Fault-tolerant model of neural computing

Abstract

A fault-tolerant model of feed-forward neural computing with mixed-mode redundancy is proposed and analyzed. A mixed-mode redundancy is a combination of spatial redundancy and temporal redundancy. The redundancy is based on the homogeneity of both structures and operations of neurons in neural networks. This fault-tolerant model can be applied to both hardware architecture and parallel software simulation. By storing multiple sets of weights in a neuron and recomputing the outputs of this neuron at other different neurons, faults in the neuron can be detected and the output errors can be corrected. The degree of the fault tolerance of this model is analyzed. Further, the sufficient conditions for detecting errors and recovering outputs are also presented. The model can highly increase the reliability of neural computing so that a fairly large number of faulty neurons can be detected and that the outputs of these faulty neurons can be recovered.<>