Aligned Contrastive Loss for Long-Tailed Recognition

TL;DR

This paper introduces an Aligned Contrastive Learning algorithm to improve long-tailed recognition, addressing gradient conflicts and imbalances in contrastive learning, and demonstrates state-of-the-art results across multiple datasets.

Contribution

The paper proposes a novel ACL algorithm that corrects gradient conflicts in contrastive learning for long-tailed recognition, achieving superior performance.

Findings

ACL outperforms existing methods on long-tailed datasets

Contrastive learning benefits from multi-view training but has limitations

Gradient conflicts are a key issue in supervised contrastive learning

Abstract

In this paper, we propose an Aligned Contrastive Learning (ACL) algorithm to address the long-tailed recognition problem. Our findings indicate that while multi-view training boosts the performance, contrastive learning does not consistently enhance model generalization as the number of views increases. Through theoretical gradient analysis of supervised contrastive learning (SCL), we identify gradient conflicts, and imbalanced attraction and repulsion gradients between positive and negative pairs as the underlying issues. Our ACL algorithm is designed to eliminate these problems and demonstrates strong performance across multiple benchmarks. We validate the effectiveness of ACL through experiments on long-tailed CIFAR, ImageNet, Places, and iNaturalist datasets. Results show that ACL achieves new state-of-the-art performance.