Sensitivity of trained neural networks with threshold functions

Abstract

In this paper, we derive the sensitivity of single hidden-layer networks with threshold functions, called "Madaline", as a function of the trained weights, the input pattern, and the variance of weight perturbation or the bit error probability of the binary input pattern. The derived results are verified with a simulation of the Madaline recognizing handwritten digits. Our result show that the sensitivity in a trained network is far different from that of networks with random weights.<>