A Bayesian Regularization Method for the Probabilistic RBF Network

The Probabilistic RBF (PRBF) network constitutes a recently proposed classification network that employs Gaussian mixture models for class conditional density estimation. The particular characteristic of this model is that it allows the sharing of the Gaussian components of the mixture models among all classes, in the same spirit that the hidden units of a classification RBF network feed all output units. Training of the PRBF network is a likelihood maximization procedure based on the Expectation - Maximization (EM) algorithm. In this work, we propose a Bayesian regularization approach for training the PRBF network that takes into account the existence of ovelapping among classes in the region where a Gaussian component has been placed. We also propose a fast and iterative training procedure (based on the EM algorithm) to adjust the component parameters. Experimental results on well-known classification data sets indicate that the proposed method leads to superior generalization performance compared to the original PRBF network with the same number of kernels.