Weight initialization of MLP classifiers using boundary-preserving patterns

This paper presents a new weight initialization technique for three layer feedforward neural network classifiers. The method estimates the subsidiary discriminant functions, represented by middle layer node activations, using a priori information about the class boundaries. A set of boundary-preserving patterns are extracted from the original training set using a modified condensed nearest neighbor algorithm. Unlike the approach proposed by Smyth (1992), this method does not require an initial guess of the appropriate number of cluster centers needed to correctly estimate the class boundaries