THE INFLUENCE OF COMPONENT IMPERFECTIONS ON TRAINABLE SYSTEM PERFORMANCE.
Interim technical rept., Aug 61-June 63,
STANFORD UNIV CALIF STANFORD ELECTRONICS LABS
Pagination or Media Count:
Numerous investigators have considered the behavior of trainable threshold-logic systems using somewhat idealized training procedures. Actual systems built with present-day components seldom implement the ideal training rules. This report considers the effect of such component defects as nonlinearity, nonuniformity, and hysteresis on the performance of trainable systems. The four areas of system performance considered are the ability of the system to solve a given problem, the stability of the resulting solution, the generalizing ability of the system, and the time required to find a solution. A convergence proof is presented which shows that even a defective system will converge to a solution if one exists. An analytical technique called the rate vector analysis is developed which can be used to predict with reasonable accuracy the time required for a defective system to converge. When used on a 16-input system in which the weights have nonlinear adaption characteristics, the increase in training time due to the defect was predicted within 15 percent. Author