Collective Computation of Many-Body Properties by Neural Networks

Abstract

Artificial neural networks constitute a novel class of many-body systems in which the particles are neuron-like units and the interactions are weighted synapse-like connections between these units.^1,2 The most extraordinary feature of these systems is that the interactions are subject to modification, depending on the states recently visited by the system. Thus, as the network experiences varied stimuli, knowledge can be stored in the neuron-neuron interactions, for later retrieval in some information-processing task. Indeed, multilayered, feedforward networks of analog neurons can be taught by example to solve complex pat tern-categorization problems using the backpropagation learning algorithm³ or other procedures for modifying connection weights. During the learning process, inner neurons may evolve into useful feature detectors tailored to regularities or correlations inherent in the ensemble of input stimulus patterns and desired output response patterns used for training. The system builds an internal representation, or model, of its pattern environment, which may provide a good approximation to the actual rules determining the underlying input-output map. Accordingly, the artificial neural network may possess a useful generalization or predictive ability, as demonstrated by a high percentage of correct responses when presented with unfamiliar input patterns absent from the training set.