Decoupled Contrastive Learning

Chun Hsiao Yeh, Cheng Yao Hong, Yen Chi Hsu, Tyng Luh Liu, Yubei Chen, Yann LeCun

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

Contrastive learning (CL) is one of the most successful paradigms for self-supervised learning (SSL). In a principled way, it considers two augmented “views” of the same image as positive to be pulled closer, and all other images as negative to be pushed further apart. However, behind the impressive success of CL-based techniques, their formulation often relies on heavy-computation settings, including large sample batches, extensive training epochs, etc. We are thus motivated to tackle these issues and establish a simple, efficient, yet competitive baseline of contrastive learning. Specifically, we identify, from theoretical and empirical studies, a noticeable negative-positive-coupling (NPC) effect in the widely used InfoNCE loss, leading to unsuitable learning efficiency concerning the batch size. By removing the NPC effect, we propose decoupled contrastive learning (DCL) loss, which removes the positive term from the denominator and significantly improves the learning efficiency. DCL achieves competitive performance with less sensitivity to sub-optimal hyperparameters, requiring neither large batches in SimCLR, momentum encoding in MoCo, or large epochs. We demonstrate with various benchmarks while manifesting robustness as much less sensitive to suboptimal hyperparameters. Notably, SimCLR with DCL achieves 68.2% ImageNet-1K top-1 accuracy using batch size 256 within 200 epochs pre-training, outperforming its SimCLR baseline by 6.4%. Further, DCL can be combined with the SOTA contrastive learning method, NNCLR, to achieve 72.3% ImageNet-1K top-1 accuracy with 512 batch size in 400 epochs, which represents a new SOTA in contrastive learning. We believe DCL provides a valuable baseline for future contrastive SSL studies.

Original languageEnglish (US)
Title of host publicationComputer Vision – ECCV 2022 - 17th European Conference, 2022, Proceedings
EditorsShai Avidan, Gabriel Brostow, Moustapha Cissé, Giovanni Maria Farinella, Tal Hassner
PublisherSpringer Science and Business Media Deutschland GmbH
Pages668-684
Number of pages17
ISBN (Print)9783031198083
DOIs
StatePublished - 2022
Event17th European Conference on Computer Vision, ECCV 2022 - Tel Aviv, Israel
Duration: Oct 23 2022Oct 27 2022

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume13686 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference17th European Conference on Computer Vision, ECCV 2022
Country/TerritoryIsrael
CityTel Aviv
Period10/23/2210/27/22

Keywords

  • Contrastive learning
  • Self-supervised learning

ASJC Scopus subject areas

  • Theoretical Computer Science
  • Computer Science(all)

Fingerprint

Dive into the research topics of 'Decoupled Contrastive Learning'. Together they form a unique fingerprint.

Cite this