A method of BP network learning by expanding the distribution of category

Abstract

In backpropagation networks, unlearned regions are left between categories if the learning samples are comparatively small. Such unlearned regions are one of the reasons for the degradation of network generalization ability. To improve the generalization ability, it is preferable that the boundaries of the categories are more accurately reflected by the pattern distribution. This article presents the method of expansion of the category distribution by adding displacements proportional to the distance from the center of gravity of the category to learning samples, and a backpropagation (BP) learning method using given learning samples and those displaced samples. The method is applied to the recognition of handwritten Kanji characters. We confirm increased generalization abilities as a result of increased recognition performance of unlearned samples in comparison to the normal learning method. © 1999 Scripta Technica, Syst Comp Jpn, 30(12): 16–24, 1999