Abstract

Optical implementations of neural networks have attracted substantial attention because of the promise of relative ease in achieving a high degree of parallelism. First demonstration of optical neural net systems include those based on the Hopfield associative memory model. However, the Hopfield associative memory has found limited application due to its relatively small storage capacity. This has given impetus to the investigation of alternative associative memory algorithms which yield increased storage capacity. Among these is the high-order associative memory, in which the increase in degrees of freedom provides large additional memory capacity. In this paper we describe the first optical implementation of a quadratic associative memory based on an inner product representation. Our experimental system is the first execution of an optical chip design, which eliminates the need for bulk optics, and results in an extremely compact network. The system utilizes liquid crystal modulators to represent input vectors and ambient room light as the light source. The experimental system also incorporates bipolar storage of memory vectors. The ten-input system stably stores up to four memory vectors and exhibits error correction capability.

© 1989 Optical Society of America