APLA: Class-imbalanced Semi-supervised Learning with Adapative Pseudo-labeling and Loss Adjustment

Qian Gui, Baoning Niu

Keywords: semi-supervised learning, class-imbalanced learning, class-imbalanced semi-supervised learning

Abstract: Semi-supervised learning (SSL) can substantially improve the performance of deep neural networks by utilizing unlabeled data when labeled data is scarce. Existing SSL algorithms implicitly assume that the class distribution of labeled datasets and unlabeled datasets are balanced, which means the different classes have the same numbers of training samples. However, they can hardly perform well on minority classes(the classes with few training examples) when the class distribution of training data is imbalanced, since the pseudo-labels learned from unlabeled data tend to be biased toward majority classes(the classes with a large number of training examples). To alleviate this issue, we propose a method called Adaptive Pseudo-labeling and Loss Adjustment (APLA) for class-imbalanced semi-supervised learning (CISSL), which includes Class-Aware Pseudo-label Thresholding (CAPT) that can utilize the imbalanced unlabeled data by dynamically adjusting the threshold for selecting pseudo-labels, and Class-Aware Loss Adjustment (CALA) that can mitigate the bias in both supervised loss and unsupervised loss. According to the experiments, APLA can deliver much higher accuracy than benchmark methods under various CISSL scenarios.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning

TL;DR: We use Class-Aware Pseudo-label Thresholding and Class-Aware Loss Adjustment to improve the performance of existing SSL algorithm in Class-imbalanced setting.