre price 150offLenovoK300desktopusingcouponcode LenovoThinkPadThinkCentre price 200offLenovoIdeaPadU450plaptopusing theirperformance HPnews Lenovonews Pr
9、oduct wordconcept LaserJet printer price performance ThinkPad ThinkCentre price performance Related Productannouncement documentclass 12 Sharesomecommonwords announcement price performance indicate ConceptLearningforTransferLearning 2020 4 24 Motivation 2 3 ExampleAnalysis Thewordsexpressingthesamew
10、ordconceptaredomain dependent 13 Product Productannouncement wordconcept indicates Theassociationbetweenwordconceptsanddocumentclassesisdomain independent ConceptLearningforTransferLearning 2020 4 24 Motivation 3 3 14 Furtherobservations Differentdomainsmayusesamekeywordstoexpressthesameconcept deno
11、tedasidenticalconcept Differentdomainsmayalsousedifferentkeywordstoexpressthesameconcept denotedasalikeconcept Differentdomainsmayalsohavetheirowndistinctconcepts denotedasdistinctconcept TheidenticalandalikeconceptsareusedasthesharedconceptsforknowledgetransferWetrytomodelthesethreekindsofconceptss
12、imultaneouslyfortransferlearningtextclassification ConceptLearningforTransferLearning 2020 4 24 PreliminaryKnowledge Basicformulaofmatrixtri factorization wheretheinputXistheword documentco occurrencematrix F G S 15 ConceptLearningforTransferLearning 2020 4 24 Previousmethod MTrickinSDM2010 1 2 Sket
13、chmapofMTrick SourcedomainXs Fs Gs Ft Gt TargetdomainXt S KnowledgeTransfer 16 ConceptLearningforTransferLearning 2020 4 24 Consideringthealikeconcepts MTrick 2 2 OptimizationproblemforMTrick G0isthesupervisioninformation theassociationSissharedasbridgetotransferknowledge 17 ConceptLearningforTransf
14、erLearning DualTransferLearning Longetal SDM2012 consideringidenticalandalikeconcepts 2020 4 24 TriplexTransferLearning TriTL 1 5 Furtherdividethewordconceptsintothreekinds 18 F1 identicalconcepts F2 alikeconcepts F3 distinctconcepts Input ssourcedomainXr 1 r s withlabelinformation ttargetdomainXr s
15、 1 r s t WeproposeTriplexTransferLearningframeworkbasedonmatrixtri factorization TriTLforshort 2020 4 24 ConceptLearningforTransferLearning F1 S1andS2aresharedasthebridgeforknowledgetransferacrossdomains ThesupervisioninformationisintegratedbyGr 1 r s insourcedomains TriTL 2 5 OptimizationProblem 19
16、 2020 4 24 ConceptLearningforTransferLearning TriTL 3 5 Wedevelopanalternativelyiterativealgorithmtoderivethesolutionandtheoreticallyanalyzeitsconvergence 20 2020 4 24 ConceptLearningforTransferLearning TriTL 4 5 Classificationontargetdomains When1 r s Grcontainsthelabelinformation soweremainitunchangedduringtheiterationswhenxibelongstoclassj thenGr i j 1 elseGr i j 0Aftertheiteration weobtaintheoutputGr s 1 r s t thenwecanperformclassificationaccordingtoGr 21 2020 4 24 ConceptLearningforTransfe
