NeurIPS 2020

Rethinking the Value of Labels for Improving Class-Imbalanced Learning


Review 1

Summary and Contributions: The paper examines the usefulness of labels in an imbalanced setting. The authors reflect on the positive and negative aspects of using labels, notably when engaging in SSL and self-supervision.

Strengths: This topic is of much interest to the NeurIPS community. The paper combines theoretical justifcations and empirical evaluations.

Weaknesses: It is possible that self-supervision can re-inforce errors. The authors thus have to reflect on using this SSL learning algorithm in their paper. (I understand that other algorithms are used in the Appendix, but this is still the pillar of their contribution.)

Correctness: The authors compare their work to one other algorithm and ignore many other significant contributions in this area. As such, it is difficult to assess the full impact of the contribution.

Clarity: The paper is generally well-written, but the style is very condense, making it sometimes difficult to navigate.

Relation to Prior Work: This is a drawback of the paper; related work are generally ignored or skipped over, even though the list of references is very long.

Reproducibility: Yes

Additional Feedback: The authors are urged to better position their work w.r.t. the substantial works already done in the SSL commuinity.


Review 2

Summary and Contributions: This is a very interesting paper. The contribution of the paper does not rely on proposing a specific semi-supervised and/or self-supervised learning method. Instead, this paper systematically studies both semi-supervised learning (pseudo-labeling) and self-supervised learning in the scenario of class-imbalanced learning, theoretically and empirically. In theory side, the authors propose a few theorems under some simplified conditions to motivate why pseudo-labeling and self-supervised learning can help the class-imbalanced learning. Theorem 1 demonstrates that, in binary classification scenario, pseudo-labeling on more unlabeled data would definitely help; Theorem 3 says, self-supervised learning can improve linear classifier given more data. In term of empirical verification, the authors run quite a lot experiments, in different scenarios and also tried a few self-supervised learning algorithms. The experimental results aligned with the proposed theorems and common sense.

Strengths: The proposed theorems, though under simplified conditions, are very motivating and intuitively explain why data matters for class-imbalanced learning. For experiments, the quantitative evaluation and ablation study all supports the proposed theorems. The results using different loss, different self-supervised learning algorithms, across different data sets are overall consistent. The t-sne plots (figure 1 and 4) are very impressive.

Weaknesses: 1. The authors did not clearly define 'relevance score'. 2. The proposed theorems, do not address the 'relevance score' effect for either semi-supervised or self-supervised learning scenarios. 3. I guess section 3 is largely used to motivate section 4, but I am not sure if you have experiments showing that self-supervised learning is more resistant to 'non-relevant unlabeled data'. I guess self-supervised learning would still be affected if the unlabeled data is totally non-relevant. Like using sounds to improve speech recognition. 4. Title seems to be a little bit misleading, actually, in the SSL scenarios, pseudo labels (and also relevance score) matters; While in the self-supervised learning scenario, the amount of unlabeled data matters. 5. Though very motivating, the theorems need to be further developed to explain more complex models.

Correctness: Yes, the claims be validated by theorems and experiments.

Clarity: Yes, the paper is self-motivating and well-written

Relation to Prior Work: The paper cited recent works in semi-supervised learning, self-supervised learning and class-imbalance learning. There are also some recent semi-supervised learning papers the author can consider to cite, like 'Unsupervised Data Augmentation for Consistency Training'

Reproducibility: Yes

Additional Feedback: Read the authors' and other reviewers' feedback/comments. Intend to keep my evaluation unchanged.


Review 3

Summary and Contributions: This paper provides extensive analysis on how to utilize data and labels in class-imbalanced classification. Through theoretical analysis and empirical evaluations, the authors demonstrate the benefits and potentials of employing semi-supervised learning and self-supervised learning in class-imbalanced classification.

Strengths: 1. The paper focus on analyzing an interesting problem: class-imbalanced classification. The problem is related to commonly existing long-tail issues in many machine learning tasks. The paper provides insightful comments on the effect of available labels in class-imbalanced learning from two different aspects. The results could be of interest to even broader area of different applications. 2. The theoretical analysis is sound. Different factors are considered, such as the class distribution (imbalanceness) and the relevance between training and testing data. Their effects on the learnability and estimation accuracy are both analyzed. 3. Inspired by the theoretical analyzing results, the authors propose to employ pseduo-label strategy to enhance the classification accuracy by harnessing the unlabeled data. Furthermore, self-supervised learning techniques are applied to alleviate the class bias introduced by the imbalanced class distribution. Both methods are proved to be effective through empirical evaluations. 4. The experiments are thorough. Extensive results demonstrate the authors claims clearly.

Weaknesses: Since the paper focuses on the value of labels, it will be stronger if more analysis on effect of different label distributions under class-imbalanced setup. For example, for the long-tail classes, what label distribution is preferable, i.e., should they have same proportion as the data distribution, or they are better to be uniform? Questions like these may provide insightful instructions on how to collect labels for class-imbalanced learning, which I find useful for many real-world applications. --------- Updates: The authors response has addressed my concerns. I recommend acceptance for this paper.

Correctness: The theoretical analysis is correct and the empirical evaluations are solid in this paper.

Clarity: The paper is written excellent.

Relation to Prior Work: The paper cited adequate related works and clearly describe its own novelties.

Reproducibility: Yes

Additional Feedback:


Review 4

Summary and Contributions: The paper studies the problem of long-tailed recognition. It theoretically and empirically demonstrates that (1) in most of the cases, semi-supervised learning can help with long-tailed recognition, unless the unlabeled data is heavily imbalanced (2) the representation learned by self-supervised pretraining can improve class-imbalanced learning. The proposed self-supervised imbalanced learning framework (SSP) achieves new SOTA on a number of long-tailed recognition benchmarks: CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT and iNaturalist 2018.

Strengths: (1) If I understand correctly, the proposed SSP method follows a similar paradigm with cRT[24]: learning the representation and classifier in different stages. While cRT[24] relies on imbalanced labels to learn the representation, this paper proposes to pretrain the feature representation in a self-supervised manner. It shows that the representation learned by self-supervised pretraining can improve class-imbalanced learning (or at least is complementary to other SOTA methods, as shown in table 3), which is very interesting. (2) The paper shows that, in most of the cases, leveraging unlabeled data in a semi-supervised setting can improve long-tailed recognition, unless the unlabeled data is heavily imbalanced (p_{U} > 50). Although the experiments are conducted on small-scale benchmarks only (e.g. CIFAR-10-LT, SVHN-LT), the finding can still provide some guidance for collecting unlabeled data for semi-supervised learning, especially when the available data is imbalanced.

Weaknesses: The paper claims that, the semi-supervised experiment demonstrates the value of labeled data. This claim looks odd. The labeled data is the “control variable” (remains unchanged) of the experiment, while the unlabeled data is the ‘independent variable’ of the experiment. I believe any conclusion drawn from the experiment should be associated with the “independent variable”. In other words, the experiment only demonstrates the value of unlabeled data for long-tailed recognition.

Correctness: The proposed method is technically sound and interesting. The claim that the semi-supervised experiment demonstrates the value of labeled data looks odd. Please refer to the weakness part.

Clarity: The paper is well written.

Relation to Prior Work: I feel like the proposed SSP method shares some similarities with cRT[24]. It would be better if more discussion on the difference between cRT[24] and the proposed SSP is included.

Reproducibility: Yes

Additional Feedback: -------------------------- Post-rebuttal: The rebuttal resolves some of my concerns. I’d like to keep my score (6), and vote for acceptance.