Accuracy vs. precision in general purpose neural digital VLSI architectures

The paper provides a sensitivity analysis to measure the loss in accuracy induced by perturbations affecting multilayered perceptrons. Differing from that suggested in the related literature, we do not assume a large number of weights or independence among inputs for the neurons nor again strict requirements on the nonlinear function. Results are then tailored to digital VLSI implementations where perturbations, caused by quantisation, affect the device inputs, weights, internal values and outputs