Training Autoencoders Using Relative Entropy Constraints

Autoencoders are widely used for dimensionality reduction and feature extraction. The backpropagation algorithm training the parameters of autoencoder model suffers from problems such as slow convergence. Therefore, researchers propose forward propagation algorithms. However, existing algorithms do not consider characteristics data itself. This paper proposes an based on relative entropy constraints, called (REAE). When solving map parameters, REAE imposes different constraints average activation value hidden layer outputs obtained by sets. In experimental section, compared applying features extracted to image classification task. results three datasets show that performance constructed is better than other