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-外文翻译》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。

1、英文原文Aninformation-maximisationapproachtoblindseparationandblinddeconvolutionAnthonyJ.BellandTerrenceJ.SejnowskiComputationalNeurobiologyLaboratoryTheSalkInstitute100010N.TorreyPinesRoadLaJolla,California92037AbstractWederiveanewself-organisinglearningalgorithmwhichmaximisestheinf

2、or-mationtransferredinanetworkofnon-linearunits.Thealgorithmdoesnotassumeanyknowledgeoftheinputdistributions,andisdefinedhereforthezero-noiselimit.Undertheseconditions,informationmaximisationhasextrapropertiesnotfoundinthelinearcase(Linsker1989).Thenon-linearitiesinthetransferfun

3、ctionareabletopickuphigher-ordermomentsoftheinputdistributionsandperformsomethingakintotrueredundancyreductionbetweenunitsintheoutputrepresentation.Thisenable-sthenetworktoseparatestatisticallyindependentcomponentsintheinputs:ahigher-ordergeneralisationofPrincipalComponentsAnalys

4、is.Weapplythenetworktothesourceseparation(orcocktailparty)problem,succe-ssfullyseparatingunknownmixturesofuptotenspeakers.Wealsoshowthatavari-antonthenetworkarchitectureisabletoperformblinddeconvolution(cancellationofunknownechoesandreverberationinaspeechsignal).Finally,wederived

5、epend-enciesofinformationtransferontimedelays.Wesuggestthatinformationmaximi-sationprovidesaunifyingframeworkforproblemsin‘blind’signalprocessing.1IntroductionThispaperpresentsaconvergenceoftwolinesofresearch.Thefirst,thedevelopme-ntofinformationtheoreticunsupervisedlearningrules

6、forneuralnetworkshasbeenpioneeredbyLinsker1992,Becker&Hinton1992,Atick&Redlich1993,Plumble-y&Fallside1988andothers.Thesecondistheuse,insignalprocessing,ofhigher-orderstatistics,forseparatingoutmixturesofindependentsources(blindsep-aration)orreversingtheeffectofanunknownfilter(bli

7、nddeconvolution).Methodsexistforsolvingtheseproblems,butitisfairtosaythatmanyofthemareadhoc.Theliteraturedisplaysadiversityofapproachesandjustifications—forhistoricalre-viewssee(Comon1994)and(Haykin1994a).Inthispaper,wesupplyacommontheoreticalframeworkfortheseproblemsthroughtheus

8、eofinformation-theoreticobjectivefunctio