Sensitivity of Madalines to input and weight perturbations

The sensitivity of a neural network's output to its input and weight perturbations is an important measure for evaluating the network's performance. In this paper, an approach to quantify the sensitivity of the feedforward network-Madeline is proposed. The sensitivity is defined as the probability of output error due to input and weight perturbations with respect to overall input patterns. Based on the structural characteristics of the Madaline, a bottom-up approach is adopted. The sensitivity of a single neuron, i.e. an Adaline, is considered first, and an algorithm is given for the computation of the sensitivity. Then followed is the sensitivity of the entire Madaline network, and another algorithm is given to compute the sensitivity. Computer simulations are run to verify the effectiveness of the algorithms. The theoretical results are in good agreement with the computer simulation results.