Learning deep representations by mutual information estimation and maximization

TLDR

The authors aim to learn unsupervised representations by maximizing mutual information between inputs and a deep neural network encoder. Their method maximizes this mutual information and further shapes the representation by adversarially matching it to a prior distribution. They show that incorporating input locality improves downstream task suitability, that DIM outperforms many unsupervised methods and rivals supervised learning on classification, and that it opens new avenues for flexible representation‑learning objectives.

Abstract

In this work, we perform unsupervised learning of representations by maximizing mutual information between an input and the output of a deep neural network encoder. Importantly, we show that structure matters: incorporating knowledge about locality of the input to the objective can greatly influence a representation's suitability for downstream tasks. We further control characteristics of the representation by matching to a prior distribution adversarially. Our method, which we call Deep InfoMax (DIM), outperforms a number of popular unsupervised learning methods and competes with fully-supervised learning on several classification tasks. DIM opens new avenues for unsupervised learning of representations and is an important step towards flexible formulations of representation-learning objectives for specific end-goals.