Stochastic spin models for pattern recognition

Abstract
We exploit for the recognition of patterns the properties of physical spin systems to assume long range order and, thereby, to establish a global interpretation of patterns. For this purposes we choose spins which can take a discrete set of values to code for local features of the patterns to be processed (feature spins). The energy of the system entails a field contribution and interactions between the feature spins. The field incorporates the information on the input pattern. The spin‐spin interaction represents ‘a priori’ knowledge on relationships between features, e.g. continuity properties. The energy function is choosen such that the ground state of the feature spin system corresponds to the best global interpretation of the pattern. The ground state is reached in the course of local stochastic dynamics, this process being simulated by the method of Monte Carlo annealing. Our study is related to work presented in Refs. 2, 3.