Learning in neural networks with Bayesian prototypes

Given a set of samples of a probability distribution on a set of discrete random variables, we study the problem of constructing a good approximate neural network model of the underlying probability distribution. Our approach is based on an unsupervised learning scheme where the samples are first divided into separate clusters, and each cluster is then coded as a single vector. These Bayesian prototype vectors consist of conditional probabilities representing the attribute-value distribution inside the corresponding cluster. Using these prototype vectors, it is possible to model the underlying joint probability distribution as a simple Bayesian network (a tree), which can be realized as a feedforward neural network capable of probabilistic reasoning. We describe how the prototypes can be determined, given a partition of the samples, and present a method for evaluating the likelihood of the corresponding Bayesian tree. We also present a greedy heuristic for searching through the space of different partition schemes with different numbers of clusters, aiming at an optimal approximation of the probability distribution.