摘要: The author presents a survey of the basic theory backpropagation neural network architecture covering architectural design, performance measurement, function approximation capability, and learning. includes previously known material, as well some new results, namely, formulation to make it valid (past formulations violated locality processing restriction) proof that mean-squared-error exists is differentiable. Also included theorem showing any L/sub 2/ from (0, 1)/sup n/ R/sup m/ can be implemented desired degree accuracy with three-layer network. speculative neurophysiological model illustrating how might plausibly in mammalian brain for corticocortical learning between nearby regions cerebral cortex. >