[PDF] Blind Source Separation and Independent Component Analysis

《[PDF] Blind Source Separation and Independent Component Analysis》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、NeuralInformationProcessing-LettersandReviewsVol.6,No.1,January2005REVIEWBlindSourceSeparationandIndependentComponentAnalysis:AReviewSeungjinChoiDepartmentofComputerSciencePohangUniversityofScienceandTechnologySan31,Hyoja-dong,Nam-gu,Pohang,Gyungbuk790-784,KoreaE-mail:seungjin@postech.ac

2、.krAndrzejCichockiRIKEN,BrainScienceInstitute,2-1Hirosawa,Wako,Saitama351-0198,JapanWarsawUniversityofTechnology,PolandE-mail:cia@bsp.brian.riken.go.jpHyung-MinParkandSoo-YoungLeeDepartmentofBioSystems,DepartmentofElectricalEngineeringandComputerScience,andCHUNGMoonSoulCenterforBioInform

3、ationandBioElectronics,KoreaAdvancedInstituteofScienceandTechnology373-1Guseong-dong,Yuseong-gu,Daejeon305-701,KoreaE-mail:fhmpark,syleeg@kaist.ac.kr(SubmittedonOctober20,2004)Abstract-Blindsourceseparation(BSS)andindependentcomponentanalysis(ICA)aregenerallybasedonawideclassofunsupervis

4、edlearningalgorithmsandtheyfoundpotentialapplicationsinmanyareasfromengineeringtoneuroscience.ArecenttrendinBSSistoconsiderproblemsintheframeworkofmatrixfactorizationormoregeneralsignalsdecompositionwithprobabilisticgenerativeandtreestructuredgraphicalmodelsandexploitaprioriknowledgeabou

5、ttruenatureandstructureoflatent(hidden)variablesorsourcessuchasspatio-temporaldecorrelation,statisticalindependence,sparseness,smoothnessorlowestcomplexityinthesensee.g.,ofbestpredictability.Thepossiblegoalofsuchdecompositioncanbeconsideredastheestimationofsourcesnotnecessarystatisticall

6、yindependentandparametersofamixingsystemormoregenerallyasfindinganewreducedorhierarchicalandstructuredrepresentationfortheobserved(sensor)datathatcanbeinterpretedasphysicallymeaningfulcodingorblindsourceestimation.Thekeyissueistofindasuchtransformationorcoding(linearornonlinear)whichhastru

7、ephysicalmeaningandinterpretation.WepresentareviewofBSSandICA,includingvariousalgorithmsforstaticanddynamicmodelsandtheirapplications.Thepapermainlyconsistsofthreeparts:(1)BSSalgorithmsforstaticmodels(instantaneousmixtures);(2)extensionofBSSandICAincorporatingwithsparsene