ABSTRACT
In this paper the sensitivity of a feed-forward neural network to weights and inputs perturbations is studied at a behavioural level. By modelling inputs and perturbations we construct a stochastic frame which provides a measure of the loss in performance at the network outputs. A suitable definition for the signal to noise ratio applied to the output neural signal and to the perturbation, here induced by a finite precision representation of neural values, provides precision requirements for inputs and weights. Results are then tailored to image processing and applied to two real-time demanding applications ship identification in radar images and defect identification in machined parts of mechanical objects, the second being realised with a dedicated digital VLSI chip.
