Estimating the number of hidden neurons of the MLP using singular value decomposition and principal components analysis: a novel approach

Abstract

This paper presents a novel technique to estimate the number of hidden neurons of an MLP classifier. The proposed approach consists in the post-training application of SVD/PCA to the backpropagated error and local gradient matrices associated with the hidden neurons. The number of hidden neurons is then set to the number of relevant singular values or eigenvalues of the involved matrices. Computer simulations using artificial and real data indicate that proposed method presents better results than obtained with the application of SVD and PCA to the outputs of the hidden neurons computed during the forward phase of the MLP training.