Optical Connectionist Machine With Polarization-Based Bipolar Weight Values

Abstract

Associative memory and pattern recognition systems have many similarities. Autoassociative memories can be viewed as image restoration, and heteroassociative memories as classification or feature extraction systems. Connectionist, or neural network, architectures are well suited to perform associative memory operations. The predominant calculation in these systems is vector-matrix multiplication. While requiring 0(N2) operations on a serial machine, a simple optical architecture can perform this calculation in nearly constant time. We have demonstrated an optoelectronic connectionist module with modifiable inputs and weight matrices that performs associative memory operations. The weight matrix can be generated with or without the optical hardware, providing a testbed for simulations of physically implemented connectionist systems. Low cost commercial liquid crystal television sets are used to encode unit activities as intensity of linearly polarized light and signed multiplication as rotation of this light. Integrated computer control allows the extension to many connectionist models.