Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Class-Aware Contrastive Semi-Supervised Learning (2203.02261v3)

Published 4 Mar 2022 in cs.CV

Abstract: Pseudo-label-based semi-supervised learning (SSL) has achieved great success on raw data utilization. However, its training procedure suffers from confirmation bias due to the noise contained in self-generated artificial labels. Moreover, the model's judgment becomes noisier in real-world applications with extensive out-of-distribution data. To address this issue, we propose a general method named Class-aware Contrastive Semi-Supervised Learning (CCSSL), which is a drop-in helper to improve the pseudo-label quality and enhance the model's robustness in the real-world setting. Rather than treating real-world data as a union set, our method separately handles reliable in-distribution data with class-wise clustering for blending into downstream tasks and noisy out-of-distribution data with image-wise contrastive for better generalization. Furthermore, by applying target re-weighting, we successfully emphasize clean label learning and simultaneously reduce noisy label learning. Despite its simplicity, our proposed CCSSL has significant performance improvements over the state-of-the-art SSL methods on the standard datasets CIFAR100 and STL10. On the real-world dataset Semi-iNat 2021, we improve FixMatch by 9.80% and CoMatch by 3.18%. Code is available https://github.com/TencentYoutuResearch/Classification-SemiCLS.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (9)
  1. Fan Yang (878 papers)
  2. Kai Wu (134 papers)
  3. Shuyi Zhang (12 papers)
  4. Guannan Jiang (24 papers)
  5. Yong Liu (721 papers)
  6. Feng Zheng (117 papers)
  7. Wei Zhang (1489 papers)
  8. Chengjie Wang (178 papers)
  9. Long Zeng (39 papers)
Citations (87)

Summary

Enhancing Semi-Supervised Learning with Class-Aware Contrastive Learning

Introduction

Semi-Supervised Learning (SSL) has witnessed substantial progress, especially in leveraging vast unlabeled data to improve learning accuracy. However, pseudo-label-based SSL, a prevalent approach in this domain, often encounters hurdles such as confirmation bias and struggles with the noisy, out-of-distribution data prevalent in real-world applications. The Class-Aware Contrastive Semi-Supervised Learning (CCSSL) method introduces an innovative approach to mitigate these challenges by integrating a class-aware contrastive learning algorithm within the SSL framework.

Class-Aware Contrastive Semi-Supervised Learning (CCSSL)

CCSSL aims to address the limitations of traditional SSL methods by discriminating between in-distribution and out-of-distribution data through a class-aware clustering and image-wise contrasting approach. Specifically, it:

  • Applies class-wise clustering for in-distribution data, facilitating its incorporation into downstream tasks.
  • Utilizes image-level contrastive learning for out-of-distribution data to enhance generalization and mitigate noise.
  • Integrates a target re-weighting mechanism to emphasize learning from clean labels while diminishing the impact of noisy labels.

The efficacy of CCSSL is demonstrated through significant performance improvements on standard datasets such as CIFAR100 and STL10, along with notable gains on the challenging real-world dataset Semi-iNat 2021, where it surpasses traditional SSL methods including FixMatch and CoMatch.

Methodology

CCSSL comprises two principal components:

  1. Semi-Supervised Module: Any end-to-end pseudo-label-based SSL method can serve as the semi-supervised module within CCSSL. This module is responsible for generating pseudo labels and training on them, despite the inherent noise.
  2. Class-Aware Contrastive Module: This module introduces contrastive learning based on class-aware clustering and image-wise contrasting. It not only alleviates noise from out-of-distribution data but also enhances the model's performance by focusing on high-confidence, in-distribution samples through a re-weighting mechanism.

CCSSL's approach ensures that SSL methods can effectively leverage the vast amounts of unlabeled data encountered in real-world settings while minimizing the adverse effects of noise and out-of-distribution data.

Experimental Results

CCSSL's performance was rigorously evaluated across various datasets, demonstrating its superior ability to mitigate confirmation bias and enhance model robustness. Notably:

  • On CIFAR100 and STL10, CCSSL significantly outperformed state-of-the-art SSL methods by effectively addressing noise from out-of-distribution data.
  • On the real-world dataset Semi-iNat 2021, CCSSL achieved remarkable improvements over traditional SSL methods, showcasing its practical utility in real-world scenarios.

Furthermore, CCSSL was found to accelerate the convergence of SSL methods, achieving optimal performance in fewer training epochs.

Implications and Future Directions

The introduction of CCSSL marks a noteworthy advancement in the field of semi-supervised learning, particularly in addressing the challenges posed by noisy and out-of-distribution data. Its ability to seamlessly integrate with existing SSL methods without extensive modifications highlights its potential as a versatile tool for enhancing SSL applications.

The promising results achieved by CCSSL on both in-distribution and real-world datasets advocate for its broader adoption and further exploration in various SSL contexts. Future research could explore the integration of CCSSL with other machine learning paradigms and its adaptation to a wider range of real-world applications.

Conclusion

CCSSL represents a significant step forward in semi-supervised learning, providing a robust framework for enhancing the effectiveness of SSL methods in the presence of noisy and out-of-distribution data. Its innovative class-aware contrastive learning approach, combined with a re-weighting mechanism, offers a powerful solution to the challenges of confirmation bias and data noise that often hinder SSL applications. The demonstrated success of CCSSL across multiple datasets underscores its potential to improve the practicality and performance of semi-supervised learning in a variety of settings.