Sensitivity Analysis of Multilayer Perceptron

Abstract

Sensitivity analysis of neural network is useful for network design. Piché used a stochastic model to describe the Multilayer Perceptron (MLP), but it doesn’t match the true MLP closely, and too severe limitations are imposed on both input and weight perturbations. This paper attempts to generalize Piché’s stochastic model of MLP, and derive an universal expression of MLP’s sensitivity for all sigmoidal activation functions, without any restriction on input and output perturbations. The effects of network design parameters such as the number of layer, the number of neuron per layer and the chosen activation function are analyzed, and they provide useful information for network design decision-making. Furthermore, we use our sensitivity expression to design MLP for a given application. It can help to design the network structure, as well as the training of MLP.