Direct validation of the information bottleneck principle for deep nets

The information bottleneck (IB) has been suggested as a fundamental principle governing performance in deep neural nets (DNNs). This idea sparked research on the information plane dynamics during training with the cross-entropy loss, and on using the IB of some 'bottleneck' layer as a loss function. However, the claim that reaching the maximal value of the IB Lagrangian in each layer leads to optimal performance, was in fact never directly confirmed. In this paper, we propose a direct way of validating this hypothesis, using layer-by-layer training with the IB loss. In accordance with the original theory, we train each DNN layer explicitly with the IB objective (and without any classification loss), and freeze it before moving on to train the next layer. While mutual information (MI) is generally hard to estimate in high dimensions, we show that in the case of MI between DNN layers, this can be done quite accurately using a modification of the recently proposed mutual information neural estimator. Interestingly, we find that layer-by-layer training with the IB loss leads to accuracy which is on-par with end-to-end training with the cross entropy loss. This is, thus, the first direct experimental illustration of the link between the IB value in each layer, and a net's performance.