A Broad Study on the Transferability of Visual Representations with Contrastive Learning

TL;DR

This study comprehensively evaluates the transferability of contrastive learning-based visual representations across various domains and tasks, showing they outperform supervised models in transferability and contain more adaptable low/mid-level semantics.

Contribution

It provides a detailed analysis of contrastive learning's transferability, demonstrating its advantages over supervised methods and revealing the semantic properties of learned representations.

Findings

Contrastive approaches transfer better across domains and tasks.

Joint contrastive and cross-entropy objectives improve transferability.

Contrastive models encode more low/mid-level semantics.

Abstract

Tremendous progress has been made in visual representation learning, notably with the recent success of self-supervised contrastive learning methods. Supervised contrastive learning has also been shown to outperform its cross-entropy counterparts by leveraging labels for choosing where to contrast. However, there has been little work to explore the transfer capability of contrastive learning to a different domain. In this paper, we conduct a comprehensive study on the transferability of learned representations of different contrastive approaches for linear evaluation, full-network transfer, and few-shot recognition on 12 downstream datasets from different domains, and object detection tasks on MSCOCO and VOC0712. The results show that the contrastive approaches learn representations that are easily transferable to a different downstream task. We further observe that the joint objective of self-supervised contrastive loss with cross-entropy/supervised-contrastive loss leads to better transferability of these models over their supervised counterparts. Our analysis reveals that the representations learned from the contrastive approaches contain more low/mid-level semantics than cross-entropy models, which enables them to quickly adapt to a new task. Our codes and models will be publicly available to facilitate future research on transferability of visual representations.