METHOD FOR STRUCTURING FEEDFORWARD TYPE NEURAL NETWORK

摘要

PURPOSE: To efficiently structure the feedforward type neural network by determining and removing a neuron which can be omitted. ;CONSTITUTION: The feedforward type neural network is made to learn with plural learning data and a matrix O is generated from the feedforward type neural network obtained by the learning. Then the generated matrix is decomposed into column vectors as many as hidden layer neurons, and a specific number of column vectors which are high in primary independency are found among the decomposed column vectors. Hidden layer neurons corresponding to the column vectors which are high in primary independency are left in a hidden layer and other hidden layer neurons are removed from the hidden layer. New link weight is determined for the feedforward type neural network having the hidden layer consisting of the left hidden layer neurons.;COPYRIGHT: (C)1995,JPO