Perturbation response in feed-forward neural networks

Two advantages claimed for feedforward neural networks with continuous-valued activation functions are robustness and a distributed nature. These issues are addressed from a very specific perspective: How sensitive is a given, trained network to perturbations in individual internal neurons? Using first-order approximations, a tractable model that predicts useful statistics of the desired sensitivities from basic information about the perturbing process is derived. The model has been tested on several trained and random network architectures. The particular case investigated considers perturbations on non-output neurons only, and for simple uniform distributions