Nonlinear component analysis as a kernel eigenvalue problem-英文文献.pdfVIP

Max-Planck-Institut für biologische Kybernetik Arbeitsgruppe Bülthoff Spemannstraße 38 • 72076 Tübingen • Germany Technical Rep ort No Decemb er Nonlinear Comp onent Analysis as a Kernel Eigenvalue Problem Bernhard Scholkopf Alexander Smola and KlausRob ert Muller Abstract Wedescribe anewmethodforperforminganonlinearformofPrincipalComponentAnal ysis Bytheuse ofintegraloperator kernelfunctionswecan ecientlycomputeprincipal componentsinhighdimensionalfeature spaces related to input space by somenonlinear mapforinstancethespace ofallpossiblepixelproductsinimages Wegivethe derivationofthe method alongwith adiscussion of other techniques whichcan be made nonlinearwiththekernelapproachandpresentrstexperimentalresultsonnonlinearfea tureextractionforpatternrecognition ASandKRMarewithGMDFirstForschungszentrumInformationstechnikRudowerChaussee Berlin ASandBSweresupported bygrantsfromtheStudienstiftungdesdeutschenVolkes BSthanks the GMD First for hospitality during two visits AS and BS thank V Vapnik for introducing them to kernel representations of dot products during joint work on Support Vector machines This work protedfromdiscussionswithVBlanzLBottouCBurgesHBulthoKGegenfurtnerPHaner NMurataPSimardSSollaVVapnikandTVetter WearegratefultoVBlanzCBurges and SSollaforreadingapreliminaryversionofthemanuscript This do cument is available as pubmpimemosTRps via anonymous ftp from ftpmpiktuebmpgde or from the World Wide Web httpwwwmpiktuebmpg debuhtml Intro duction