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

Fault tolerance is a frequently cited advantage of artificial neural nets, yet it has rarely been the subject of specific study. In this paper, we investigate empirically the performance under damage conditions of single- and multilayer perceptrons (MLP's), with various numbers of hidden units, in a representative pattern-recognition task. While some degree of graceful degradation was observed, the single-layer perceptron was considerably less fault tolerant (at least, as far as the performance metric employed here indicates) than any of the multilayer perceptrons, including one with fewer adjustable weights. Our initial hypothesis that fault tolerance would be significantly improved for multilayer nets with larger numbers of hidden units proved incorrect. Indeed, there appeared to be a liability to having excess hidden units. A simple technique (called augmentation) is described, however, which was succesful in translating excess hidden units into improved fault tolerance. Finally, our results were supported by applying singular value decomposition (SVD) analysis to the MLP's internal representations.