Two-pattern classification and feature extraction based on minimum error decision boundary using neural networks

A new method is proposed for two pattern classification and feature extraction based directly on an optimum decision boundary using neural networks (NN). The proposed approach has several desirable properties: (1) it predicts an optimum decision boundary which provides a classification accuracy at least as good as as that of an optimum global decision hyperplane; (2) it extracts optimum discrimination features even though the joint probability distribution of features is unknown; and (3) it determines the minimum number of discriminating features