Multi-category Bayesian Decision by Neural Networks

For neural networks, learning from dichotomous random samples is difficult. An example is learning of a Bayesian discriminant funciton. However, one-hidden-layer neural networks with fewer innner parameters can learn from such signals better thatn ordinary oens. 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 simple simulation are shown as examples.