A method for training feed forward neural network to be fault tolerant

A method for training a feedforward neural network to be fault tolerant against weight perturbations is described. The measure for fault tolerance is the deviation of the network's output after training, when each interconnection weight is perturbed, from that output without perturbation. In this method, an attempt is made to keep that deviation as low as possible. This measure is used because it can represent that kinds of error which arises when neural networks are implemented in hardware.<
>