ELM-Randomness-Kernel

《ELM-Randomness-Kernel》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、Author'spersonalcopyCognComput(2014)6:376–390DOI10.1007/s12559-014-9255-2AnInsightintoExtremeLearningMachines:RandomNeurons,RandomFeaturesandKernelsGuang-BinHuangReceived:25January2014/Accepted:13March2014/Publishedonline:3April2014ÓSpringerScience+BusinessMediaNewYork2014AbstractExtremelearningmach

2、ines(ELMs)basicallyroleinbothresearchandclassificationrelatedapplicationsgiveanswerstotwofundamentallearningproblems:(1)inthepasttwodecades.SVMachieveshighergeneral-Canfundamentalsoflearning(i.e.,featurelearning,clus-izationperformancethanconventionalartificialneuraltering,regressionandclassification)b

3、emadewithoutnetworksinmostclassificationapplications.Intheoriginaltuninghiddenneurons(includingbiologicalneurons)evenimplementationofSVM,onehastohandleaquadraticwhentheoutputshapesandfunctionmodelingoftheseprogramming(QP)problemwhichisusuallytediousandneuronsareunknown?(2)Doesthereexistunifiedframe-ti

4、meconsuming.AsoneofthemainvariantofSVM,Leastworkforfeedforwardneuralnetworksandfeaturespacesquaresupportvectormachine(LS-SVM)[2]aimstoavoidmethods?ELMsthathavebuiltsometangiblelinkstheQPproblemusingequalityconstraintsinsteadofthebetweenmachinelearningtechniquesandbiologicalinequalityconstraintadopte

5、dinconventionalSVM.Com-learningmechanismshaverecentlyattractedincreasingparedwithSVM,LS-SVMiseaseofimplementation.attentionofresearchersinwidespreadresearchareas.ThisExtremelearningmachines(ELMs)[3–7]becomeattrac-paperprovidesaninsightintoELMsinthreeaspects,viz:tivetomoreandmoreresearchersrecently[8

6、–20].Thisrandomneurons,randomfeaturesandkernels.ThispaperpaperaimstoreviewtheELMfromrandomneuronsandalsoshowsthatintheoryELMs(withthesamekernels)kernelspointofviewandtobuildsomerelationshipandtendtooutperformsupportvectormachineanditsvariantslinksbetweenELM,SVMandotherrelatedmachineinbothregressiona

7、ndclassificationapplicationswithlearningtechniques.Althoughitisoutofquestionthatmucheasierimplementation.SVManditsvariantsachievesurprisingperformanceinmostapplications,differentfromsomecommonconceptin