课堂展示简约PPT模板课件.ppt

《课堂展示简约PPT模板课件.ppt》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、Learningthepartsofobjectsbynon-negativematrixfactorizationContents1Threemethodslearntorepresentaface2Thefeatureofthreedifferentmethods3NMFalgorithmappliedtosemanticanalysis4Dosanddon’tsProfessionalEnglish5OtherusesofNMF1ThreemethodslearntorepresentafaceProfessionalEnglis

2、hNMFPCAVQNon-negativematrixfactorizationPrincipalcomponentsanalysisVectorquantization2ThefeatureofthreedifferentmethodsProfessionalEnglishNMFPCAVQItsimagesarelocalizedfeaturesthatcorrespondbetterwithintuitivenotionsofthepartsoffacesThebasisimagesforPCAare‘eigenfaces’,som

3、eofwhichresembledistortedversionsofwholefacesVQdiscoversabasisconsistingofprototypes,eachofwhichisawholefaceFigure1Figure2Figure3ThreemethodsinamatrixfactorizationframeworkThercolumnsofWarecalledbasisimages.EachcolumnofHiscalledanencodingandisinone-to-onecorrespondencewi

4、thafaceinV.ThedifferencesbetweenPCA,VQandNMFarisefromdifferentconstraintsimposedonthematrixfactorsWandHInVQ,eachcolumnofHisconstrainedtobeaunaryvector,withoneelementequaltounityandtheotherelementsequaltozeroPCAconstrainsthecolumnsofWtobeorthonormalandtherowsofHtobeorthon

5、ormaltoeachotherNMFdoesnotallownegativeentriesinthematrixfactorsWandH3NMFalgorithmappliedtootherdomainsApplyingNMFtoacompletelydifferentproblem,thesemanticanalysisoftextdocuments.VimisthenumberoftimestheithwordinthevocabularyappearsinthemthdocumentIneachsemanticfeature,t

6、healgorithmhasgroupedtogethersemanticallyrelatedwordsFigure4AlthoughNMFissuccessfulinlearningfacialpartsandsemantictopics,thissuccessdoesnotimplythatthemethodcanlearnpartsfromanydatabase,suchasimagesofobjectsviewedfromextremelydifferentviewpoints,orhighlyarticulatedobjec

7、ts4Dosanddon’tsAneuralnetworkthatinfersthehiddenfromthevisiblevariablesrequirestheadditionofinhibitoryfeedbackconnections.NMFlearningisthenimplementedthroughplasticityinthesynapticconnections.5OtherusesofNMFThankyou