Arithmetic codes for concurrent error detection in artificial neural networks: the case of AN+B codes

Abstract

A number of digital implementations of neural networks have been presented in recent literature. Moreover, several authors have dealt with the problem of fault tolerance; whether such aim is achieved by techniques typical of the neural computation (e.g. by repeated learning) or by architecture-specific solutions, the first basic step consists clearly in diagnosing the faulty elements. The present paper suggests adoption of concurrent error detection; the granularity chosen to identify faults is that of the neuron. An approach based on a class of arithmetic codes is suggested; various different solutions are discussed, and their relative performances and costs are evaluated. To check the validity of the approach, its application is examined with reference to multi-layered feed-forward networks.<>