A modular neural network with RBF output units

Modular-type neural networks have been proposed for solving large-scale classification problems efficiently. They divide an original problem into a set of relatively small two-class classification problems. It has been shown that modular-type neural networks are more efficient from the training time and recognition rates point of view than the usual layered neural networks. They, however, still have the following problems: 1) the rejection rate on patterns in unlearned classes is low; and 2) the incremental learning for newly added classes is not efficient. In this paper, we propose a modular-type neural network model with RBF output units and an algorithm of the incremental learning to improve these problems.