Design of fault tolerant multilayer perceptron with a desired level of robustness

The definition of a fault tolerant neural network is presented. The definition makes it possible to design a network with a desired level of robustness. Based on this definition, an efficient method is proposed: called selective augmentation, which transforms a trained network into one that is fault tolerant against a stuck at 0 fault at the hidden neurons. It is shown, through an example, that the resulting network designed by the proposed method is not only fault tolerant, but also much less redundant than a network designed by uniform augmentation