Multi-category bayesian decision by neural networks

Abstract

For neural networks, learning from dichotomous random samples is difficult. An example is learning of a Bayesian discriminant function. However, one-hidden-layer neural networks with fewer inner parameters can learn from such signals better than ordinary ones. We show that such neural networks can be used for approximating multi-category Bayesian discriminant functions when the state-conditional probability distributions are two dimensional normal distributions. Results of a simple simulation are shown as examples. © Springer-Verlag Berlin Heidelberg 2008.