Computing the Information Content of Trained Neural Networks

TL;DR

This paper develops methods to measure the true information content stored in neural networks, revealing how it relates to generalisation and architecture, and providing bounds that align with empirical observations.

Contribution

It introduces a consistent estimator and a closed-form upper bound for the information content in infinitely wide neural networks, linking architecture, data, and generalisation.

Findings

Bounds are non-vacuous and empirically correlated with generalisation.

Information content can be analytically controlled by network architecture and data.

The approach extends previous Gaussian-based analyses to infinite-width networks.

Abstract

How much information does a learning algorithm extract from the training data and store in a neural network's weights? Too much, and the network would overfit to the training data. Too little, and the network would not fit to anything at all. Na\"ively, the amount of information the network stores should scale in proportion to the number of trainable weights. This raises the question: how can neural networks with vastly more weights than training data still generalise? A simple resolution to this conundrum is that the number of weights is usually a bad proxy for the actual amount of information stored. For instance, typical weight vectors may be highly compressible. Then another question occurs: is it possible to compute the actual amount of information stored? This paper derives both a consistent estimator and a closed-form upper bound on the information content of infinitely wide neural networks. The derivation is based on an identification between neural information content and the negative log probability of a Gaussian orthant. This identification yields bounds that analytically control the generalisation behaviour of the entire solution space of infinitely wide networks. The bounds have a simple dependence on both the network architecture and the training data. Corroborating the findings of Valle-P\'erez et al. (2019), who conducted a similar analysis using approximate Gaussian integration techniques, the bounds are found to be both non-vacuous and correlated with the empirical generalisation behaviour at finite width.