Guiding Artificial Neural Networks Using Discriminatory Information In Hidden Layers

Artificial Neural Networks (ANN) are gaining much of popularity in recent years due to the advancement of fast GPU based machines and tremendous growth in data to train the networks. Traditionally back propagation algorithms are used to learn network weights that minimizes the error in the training set. Recent advances in Deep Neural Networks have shown that intermediate hidden network are able to learn structural features even from unstructured input data. In this paper, we propose a novel method to guide ANN learning algorithms using the discriminatory information available in the intermediate hidden layers of the network. For a binary classification problem, we guide the learning algorithm minimizing the discriminatory information of each individual neurons along with traditional error function. We have tested our proposed method on several standard benchmark datasets. Experimental results are promising and show significant improvement over traditional error functions.