Optical Neural Networks Using Competitive Learning

Abstract

Among the neural network models proposed, some of the most promising are those that use an optical correlator architecture. The fastest optical correlators are acousto-optical, but these are one-dimensional systems. Two-dimensional optical correlators can handle 2-D data and are particularly appropriate for the kinds of application for which optical neural nets are most promising, that is, recognition and classification of image formatted data. Because of the lack of suitable non-linear devices, present systems must use a combination of optics and electronics. Our hybrid system, which stores interconnect weights in a computer-generated hologram, and which includes a computer for thresholding and feedback, can implement a number of neural models including Hopfield, Hamming, adaptive resonance theory, and competitive learning. We report here on using this system for competitive learning applied to recognition of alphabetical characters.