NEURAL NETWORK LEARNING DEVICE HAVING PAST HYSTORY PRESERVATION FUNCTION

摘要

PURPOSE:To improve hystory preserving capacity by copying the output values of a hidden layer for plural past hours in a hystory preserving part in a learning pattern execution part and providing plural context layers becoming the input of a next network. CONSTITUTION:The hystory preserving part 7 copies the output values of the hidden layer 5 on the context layers 8a-8n and the part 7 is that becoming the input of the next network with an input layer 4 in the neural network. The layers (n) of the layers 8a-8n are decided by layer number information on the context layers in the learning pattern preserving part. The output of the layer 5 is multiplied by a copy coefficient (a) and it is copied on the layer 8a. Then, data corresponding to the output state of the hidden layer previous by one hour is stored. The copied value of the layer 8a is multiplied by a prescribed copy coefficient (b) and it is copied on the layer 8b. Thus, data corresponding to the output state of the hidden layer previous by two hours is stored. Thus, past hystory can strongly affect the learning function of the neutral network.