资源描述：
《ChapterEnsembleMLPClassifierDesignSurrey章合奏MLP分类器设计——萨里》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、Chapter:EnsembleMLPClassifierDesignTerryWindeattCentreforVision,SpeechandSignalProc.,DepartmentofElectronicEngineering,UniversityofSurrey,Guildford,Surrey,UnitedKingdomGU27XHt.windeatt@surrey.ac.ukAbstract.Multi-layerperceptrons(MLP)makepowerfulclassifiersthatmayprovidesuperiorperformancecomparedw
2、ithotherclassifiers,butareoftencriticizedforthenumberoffreeparameters.Mostcommonly,parametersaresetwiththehelpofeitheravalidationsetorcross-validationtechniques,butthereisnoguaranteethatapseudo-testsetisrepresentative.FurtherdifficultieswithMLPsincludelongtrainingtimesandlocalminima.Inthischapter,
3、anensembleofMLPclassifiersisproposedtosolvetheseproblems.Parameterselectionforoptimalperformanceisperformedusingmeasuresthatcorrelatewellwithgeneralisationerror.1IntroductionThetopicofthischapterconcernssolvingproblemsinpatternrecognitionusingacombinationofneuralnetworkclassifiers.Patternclassific
4、ationinvolvesassignmentofanobjecttooneofseveralpre-specifiedcategoriesorclasses,andisakeycomponentinmanydatainterpretationactivities.Herewefocusonclassifiersthatlearnfromexamples,anditisassumedthateachexamplepatternisrepresentedbyasetofnumbers,whichareknownasthepatternfeatures.Inthecaseoffacerecog
5、nition(Section5),thesefeaturesconsistofnumbersrepresentingdifferentaspectsoffacialfeatures.Inordertodesignalearningsystemitiscustomarytodividetheexamplepatternsintotwosets,atrainingsettodesigntheclassifierandatestset,whichissubsequentlyusedtopredicttheperformancewhenpreviouslyunseenexamplesareappl
6、ied.Aproblemariseswhentherearemanyfeaturesandrelativelyfewtrainingexamples,andtheclassifiercanlearnthetrainingsettoowell,knownasover-fittingsothatperformanceonthetestsetdegrades.Automatingtheclassificationtasktoachieveoptimalperformancehasbeenstudiedinthetraditionalfieldsofpatternrecognition,machi
7、nelearningandneuralnetworksaswellasnewerdisciplinessuchasdatafusion,dataminingandknowledgediscovery.Traditionally,theapproachthathasbeenusedinthedesignofpatternclassificationsystemsistoexperimentallyassesstheperf
