资源描述：
《Pruning least objective contribution in KMSE》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、Neurocomputing74(2011)3009–3018ContentslistsavailableatScienceDirectNeurocomputingjournalhomepage:www.elsevier.com/locate/neucomPruningleastobjectivecontributioninKMSEa,nbaa,cbYong-PingZhao,Jian-GuoSun,Zhong-HuaDu,Zhi-AnZhang,Hai-BoZhangaZNDYofMinist
2、erialKeyLaboratory,NanjingUniversityofScienceandTechnology,Nanjing210094,PRChinabCollegeofAutomationandEngineering,NanjingUniversityofAeronauticsandAstronautics,Nanjing210016,PRChinacCollegeofEnergyandPowerEngineering,NanjingUniversityofAeronauticsan
3、dAstronautics,Nanjing210016,PRChinaarticleinfoabstractArticlehistory:Althoughkernelminimumsquarederror(KMSE)iscomputationallysimple,i.e.,itonlyneedssolvingaReceived10October2010linearequationset,itsuffersfromthedrawbackthatinthetestingphasethecomputa
4、tionalefficiencyReceivedinrevisedformdecreasesseriouslyasthetrainingsamplesincrease.TheunderlyingreasonisthatthesolutionofNaı¨ve18February2011KMSEisrepresentedbyallthetrainingsamplesinthefeaturespace.Hence,inthispaper,amethodofAccepted7April2011select
5、ingsignificantnodesforKMSEisproposed.Duringeachcalculationround,thepresentedCommunicatedbyS.Choialgorithmprunesthetrainingsamplemakingleastcontributiontotheobjectivefunction,hencecalledAvailableonline23May2011asPLOC-KMSE.Toacceleratethetrainingprocedu
6、re,abatchofso-callednonsignificantnodesisprunedKeywords:insteadofonebyoneinPLOC-KMSE,andthisspeedupalgorithmisnamedMPLOC-KMSEforshort.ToKernellearningshowtheefficacyandfeasibilityoftheproposedPLOC-KMSEandMPLOC-KMSE,theexperimentsonMinimumsquarederrorbe
7、nchmarkdatasetsandreal-worldinstancesarereported.TheexperimentalresultsdemonstratethatRegressionPLOC-KMSEandMPLOC-KMSErequirethefewestsignificantnodescomparedwithotheralgorithms.ClassificationSignificantnodesThatistosay,theircomputationalefficiencyinthet
8、estingphaseisbest,thussuitableforenvironmentshavingastrictdemandofcomputationalefficiency.Inaddition,fromtheperformedexperiments,itiseasilyknownthattheproposedMPLOC-KMSEacceleratesthetrainingprocedurewithoutsacrificingthecomputationalefficiencyoftesting
