AN OPTOELECTRONIC IMPLEMENTATION OF THE ADAPTIVE RESONANCE NEURAL-NETWORK

Implementation of the adaptive resonance theory (ART) of neural networ ks has been a thorny problem for several years. This work presents a n ovel solution to the problem by using an optical correlator, allowing the large body of correlator research to be leveraged in the implement ation of ART. The implementation takes advantage of the fact that one ART-based architecture, known as ART1, can be broken into several part s, some of which are better to implement in parallel. The control stru cture of ART, often regarded as its most complex part, is actually not very time consuming and can be done in electronics. The bottom-up and top-down gated pathways, however, are very time consuming to simulate and are difficult to implement directly in electronics due to the hig h number of interconnections. Two facts simplify this. The first is th at the pathways are computing a set of inner products. These inner pro ducts represent as least 80% of the computation time of the ART1 imple mentation. The second insight, our contribution, is that implementing the inner products optically, and the rest of the network in electroni cs, is a very effective marriage of the two technologies to realize th e ART1 network. In addition to the design, we present experiments with a laboratory prototype to illustrate its feasibility and to discuss i mplementation details that arise in practice. This device potentially can significantly outperform alternative implementations of ART1 by as much as two to three orders of magnitude in problems requiring especi ally large input fields. It should be noted that all of these results apply to just one of the various ART architectures, known as ART1, but that other ART networks and other neural nets in general also use inn er products and could benefit from this work as well.