Insight into the learning process of neural networks for classifying acoustic signatures

Abstract

Feedforward networks employing the backward propagating delta rule for error correction were tested utilizing simulated target signatures and noise to provide insight into the network learning process. Network training histories and weight evolutions were studied for alternating signal and noise input vectors for two network architectures. Contour plots of the input‐to‐hidden layer weights clearly indicate the relationship between the evolving features of the network weights as they respond to the input signatures during the learning process. Singular value decomposition of the input‐to‐hidden layer transfer matrix provides insight into the similarities of a trained neural network and a classical matched filter.