Attractor Manipulation in Denoising Autoencoders for Robust Phone Recognition

Abstract

Autoencoder Neural Networks can filter unwanted variabilities; however, their performance will degrade if their attractors and their basins of attraction are not correctly adjusted. This paper proposes a heuristic method to increase attractors shaped in desired points and expand their basins of attraction. These well-formed attractors can compensate variabilities and hence increase the chance of robust recognition. This method's effectiveness is shown on synthetic data and is compared with another attractor manipulation method called the cyclic method. This method's performance on the phone recognition task has shown 22.1 percent relative increase in the number of attractors and 4.2 percent relative improvement in the phone error rate on the Farsdat database.