欢迎来到天天文库
浏览记录
ID:30814474
大小:117.00 KB
页数:13页
时间:2019-01-03
《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,更多相关内容在工程资料-天天文库。
1、英文原文Aninformation・maximisationapproachtoblindseparationandblinddeconvolutionAnthonyJ.BellandTerrenceJ.SejnowskiComputationalNeurobiologyLaboratoryTheSalkInstitute100010N.TorreyPinesRoadLaJolla,California92037AbstractWederiveanewself-organisinglearningalgorithmwhichmaximisestheinfor-mationtransferre
2、dinanetworkofnon-linearunits.Thealgorithmdoesnotassumeanyknowledgeoftheinputdistributions,andisdefinedhereforthezero-noiselimit.Undertheseconditions,informationmaximisationhasextrapropertiesnotfoundinthelinearcase(Linsker1989).Thenon-linearitiesinthetransferfunctionareabletopickuphigher-ordermoment
3、softheinputdistributionsandperformsomethingakintotrueredundancyreductionbetweenunitsintheoutputrepresentation.Thisenable-sthenetworktoseparatestatisticallyindependentcomponentsintheinputs:ahigher-ordergeneralisationofPrincipalComponentsAnalysis.Weapplythenetworktothesourceseparation(orcocktailparty
4、)problem,succe-ssfullyseparatingunknownmixturesofuptotenspeakers.Wealsoshowthatavari-antonthenetworkarchitectureisabletoperformblinddeconvolution(cancellationofunknownechoesandreverberationinaspeechsignal).Finally,wederivedepend-enciesofinformationtransferontimedelays.Wesuggestthatinformationmaximi
5、-sationprovidesaunifyingframeworkforproblemsin'blind'signalprocessing.1IntroductionThispaperpresentsaconvergenceoftwolinesofresearch.Thefirst,thedevelopme-ntofinformationtheoreticunsupervisedlearningrulesforneuralnetworkshasbeenpioneeredbyLinsker1992,Becker&Hinton1992,Atick&Redlich1993,Plumble-y&Fa
6、llside1988andothers.Thesecondistheuse,insignalprocessing,ofhigher-orderstatistics,forseparatingoutmixturesofindependentsources(blindsep・aration)orreversingtheeffectofanunknownfilter(blinddeconvolution).Methodsexistforsolvingtheseproblems,butitisfairtosaythatmanyofthemareadhoc.Theliteraturedisplaysa
7、diversityofapproachesandjustifications一forhistoricalre-viewssee(Comon1994)and(Haykin1994a).Inthispaper,wesupplyacommontheoreticalframeworkfortheseproblemsthroughtheuseofinformation-theoreticobjectivefunctio
此文档下载收益归作者所有