Manipulation of hidden units activities for fault tolerant multi-layer neural networks

We propose a new training algorithm to enhance fault tolerance of multi-layer neural networks (MLNs). This method is based on the fact that faults on connections between hidden layer and output layer have a harmful effect on fault tolerance of MLNs. to decrease these effects, we introduced two approaches, (1) reduce the number of strong connections between hidden layer and output layer, (2) neutralize the activities of hidden units. The first approach aims to reduce the undesirable connections. The second one aims to increase redundancy of internal representation.