A FAULT-TOLERANT MULTILAYER NEURAL-NETWORK MODEL AND ITS PROPERTIES

Although it is pointed often that multilayer neural networks should have a certain degree of fault tolerance, very few discussions based on the rigorous definition of fault tolerance have been made so far. Also, there have been few discussions on the mechanisms that bring out the fault tolerance. This paper shows that a learning algorithm that directly reduces a measure of fault tolerance can be derived in a similar way to the conventional back-propagation. By analyzing the resulting networks, the mechanism that realizes the fault tolerance and the properties of the fault-tolerant networks are investigated. Simulation results show the effectiveness of the proposed learning algorithm. It also is revealed that the utilization of the redundant hidden units and the saturation property of the sigmoid function realizes the fault tolerance. Moreover, it is shown that a good influence on generalization ability can be expected from the learning algorithm for fault tolerance.