Training of neural classifiers by separating distributions at the hidden layer

A new cost function for training of binary classifiers based on neural networks is proposed. This cost function aims at separating the distributions for patterns of each class at the output of the hidden layer of the network. It has been implemented in a Generalized Radial Basis Function (GRBF) network and its performance has been evaluated under three different databases, showing advantages with respect to the conventional Mean Squared Error (MSE) cost function. With respect to the Support Vector Machine (SVM) classifier, the proposed method has also advantages both in terms of performance and complexity.