SHADE: Information Based Regularization for Deep Learning

TL;DR

SHADE is a novel information-theoretic regularization method for deep learning that improves classification performance by explicitly decoupling invariant representation learning from data fitting.

Contribution

It introduces a prior based on conditional entropy and derives a stochastic, tractable regularizer compatible with deep neural networks.

Findings

SHADE outperforms common regularizers on standard architectures

The method effectively decouples invariant representation learning from data fitting

Empirical results show improved classification accuracy

Abstract

Regularization is a big issue for training deep neural networks. In this paper, we propose a new information-theory-based regularization scheme named SHADE for SHAnnon DEcay. The originality of the approach is to define a prior based on conditional entropy, which explicitly decouples the learning of invariant representations in the regularizer and the learning of correlations between inputs and labels in the data fitting term. Our second contribution is to derive a stochastic version of the regularizer compatible with deep learning, resulting in a tractable training scheme. We empirically validate the efficiency of our approach to improve classification performances compared to common regularization schemes on several standard architectures.