Abstract

Adaptive resonance theory (ART) neural networks are becoming an important component of neural network industrial applications. Electronic implementation of these network models has proven difficult to scale to practical input dimensionality. In this paper, a new implementation of ART1 is proposed that efficiently combines optical and electronic devices. Global computations are performed by the optics, while local operations are performed in electronics. A physical implementation of this architecture that uses ferroelectric liquid crystal modulators integrated with VLSI circuitry is presented.¹ The system has the capacity to learn in real time and to store the neural weights in a two dimensional optically addressed smart spatial light modulator (SSLM). The implementation can be packaged in a multi-chip module of small physical dimensions. Macro-circuits can be constructed from these modules to perform complex logical functions. This system can also be modified to allow read-out and readin of the learned weights stored in the SSLM, for replication of trained systems. The sensitivity of the ART1 network algorithm to variations in modulator and detector device characteristics is discussed.

© 1992 Optical Society of America