Optical learning neural networks with two-dimensional structure

We constructed optical systems for learning neural networks with 2D structure sing Selfoc microlens arrays. On the systems, pattern recognitions of typed alphabetic characters which were directly detected with a CCD camera were realized. Liquid crystal devices, an electron-beam addressed spatial light modulator and a Pockels readout optical modulator were used for displaying weight matrices of the neural networks. The weights were renewed according to the random search algorithm or the delta rule with error signals calculated optically. The 2D structure for image processing can be implemented with large scale networks that consist of several thousands input neurons.