Aligned Contrastive Loss for Long-Tailed Recognition

2025-06-01

AI Summary

In long-tailed recognition, supervised contrastive learning suffers from degraded generalization due to gradient conflicts and imbalanced attraction/repulsion between positive and negative pairs—especially for tail classes. To address this, we propose Alignment Contrastive Learning (ACL), the first framework to theoretically analyze gradient dynamics in long-tailed settings, revealing the root cause of performance collapse. Based on this analysis, ACL introduces an explicit gradient alignment strategy that mitigates gradient imbalance for tail-class positive/negative pairs. Furthermore, ACL integrates multi-view representation alignment with long-tail-aware data augmentation and re-sampling. Extensive experiments on long-tailed CIFAR, ImageNet, Places, and iNaturalist demonstrate state-of-the-art performance: ACL achieves significant gains in tail-class accuracy and exhibits superior generalization stability compared to existing contrastive learning methods.

Technology Category

• Computer Vision: Representation Learning for Vision
• Machine Learning: Multi-instance/Multi-view Learning
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Web Mining and Content Analysis: Large pretrained models with web data

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.

Problem

• Addresses long-tailed recognition using Aligned Contrastive Learning
• Identifies gradient conflicts in supervised contrastive learning
• Improves model generalization across multiple benchmark datasets

Innovation

• Aligned Contrastive Learning for long-tailed recognition
• Addresses gradient conflicts in contrastive learning
• Improves performance on multiple benchmark datasets