Nonlinear discriminant features constructed by using outputs of multilayer perceptron

The paper proposes a method to extract nonlinear discriminant features from given input measurements by using outputs of a multilayer perceptron (MLP). Linear discriminant analysis (LDA) is one of the best known methods to construct linear features which are suitable for class discrimination. Otsu (1975, 1981) showed that LDA can be extended to nonlinear if one can estimate Bayesian a posteriori probabilities. Previously, MLPs have been successfully applied to many kinds of pattern recognition problems. It is also regarded that outputs of MLPs trained for pattern classification approximate Bayesian a posteriori probabilities. Thus one can construct nonlinear discriminant features that maximize the discriminant criterion by using outputs of MLPs as estimates of Bayesian a posteriori probabilities