Contrastive Representation Learning with Trainable Augmentation Channel

TL;DR

This paper introduces a novel contrastive learning framework that adaptively learns augmentation strategies to prevent representation collapse, improving the robustness of learned image representations.

Contribution

It formalizes a stochastic encoding process with a data-dependent augmentation distribution, addressing the issue of augmentation-induced information loss in contrastive learning.

Findings

Learned augmentation distributions prevent representation collapse.

The infoMax objective guides adaptive augmentation selection.

Enhanced robustness of image representations in contrastive learning.

Abstract

In contrastive representation learning, data representation is trained so that it can classify the image instances even when the images are altered by augmentations. However, depending on the datasets, some augmentations can damage the information of the images beyond recognition, and such augmentations can result in collapsed representations. We present a partial solution to this problem by formalizing a stochastic encoding process in which there exist a tug-of-war between the data corruption introduced by the augmentations and the information preserved by the encoder. We show that, with the infoMax objective based on this framework, we can learn a data-dependent distribution of augmentations to avoid the collapse of the representation.