Abstract

The back-propagation algorithm¹ has become increasingly popular in the neural-net research community. Various optical implementations have been proposed, with the hope of increased performance by means of the parallelism of optics. Realistic models of optoelectronic implementations must include the effects of noise. Although noise often anneals, increasing the convergence rate, excessive noise can effectively transform any updating algorithm into a random search among weight configurations. For the purpose of evaluating the effects of component noise on the performance of the back-propagation algorithm, we have developed a simulation program that permits insertion of noise terms wherever appropriate. The program introduces noise processes into the weight-updating process during the learning phase. The learning curve is defined as the mean-square error versus the iteration number. Our analysis emphasizes examination of learning curves rather than simply the probability of convergence.

© 1990 Optical Society of America