The eigenspace separation transform for neural-network classifiers

This paper presents a linear transform that compresses data in a manner designed to improve the performance of a neural network used as a binary classifier. The classifier is intended to accommodate data distributions that may be non-normal, may have equal class means, may be multimodal, and have unknown a priori probabilities for the two classes. The transform, which is called the eigenspace separation transform, allows the reduction of the size of a neural network while enhancing its generalization accuracy as a binary classifier. Published by Elsevier Science Ltd.