LEARNING PROCESSING SYSTEM OF NETWORK STRUCTURE DATA PROCESSING UNIT

摘要

A basic unit for obtaining an output is constituted by multiplying a plurality of inputs by weights, totaling the results of multiplications to obtain a product sum and obtaining the output from this product sum by use of a threshold function. A hierarchical network consisting of an input layer, an intermediate layer and an output layer is constituted by use of such a basic unit. The hierarchical network conducts learning by determining the updating quantity of the weights from errors between the output signals obtained from the output layer for a plurality of inputs and teacher's signals representing the values which the output signals should take on, in accordance with a back-propagation method in order to obtain the weight value which keeps the sum of the error values within a predetermined allowance range. In learning of such a hierarchical network, the updating quantity of the weight to be obtained at the time of the current weight updating cycle is determined in consideration of the updating quantity of the weight obtained at the time of one cycle earlier updating cycle than the last updating cycle in addition to the updating quantity of the weight determined at the time of the last weight updating cycle.