A training algorithm for optimum margin classifiers

DB DB error m error m no smo othing The p erformance impro v ed considerably for DB F or DB the impro v emen t is less signi can t and the opti m um w as obtained for less smo othing than for DB This is exp ected since the n um b er of training patterns in DB is m uc h larger than in DB v ersus A higher p erformance gain can b e exp ected for more selec tiv e hin ts than smo othing suc ha s i n v ariance to small rotations or scaling of the digits SLD Better p erformance migh tb e a c hiev ed with other sim ilarit y functions K x x Figure sho ws the decision b oundary obtained with a second order p olynomial and a radial basis function RBF maxim um margin classi er with K x x e x p k x x k The decision b oundary of the p olynomial classi er is m uc h closer to one of the t w o classes This is a consequence of the non linear transform from space to x space of p olynomials whic h realizes a p osition dep enden t scaling of distance Radial Basis F unctions do not exhibit this problem The ...