Collective Computation of Many-Body Properties by Neural Networks

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