Training by adaptive gain (TAG) neural network model, which had been developed for optical implementation of large-scale artificial neural networks, is further extended for better performance and its feasibility is demonstrated by a small-scale electro-optic implementation. For fully interconnected single-layer neural networks with N input and M output neurons the modified TAG model contains two different types of interconnections, i.e., MN fixed global interconnections and (beta) N + M adaptive local interconnections. For the original TAG model the number of adaptive local interconnections (beta) was set to 1, and the interconnections were understood as adaptive gain. For 2-dimensional input and output patterns the fixed global interconnections may be achieved by page-oriented holograms, and the adaptive local interconnections by spatial light modulators. The original and modified TAG models require much less adaptive elements than the popular perceptron model with fully adaptive global interconnections, and provide possibilities of implementing large-scale artificial neural networks with some sacrifice in performance. The training algorithm is based on gradient descent and error back-propagation, and is easily extensible to multi-layer architecture. Computer simulation and electro-optic implementation demonstrate much better performance of the modified TAG model compared to the original TAG model.

Date Published: 30 October 1992
PDF: 5 pages
Proc. SPIE 1812, Optical Computing and Neural Networks, (30 October 1992); doi: 10.1117/12.131195

Published in SPIE Proceedings Vol. 1812:
Optical Computing and Neural Networks
Ken Yuh Hsu; Hua-Kuang Liu, Editor(s)