Comparison of gradient descent and conjugate gradient learning algorithms for classification of electrogastrogram

Abstract

Our previous study showed that the possibility of using an optical three-layer feedforward neural network employing the gradient descent learning algorithm for automated assessment of normality of the electrogastrogram. However, problems with this algorithm are slow convergence rate and critical user-dependent parameters. In the present study, two conjugate gradient learning algorithms (quasi-Newton and scaled conjugate algorithm) were introduced and compared with the gradient descent learning algorithm for the classification of the normal and abnormal electrogastrogram. Three indexes, the convergence rate, complexity per iteration and parameter robustness, were used to evaluate the performance of each algorithm. The results showed that the scaled conjugate gradient algorithm performed the best, which was robust and provided a super linear convergence rate.