Continuous Contrastive Learning for Long-Tailed Semi-Supervised Recognition

Summary

The paper proposes a novel method to address the long-tail problem in semi-supervised learning by leveraging continuous contrastive learning on both labeled and unlabeled samples to improve model performance, particularly for minority classes.

Strengths

- Novelty: The paper introduces a novel approach to integrate contrastive learning with unlabeled samples into SSL frameworks, which is relatively unexplored in existing literature. - Technical Soundness: The theoretical formulation of learning good representations and the unified framework of many existing long-tail learning methods are well-explained and logically sound. - Experiments: The experiments are comprehensive, covering various imbalanced ratios and labeled ratios on CIFAR-10-LT and CIFAR-100-LT datasets. The results are statistically significant and show clear improvements over baseline methods. - Clarity: The paper is well-structured and clearly written. The introduction, methodology, and experimental sections are easy to follow, and the figures and tables effectively illustrate the results.

Weaknesses

- In line 189, it is unclear why selecting pseudo-labels using energy score can ensure better model calibration.

Questions

A minor suggestion: the paper can include some demonstrations of unlabeled images with their associated continuous pseudo-labels, including both reliable and smoothed versions. This can help readers identify cases in which the model tends to make mistakes.

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have adequately addressed the limitations and potential negative societal impact of their work.

Summary

The author proposes a probabilistic framework that unifies many recent proposals in long-tail learning. Specifically, for long-tailed semi-supervised learning, a continuous contrastive learning method based on reliable and smoothed pseudo-labels to address confirmation bias and improve the quality of learned representations is proposed. The experiments demonstrate their effectiveness.

Strengths

1) The author notices an important problem that existing approaches often assume the equivalence of the unlabeled data distribution with the labeled data, which is more practical. 2) The proposed probabilistic framework which unifies many recent proposals in long-tail learning is interesting. 3) The paper is well-written and easy to follow.

Weaknesses

1) The author should discuss explicitly how the proposed method solves the unlabelled data for diverse label distributions of unlabeled data and why other methods do not. 2) How to interpret " Furthermore, these methods primarily focus on correcting model outputs without delving into the role of representation learning in improving performance." line 33-34. 3) The author may give an analysis of the relationship between reliable pseudo-labels and smoothed pseudo-labels and how they play a different role in semi-supervised learning. 4) The time and space complexity should be analyzed in experiments between the proposed method and other methods.

Questions

see weakness

Rating

6

Confidence

2

Soundness

3

Presentation

2

Contribution

3

Limitations

yes

Summary

This paper proposes a novel contrastive learning method for long-tailed semi-supervised learning (LTSSL). The method is motivated by variational information bottleneck for learning good representations and extends to unlabeled data using continuous pseudo-labels. This paper showcases strong empirical results on multiple LTSSL datasets and label distributions.

Strengths

1.The proposed method is well-motivated. Previous works in LTSSL generally make contributions in learning balanced classifiers, while this paper examines representation learning. 2.This paper summarizes many existing methods for long-tail learning in a unified framework. It is interesting to show that the equivalence between supervised contrastive learning and Gaussian kernel estimation. 3.This paper proposes a new LTSSL method by optimizing the contrastive loss using two types of continuous pseudo-labels. The proposed method is novel in LTSSL. 4.Extensive experiments showcase that the proposed method achieves best results on all datasets. The paper also provides extensive in-depth studies to understand the effectiveness of the method.

Weaknesses

1.In Table 1, the paper states that long-tail learning methods need to avoid none class samples in each min-batch during training. However, it is unclear how this issue is tackled in the proposed method CCL. 2.The training time of the proposed method is not reported. It is suggested to compare the training time with ACR.

Questions

Please see Weaknesses.

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors have adequately addressed the limitations and potential negative societal impact of their work.

Summary

This paper tackles the long-tailed semi-supervised learning problem. It first reviews recent works with a novel probabilistic framework. Based on this, it proposes a continuous contrastive learning method, CCL, to extend the framework to unlabeled data with pseudo-labels. Experiments show that it outperforms all recent works on the popular long-tailed semi-supervised learning benchmarks.

Strengths

Overall, this paper is good. 1. With a framework from information theory perspective, it unified different approaches on the long-tailed learning topic. 2. It systematically studied the key ideas to tackle the long-tailed learning and semi-supervised learning tasks. 3. By proposing CCL, it unifies different ideas from long-tailed learning and semi-supervised learning, and the result is an effective approach to solve the LTSSL problem.

Weaknesses

I only have one question: In Eq. 22, it mentioned L_{cls}, which is also mentioned in Line 171. Is that the same as P^{cls} in Eq. 15?

Questions

See weaknesses In Eq. 22, it mentioned L_{cls}, which is also mentioned in Line 171. Is that the same as P^{cls} in Eq. 15?

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

Limitations and potential impacts are addressed in Appendix,

Keep the original ratings

Thanks for your response. I would like to keep the original ratings.