Fault Tolerant Regularizers for Multilayer Feedforward Networks

In multilayer feedforward networks (MFNs), when open weight fault exists, many potential faulty networks should be considered during training. Hence; the objective function, as well as the corresponding learning algorithm; would be computationally complicated. This paper derives an objective function for improving the fault tolerance of MFNs. With the linearization technique; the objective function is decomposed into two terms; the training error and a simple regularization term. In our approach, the objective function is computational simple. Hence; the conventional backpropagation algorithm can lie simply applied to handle this fault tolerant objective function. Simulation results show that compared with the conventional approach; our approach has a better fault tolerant ability.