The improved localized generalization error model and its applications to feature selection for RBFNN

In pattern classification problems, the generalization error caused more and more attentions because of its importance for classifier´s training. Wing W.Y. NG et al. proposed localized generalization error model compared to global generalization error model. The idea is perfect, but the derivation of the error model and stochastic sensitivity measure has some flaws. In this paper, we propose an improved localized generalization error model in order to avoid these flaws of the model proposed by Wing.