A construction method of feedforward neural network for selecting effective hidden nodes

Abstract

Based on the analysis of construction methods proposed recently, a new construction method of feedforward neural network (FNN), which is called CMSEHN (construction method for selecting effective hidden nodes), is proposed for classification tasks in this paper. The most significant contribution of CMSEHN is that it guarantees the convergence of the learning procedure with fast speed, automatically determines the number of neurons in the hidden layer and selects much effective hidden nodes to improve the generalization ability of network. Also, CMSEHN can solve the two problems of the other construction methods, that are difficulty in improving generalization ability and difficulty in applying them to problems with analog input vectors of different length. In the meantime, the network constructed by CMSEHN is able to learn an arbitrary training set. Experiment results of handwritten digit recognition show that the network constructed by CMSEHN has good learning and generalization performance. © 1997 IEEE.