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

We present a new definition of fault tolerance for an MLP by introducing the concept of a desired level of robustness. Based on this definition we propose an efficient method, called selective augmentation, which transforms a trained MLP to the fault tolerant one against a stuck_at_0 fault at the hidden neurons. We show, through an example, that the resulting networks designed by the proposed method are not only fault tolerant but also less redundant than the ones by the uniform augmentation.