DETERMINING AND IMPROVING THE FAULT-TOLERANCE OF MULTILAYER PERCEPTRONS IN A PATTERN-RECOGNITION APPLICATION

Fault tolerance is a frequently cited advantage of artificial neural n ets, yet it has rarely been the subject of specific study. In this pap er, we investigate empirically the performance under damage conditions of single- and multilayer perceptrons (MLP's), with various numbers o f hidden units, in a representative pattern-recognition task. While so me degree of graceful degradation was observed, the single-layer perce ptron was considerably less fault tolerant (at least, as far as the pe rformance metric employed here indicates) than any of the multilayer p erceptrons, including one with fewer adjustable weights. Our initial h ypothesis that fault tolerance would be significantly improved for mul tilayer nets with larger numbers of hidden units proved incorrect. Ind eed, there appeared to be a liability to having excess hidden units. A simple technique (called augmentation) is described, however, which w as succesful in translating excess hidden units into improved fault to lerance. Finally, our results were supported by applying singular valu e decomposition (SVD) analysis to the MLP's internal representations.