Initial analysis on sensitivity of multilayer perceptron
D.S. Yeung, X.Z. Wang · 2003
One of the most important issues in the study of neural networks is its sensitivity arising from weight perturbation and input error, which may be caused by initial error, propagation or limited precision in data storage. It is difficult to establish a generic sensitivity formulation for the general multilayer perceptron. The paper continues our previous work on sensitivity analysis of the Neocognitron and makes an initial effort to analyze the sensitivity of a multilayer perceptron by following a similar method used in the case of the Neocognitron. Initial estimates on the expected decision error of an individual node due to the weight perturbation and on the expected decision error due to the input error are obtained. One general method is demonstrated by considering a special activation function for the multilayer perceptron. These preliminary investigations form the basis for our overall sensitivity analysis of the multilayer perceptron.