Abstract

Optical implementations of one-layer, perceptron-like neural networks have been shown to be very successful at associating pattern/target sets despite large system errors [1,2]. It has also been shown that large systems can be realized with such architectures (≥4 x 10⁴ interconnections [2,3]), and appreciable processing speeds have been demonstrated (>10⁸ interconnections/sec [4]). However, single layer networks are limited due to their inability to associate patterns that are not linearly separable. A more general network is the two layer network, which is able to model arbitrary functions, and create any decision boundary within the input vector pattern space [5]. In order to implement such a network, it is necessary to perform a nonlinearity at the hidden layer before performing a subsequent matrix multiplication. In general, optical materials performing fast nonlinear processing require high optical powers. Hybrid opto-electronic devices can perform nonlinear operations at moderate speeds and low optical powers [6].

© 1991 Optical Society of America