Pruning the Multilayer Perceptron through the Correlation of Backpropagated Errors

In this paper we introduce an efficient methodology for pruning unnecessary weights of a trained multilayer Perceptron (MLP). The proposed method is based on the correlation analysis of the errors produced by the output neurons and the backpropagated errors associated with the hidden neurons. Weight connections for which the correlations are smaller than a user-defined error tolerance are discarded. The successive application of this weight pruning methodology leads eventually to the complete elimination of all connections of a neuron. Extensive computer simulations using synthetic and real-world data indicate that the proposed method present better or equivalent results than standard pruning techniques, such as the Optimal Brain Surgeon (OBS) and Weight Decay and Elimination (WDE), with much lower computational costs. We also show that the resulting optimal architecture matches that provided by Akaike's Information Criterion (AIC).