外文翻译--一种解决盲分离和盲信号去卷积的信息最大化方法

《外文翻译--一种解决盲分离和盲信号去卷积的信息最大化方法》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、附录4英文原文Aninformation-maximisationapproachtoblindseparationandblinddeconvolutionAnthonyJ.BellandTerrenceJ.SejnowskiComputationalNeurobiologyLaboratoryTheSalkInstituteN.TorreyPinesRoadLaJolla,California92037AbstractWederiveanewself-organisinglearningal

2、gorithmwhichmaximisestheinfor-mationtransferredinanetworkofnon-linearunits.Thealgorithmdoesnotassumeanyknowledgeoftheinputdistributions,andisdefinedhereforthezero-noiselimit.Undertheseconditions,informationmaximisationhasextrapropertiesnotfoundinthel

3、inearcase(Linsker1989).Thenon-linearitiesinthetransferfunctionareabletopickuphigher-ordermomentsoftheinputdistributionsandperformsomethingakintotrueredundancyreductionbetweenunitsintheoutputrepresentation.Thisenable-sthenetworktoseparatestatistically

4、independentcomponentsintheinputs:ahigher-ordergeneralisationofPrincipalComponentsAnalysis.Weapplythenetworktothesourceseparation(orcocktailparty)problem,succe-ssfullyseparatingunknownmixturesofuptotenspeakers.Wealsoshowthatavari-antonthenetworkarchit

5、ectureisabletoperformblinddeconvolution(cancellationofunknownechoesandreverberationinaspeechsignal).Finally,wederivedepend-enciesofinformationtransferontimedelays.Wesuggestthatinformationmaximi-sationprovidesaunifyingframeworkforproblemsin‘blind’sign

6、alprocessing.附录41IntroductionThispaperpresentsaconvergenceoftwolinesofresearch.Thefirst,thedevelopme-ntofinformationtheoreticunsupervisedlearningrulesforneuralnetworkshasbeenpioneeredbyLinsker1992,Becker&Hinton1992,Atick&Redlich1993,Plumble-y&Fallsid

7、e1988andothers.Thesecondistheuse,insignalprocessing,ofhigher-orderstatistics,forseparatingoutmixturesofindependentsources(blindsep-aration)orreversingtheeffectofanunknownfilter(blinddeconvolution).Methodsexistforsolvingtheseproblems,butitisfairtosayt

8、hatmanyofthemareadhoc.Theliteraturedisplaysadiversityofapproachesandjustifications—forhistoricalre-viewssee(Comon1994)and(Haykin1994a).Inthispaper,wesupplyacommontheoreticalframeworkfortheseproblemsthroughtheuseofinformation-theoreticobjectivefunctio