核函数方法_上_.pdf

() 罗 公 亮 (, 100071) [ ] , ,, Fisher , Fisher (KFD) ( KPCA) , ,,, [ ] ; Fisher ; ; [ ] TP183 []A [] 1000-7059( 2002) 03-0001-04 Kernel-based methods(A) LUO Gon-lian g g (Auto ation Research and Desi n Institute of Metallur ical Industr , Bei in 100071, China) g g y j g Abstract: Theappealof kernel-based ethodshas beenarisen in recent yearsbythesuccessof support vector a- chines. Any standard linear technique,if applied in an high di ensional featurespaceresulted fro a nonlinear a of theori inal in ut s ace, a be transfor ed to its nonlinear version in the in ut s ace ver si l b p g p p y p p y p y y eansof the kernel trick, rovided that interactions between theco onents of theal orith areli ited to dot p p g roducts. Fisher discri inant and rinci al co onent anal sisareclassical lineartechni ues widel used in at- p p p p y q y p tern classification and feature extraction. Their corresponding nonlinear versions, the kernel Fisher discri -i nant(KFD) and kernel principal co ponent analysis(KPCA), havee erged recently with better perfor ance, widerapplication areaand oreflexibility, which arenew technologieswith pro ising prospects in applications and worthwhile to a attention to. p y Ke words: kernel function; Fisher discri inant; rinci al co onent anal sis; su ort vector achines y p p p y pp [ 1] VC , (SVM) 5 ( #) , : ( SRM)