A Novel Design Method for Multilayer Feedforward Neural Networks

Abstract

A multilayer feedforward neural network and a design method that takes the distribution of given training patterns into consideration are proposed. The size of the network, initial values of interconnection weights, and parameters defining the nonlinearities of processing elements are determined from a specially selected portion of the given training patterns, which is called a set of feature points. With these initial settings, the performance of the network is further improved by a modified error backpropagation learning process. It is shown in several examples that the proposed model and the design method are capable of rapidly learning the training patterns compared to conventional multilayer feedforward neural networks with random initialization techniques.