Quantitative Evaluation of Salient Deep Neural Network Features Using Random Forests

The Deep Neural Networks and Deep Convolutional Neural Network have the property of providing multi-scale features at different layers of the network. Combination of these large number of features is one of the attributed reasons for the performance of the Neural Network (NN) on vision problems. This work uses Random Forests to identify robust features at various layers of the NN and evaluates the classification performance of these features in isolation. We propose a method for evaluation of parts of an already trained network using the selection by entropy maximization property of the Random Forests. We define measures for saliency in terms of contribution to the final classification, and evaluate the feature saliency. Simultaneously, a measure to identify the imperativeness of network features for classification is also formalized. The experiments made on a Hand dataset and the MNIST dataset, quantitatively validate various intuitions like the discriminatory nature of the outer layer features.