Sensitivity analysis of artificial neural networks: Case study in clinical electromyography

Abstract

The usefulness of artificial neural networks (ANNs) trained with the momentum backpropagation and the conjugate gradient backpropagation (CGBP) learning algorithms in the classification of electromyography (EMG) data has recently been demonstrated. The sensitivity of feedforward-layered networks supplied with EMG data and trained with the CGBP learning algorithm to weight errors and random cutoff of connections is examined. The results suggest that ANN models are capable of tolerating weight error changes around the optimal values, as well as a limited number of disconnections.