Semi-supervised Long-tailed Recognition using Alternate Sampling

This study aims to train an algorithm with a training set that class distribution is imbalanced and it has both labeled and unlabeled sets. For the training, the authors borrowed the concept of decoupled learning of features and classifier. Specifically, the authors proposed an algorithm consists of three stage which separatively and iteratively learns features and classifier using alternative sampling and pseudo-labels. Experiments on two datasets verify effectiveness of the proposed algorithm.

This paper studies the semi-supervised long-tailed recognition problem, which aims to leverage unlabeled data to enhance performance on long-tailed recognition. An alternate sampling method is proposed to incorporate the decoupling in long-tailed recognition with semi-supervised learning. The improved performance is reported in the paper.

In this paper, the authors address long-tailed recognition in the semi-supervised learning paradigm. The authors claim two long-tailed recognition challenges: sample scarcity and imbalanced class distribution, which is solved by the available unlabeled data. This paper proposes an alternate sampling strategy and consistency regularization in a decoupling framework, i.e., iteratively optimizing representation learning and classifier learning. Experiments are conducted on CIFAR-LT and ImageNet-LT to validate the proposed method.

The semi-supervised long-tailed recognition problem is defined in this article, which also combines the long-tailed and semi-supervised problems in recognition tasks. To present a three-stage learning approach, the author integrates certain techniques from the previous long-tailed learning field and the semi-supervised learning field. Before the three stages of learning, the author, who refers to the previous methods, decouples feature embedding and classifier to initialize the model. Afterward, the model undergoes several rounds of learning, repeating the methods of these three stages in each round. In the first stage, the model generates pseudo labels on unlabeled data. In the second stage, the model uses both pseudo-label samples and labeled samples for training feature embedding. In the third stage, the model uses labeled samples to train the classifier. Among them, random sampling is used in the first and second stages, and class-balanced sampling is used in the third stage. This method is an extension of the decoupling method in the field of long-tailed learning on semi-supervised data, which helps to make model representation more generalized and classification more balanced. The author also demonstrated in the experiment the superior performance of such methods on extremely imbalanced data.

This paper studies semi-supervised long-tailed recognition problem, where the data exhibits a long-tailed imbalanced distribution, and there is unlabeled data that can be leveraged. To address this problem, the authors incorporate the idea of model decoupling and balanced sampling. Furthermore, it proposes a three-stage approach to learn from both labeled and unlabeled data. The empirical results demonstrate the effectiveness of the proposed method. The studied problem is interesting and a corresponding method is proposed. However, all of the reviewers are with negative comments. The proposed approach is a easy combination of previous methods. Moreover, there are already many long-tailed semi-supervised learning approaches, but the authors have not discussed and compared them. More concerns can be referred to from the reviewers' comments, such as unsatisfactory experiments, code reproducibility, and no theoretical ground. Moreover, the authors have not made any response to the reviewer' comments, and the concerns are not addressed at all. Therefore, I recommend rejecting this paper.

Reject