资源描述:
《[NIPS 2011] Learning Anchor Planes for Classification》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、LearningAnchorPlanesforClassificationyzyyxZimingZhangL’uborLadickýPhilipH.S.TorrAmirSaffariyDepartmentofComputing,OxfordBrookesUniversity,Wheatley,Oxford,OX331HX,U.K.zDepartmentofEngineeringScience,UniversityofOxford,ParksRoad,Oxford,OX13PJ,U.K.xSonyComputerEntert
2、ainmentEurope,London,UK{ziming.zhang,philiptorr}@brookes.ac.uklubor@robots.ox.ac.ukamir@ymer.orgAbstractLocalCoordinateCoding(LCC)[18]isamethodformodelingfunctionsofdatalyingonnon-linearmanifolds.Itprovidesasetofanchorpointswhichformalocalcoordinatesystem,suchtha
3、teachdatapointonthemanifoldcanbeapproximatedbyalinearcombinationofitsanchorpoints,andthelinearweightsbecomethelocalcoordinatecoding.Inthispaperweproposeencodingdatausingorthogonalanchorplanes,ratherthananchorpoints.Ourmethodneedsonlyafeworthogonalanchorplanesforc
4、oding,anditcanlinearizeany(;;p)-Lipschitzsmoothnon-linearfunctionwithafixedexpectedvalueoftheupper-boundapproximationerroronanyhighdimensionaldata.Inpractice,theorthogonalcoordinatesystemcanbeeasilylearnedbyminimizingthisupperboundusingsingularvaluedecomposition
5、(SVD).WeapplyourmethodtomodelthecoordinateslocallyinlinearSVMsforclassificationtasks,andourexperimentonMNISTshowsthatusingonly50anchorplanesourmethodachieves1.72%errorrate,whileLCCachieves1.90%errorrateusing4096anchorpoints.1IntroductionLocalCoordinateCoding(LCC)[
6、18]isacodingschemethatencodesthedatalocallysothatanynon-linear(;;p)-Lipschitzsmoothfunction(seeDefinition1inSection2fordetails)overthedatamanifoldcanbeapproximatedusinglinearfunctions.Therearetwocomponentsinthismethod:(1)asetofanchorpointswhichdecidethelocalcoor
7、dinates,and(2)thecodingforeachdatabasedonthelocalcoordinatesgiventheanchorpoints.Theoretically[18]suggeststhatundercertainassumptions,localityismoreessentialthansparsityfornon-linearfunctionapproximation.LCChasbeensuccessfullyappliedtomanyapplicationssuchlikeobje
8、ctrecognition(e.g.locality-constraintlinearcoding(LLC)[16])inVOC2009challenge[7].OnebigissueinLCCisthatitsclassificationperformanceishighlydependentonthenumbero