OPTICAL-PATTERN CLASSIFIER WITH PERCEPTRON LEARNING

An optical realization of a single layer pattern classifier is described in which Perceptron learning is implemented to train the system weights. Novel use of the Stokes’s principle of reversability for light is made to realize both additive and subtractive weight modifications necessary for true Perceptron learning. This is achieved by using a double Mach-Zehnder interferometer in conjunction with photorefractive hologram recording. Experimental results are given which show the high quality subtractive changes that can be made. © 1990 Optical Society of America.