文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > 工程科技 > 信息与通信 > Independent component analysis_ algorithms and applications
侵权投诉

Independent component analysis_ algorithms and applications

卷积神经网络和一些独立成分分析的外文文献

doc.guandang.net

Invitedarticle

Independentcomponentanalysis:algorithmsandapplications

¨rinen,E.Oja*A.Hyva

NeuralNetworksResearchCentre,HelsinkiUniversityofTechnology,P.O.Box5400,FIN-02015HUT,Helsinki,Finland

Received28March2000;accepted28March2000

Abstract

Afundamentalprobleminneuralnetworkresearch,aswellasinmanyotherdisciplines,is ndingasuitablerepresentationofmultivariate

data,i.e.randomvectors.Forreasonsofcomputationalandconceptualsimplicity,therepresentationisoftensoughtasalineartransformationoftheoriginaldata.Inotherwords,eachcomponentoftherepresentationisalinearcombinationoftheoriginalvariables.Well-knownlineartransformationmethodsincludeprincipalcomponentanalysis,factoranalysis,andprojectionpursuit.Independentcomponentanalysis(ICA)isarecentlydevelopedmethodinwhichthegoalisto ndalinearrepresentationofnon-Gaussiandatasothatthecomponentsarestatisticallyindependent,orasindependentaspossible.Sucharepresentationseemstocapturetheessentialstructureofthedatainmanyapplications,includingfeatureextractionandsignalseparation.Inthispaper,wepresentthebasictheoryandapplicationsofICA,andourrecentworkonthesubject. 2000PublishedbyElsevierScienceLtd.

Keywords:Independentcomponentanalysis;Projectionpursuit;Blindsignalseparation;Sourceseparation;Factoranalysis;Representation

1.Motivation

Imaginethatyouareinaroomwheretwopeoplearespeakingsimultaneously.Youhavetwomicrophones,whichyouholdindifferentlocations.Themicrophonesgiveyoutworecordedtimesignals,whichwecoulddenotebyx1(t)andx2(t),withx1andx2theamplitudes,andtthetimeindex.Eachoftheserecordedsignalsisaweightedsumofthespeechsignalsemittedbythetwospeakers,whichwedenotebys1(t)ands2(t).Wecouldexpressthisasalinearequation:

x1 t a11s1 a12s2x2 t a21s1 a22s2

1 2

wherea11,a12,a21,anda22aresomeparametersthatdependonthedistancesofthemicrophonesfromthespeakers.Itwouldbeveryusefulifyoucouldnowestimatethetwooriginalspeechsignalss1(t)ands2(t),usingonlytherecordedsignalsx1(t)andx2(t).Thisiscalledthecocktail-partyproblem.Forthetimebeing,weomitanytimedelaysorotherextrafactorsfromoursimpli edmixingmodel.Asanillustration,considerthewaveformsinFigs.1and2.Theseare,ofcourse,notrealisticspeechsignals,butsuf ceforthisillustration.Theoriginalspeechsignals

*Correspondingauthor.Tel.: 358-9-4513265;fax: 358-9-4513277.

¨rinen),erkki.oja@E-mailaddresses:aapo.hyvarinen@hut. (A.Hyva

hut. (E.Oja).

0893-6080/00/$-seefrontmatter 2000PublishedbyElsevierScienceLtd.PII:S0893-6080(00)00026-5

couldlooksomethinglikethoseinFig.1andthemixedsignalscouldlooklikethoseinFig.2.TheproblemistorecoverthedatainFig.1usingonlythedatainFig.2.Actually,ifweknewtheparametersaij,wecouldsolvethelinearequationin(1)byclassicalmethods.Thepointis,however,thatifyoudonotknowtheaij,theproblemisconsiderablymoredif cult.

Oneapproachtosolvingthisproblemwouldbetousesomeinformationonthestatisticalpropertiesofthesignalssi(t)toestimatetheaii.Actually,andperhapssurprisingly,itturnsoutthatitisenoughtoassumethats1(t)ands2(t),ateachtimeinstantt,arestatisticallyindependent.Thisisnotanunrealisticassumptioninmanycases,anditneednotbeexactlytrueinpractice.TherecentlydevelopedtechniqueofIndependentComponentAnalysis,orICA,canbeusedtoestimatetheaijbasedontheinformationoftheirindepen-dence,whichallowsustoseparatethetwooriginalsourcesignalss1(t)ands2(t)fromtheirmixturesx1(t)andx2(t).Fig.3givesthetwosignalsestimatedbytheICAmethod.Ascanbeseen,theseareveryclosetotheoriginalsourcesignals(theirsignsarereversed,butthishasnosigni cance).

Independentcomponentanalysiswasoriginallydevel-opedtodealwithproblemsthatarecloselyrelatedtothecocktail-partyproblem.SincetherecentincreaseofinterestinICA,ithasbecomeclearthatthisprinciplehasalotofotherinterestingapplicationsaswell.

Consider,forexample,electricalrecordingsofbrainactivityasgivenbyanelectroencephalogram(EEG).TheEEGdataconsistsofrecordingsofelectricalpotentialsin

Independent component analysis_ algorithms and applications

Independent component analysis_ algorithms and applications

第1页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email