Autotagging music with conditional restricted Boltzmann machines有条件限制Boltzmann机器地自动调音音乐.pdf

《Autotagging music with conditional restricted Boltzmann machines有条件限制Boltzmann机器地自动调音音乐.pdf》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、AutotaggingmusicwithconditionalrestrictedBoltzmannmachinesMichaelMandel1RazvanPascanu1HugoLarochelle2YoshuaBengio1Dept.IRO,UniversitedeMontreal1Comp.Sc.Dept.UniversityofToronto2Abstractbutfailsforitemsthatareneworniche,thisistheso-calledcoldstartproblem[2].Thispaperdescribestwoapplicationso

2、fconditionalOnepromisingwaytoovercomethecoldstartisrestrictedBoltzmannmachines(CRBMs)tothetaskthroughcontent-basedanalysisandtaggingoftheofautotaggingmusic.Therstconsistsoftrainingaitemsinthecollection,knownasautotagging.Re-CRBMtopredicttagsthatauserwouldapplytoasearchershaveinvestigatedanumb

3、erofautotaggingclipofasongbasedontagsalreadyappliedbyothertechniquesformusicoverthelastdecade[3,4,5].users.Bylearningtherelationshipsbetweentags,thisWhileafewautotaggingtechniquesattempttocap-modelisabletopre-processtrainingdatatosigni-turetherelationshipbetweentags(e.g.[6]),manycantlyimprove

4、theperformanceofasupportvectortreateachtagasaseparateclassicationorrankingmachine(SVM)autotagging.Thesecondistheuseofproblem(e.g.[7]).TheproblemofpredictingtheadiscriminativeRBM,atypeofCRBM,toautotagpresenceorrelevanceofmultipletagssimultaneouslymusic.Bysimultaneouslyexploitingtherelationship

5、sisknownasthemulti-labelclassicationproblem[8].amongtagsandbetweentagsandaudio-basedfea-Thispaperexplorestechniquesforautotaggingmu-tures,thismodelisabletosignicantlyoutperformsicthatincorporatetherelationshipsbetweentags.SVMs,logisticregression,andmulti-layerperceptrons.Weapproachthisproble

6、mintwoways,bothofwhichInordertobeappliedtothisproblem,thediscrimina-arebasedonconditionalrestrictedBoltzmannma-tiveRBMwasgeneralizedtothemulti-labelsettingchines(RBMs)describedinSection2.Therstap-andfourdierentlearningalgorithmsforitwereeval-proach,describedinSection2.1,isanovelmodeluated,th

7、erstsuchin-depthanalysisofwhichwearetrainedtopredictthetagsthatauserwillapplytoaware.musicbasedonthetagsotherusershaveappliedtoit.Itisapurelytextualmodelinthatitdoesnotutilizetheaudioatalltomakepredictions.Thesepredicted1Introductionta