Anatomize Deep Learning with Information Theory

A useful representation can discard details while preserving information needed for a particular prediction. Lilian Weng's 2017 article introduces the information-bottleneck view of neural networks through that tradeoff.

The original method makes relevance depend on the target. Its abstract defines relevant information as what one signal tells you about another, such as the information speech sounds provide about the words spoken, and seeks a short code that preserves as much of that information as possible. Applied to neural networks, this offers a way to ask what a hidden representation retains. The article's account of training first fitting and then compressing is a proposed explanation, rather than a universal law of network training.

“DNN training involves two distinct phases: First, the network is trained to fully represent the input data and minimize the generalization error; then, it learns to forget the irrelevant details by compressing the representation of the input.”