Cross-Entropy Regularization with Mutual Information in Training CNNs

We examine the learning behavior of a shallow and a deep convolutional neural network performing classification tasks on subsets of two databases. Our investigation focuses on the label, the input, and the prediction layer, and we compute the mutual information between these layers epoch-wise using Rényi’s matrix-based entropy functional. We evaluate the data processing inequality to interpret the learning behavior in a consistent information-theoretic framework. Our primary goals are to 1) clarify the relation between the two training objectives of minimizing the cross-entropy and maximizing the mutual information between the label and the prediction layer, 2) gradually switch from the first to the second training objective, and 3) interpret the impact of the latter transition. One of the main contributions is the proposed novel method for regularizing the cross-entropy objective and assessing the neural network’s learning activity.