Hidden-layer neuron redundancy-analysis and application in MLP´s fault tolerance

Redundancy on hidden-layer neurons has proven useful in the fault tolerance of neural networks. This approach has been applied successfully in the fault tolerance design of classification neural networks, thus the complete single-fault tolerance can be gained. But this approach can only be applied to the feedforward networks which has hard-limit activation functions in output layer. And it prove that this approach is valid only to single-fault. There are universal faults of neurons and weights in actual applications, so we evaluated this approach under universal faults in feedforward networks. We proved that the global fault-rate is reduced though the redundancy on hidden-layer neurons. Then we presented a practical and valid method of redundancy of hidden-layer neurons to gain fault tolerance