ABSTRACT
The Bayesian neural network implements an optimal Bayesian classifier in the case of independent evidence. If the evidence is not independent, a hidden layer with an independent representation can be used in the network, to overcome this limitation on the evidence. This Bayesian neural network has been tested on a realistic task of classifying faults in a telephone exchange computer, with very good results. The results are compared to those of a backpropagation neural network on the same data. Also presented is an explanatory mechanism for finding the primary causes for a classification.
