Renewal method of weight matrix in optical neural network

We proposed a new renewal method of the weights to realize a back propagation learning algorithm with a large number of iteration times in an optical neural network. The main purpose of this method is to use the limited dynamic range of an optical device widely. Since some weights displaying on the device are quickly saturated in the conventional method, the learning is not completed in many cases. The new technique can suppress saturation of the weights so that learning is completed. The computer simulation and the optical experiment are presented.