A poly-time analysis of robustness in feedforward neural networks

The paper provides a methodology for evaluating the performance degradation of a feedforward neural network once affected by fixed perturbations injected in the computation. The loss in performance associated with the perturbed computation can be evaluated with a polynomial time algorithm and a general family of loss functions by resorting to a probabilistic analysis based on randomized algorithms. The methodology can be used to test the impact of finite precision implementations (analog, digital or mixed) on the weights of a feedforward neural network.