Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition-' Washington , D . C . , June , 1983 OPTIMAL PERCEPTUAL INFERENCE

When a vision system creates an interpretation of some input data, it assigns truth values or probabilities to internal hypotheses about the world. \Ve present a non-deterministic method for assigning truth values that avoids many of the problems encountered by existing relaxation methods. Instead of representing probabilities with realnumbers. we use a more direct encoding in which the probability \ associated with a hypothesis is represented by the probability that it is in one of two states. true or false. We give a particular nondeterministic operator. based on statistlcal mechanics. for updating the truth values of hypotheses. The operator ensures that the probability of discovering a particular combination of hypotheses is a simple function of how good that combination is. We show that there is a simple relationship between this operato.r and Bayesian inference. and we describe a learning rule which allows a parallel system to converge on a set"of weights that optimizes its perceptual inferences.