Gaussian based neural networks applied to pattern classification and multivariate probability density estimation

A Gaussian based neural network is applied to the clustering problem. We consider the hypothesis that the samples are drawn from a finite mixture of Gaussian density functions. Each of them corresponds to one cluster. Competitive learning algorithms are then used to estimate the network parameters. The number of units in the hidden layer is determined by minimising the information criterion of Akaike. Performance evaluations using training data from mixture Gaussian densities are presented.<
>