Sensitivity to errors in artificial neural networks: a behavioral approach

The problem of sensitivity to errors in artificial neural networks is discussed here considering an abstract model of the network and the errors that can affect a neuron´s computation. Feed-forward multilayered networks are considered; the performance taken into account with respect to error sensitivity is their classification capacity. The final aim is evaluation of the probability that a single neuron´s error will affect both its own classification capacity and that of the whole network. A geometrical representation of the neural computation is adopted as the basis for such evaluation. Probability of error propagation is evaluated with respect to the single neuron´s output as well as to the complete network´s output. The information derived is used to evaluate, for a specific digital network architecture, the most critical sections of the implementation as far as reliability is concerned and thus to point out candidates for ad-hoc fault-tolerance policies