Fault tolerance in neural networks: theoretical analysis and simulation results

Work is continuing on the intrinsic capacity of survival of fault characterizing neural nets per se. The authors deal with this theme, considering in particular multilayered feedforward nets. The study is performed on the abstract neural graphs, thus involving errors rather than faults. After an initial analysis of the error model, the effects of errors are mathematically derived and the conditions allowing the complete recovery from faults through redistribution of weights in the network (or otherwise allowing predetermined upper bounds on errors) are derived. Simulation results are presented identifying the effect of such errors on the neural computation. It is seen that (unless a good measure of redundancy is present in the net from the beginning) even single errors affect in a relevant way the computation. Correction of this effect is sought through repeated learning, i.e. an operation leading to the weight adjustment previously discussed in theoretical terms