On two-pattern classification and feature selection using neural networks

Two new methods for feature selection and two-pattern classification by neural networks (NN) are presented. The network utilized is a single-neuron classifier. In the first approach the pocket perceptron learning algorithm is used for feature selection, and NN classifier training as well. The goals are to reduce the dimension of a feature set by selecting a subset of features of high discrimination power and to minimize the frequency of misclassification. In the second approach a modified pocket perceptron learning algorithm is used. In addition to select a subset of high discriminating features, one of the main goals of this approach is to minimize the total misclassification frequency and the false rejection error (or false acceptance error) simultaneously. The proposed methods were applied to a signature verification problem