Debiased Imbalanced Pseudo-Labeling for Generalized Category Discovery

Xinzi Cao, Xiawu Zheng, Feidiao Yang, Quanmin Liang, Guanhong Wang, Yutong Lu, Yonghong Tian

Keywords: Generalized Category Discovery, Pseudo-Labeling

Abstract:

Generalized Category Discovery (GCD) is a challenging task that aims to recognize seen and novel categories within unlabeled data by leveraging labeled data.
Designing a prototype classifier to identify unlabeled samples instead of relying on traditional time-consuming clustering is well recognized as a milestone in GCD.
However, we discover there exists a bias in this classifier: some seen categories are mistakenly classified as novel ones, leading to imbalanced pseudo-labeling during classifier learning.
Based on this finding, we identify the low discriminability between seen and novel prototypes as the key issue.
To address this issue, we propose DebiasGCD, an effective debiasing method that integrates *dynamic prototype debiasing* (DPD) and *local representation alignment* (LRA).
DPD dynamically maintains inter-prototype margins, encouraging the network to strengthen the learning of class-specific features and enhance prototype discrimination.
Additionally, LRA promotes local representation learning, enabling DPD to capture subtle details that further refine the understanding of class-specific features.
In this way, it successfully improves prototype discriminability and
generates more reliable predictions for seen classes.
Extensive experiments validate that our method effectively mitigates pseudo-labeling bias across all datasets, especially on fine-grained ones. For instance, it delivers a 10.7\% boost on `Old' classes in CUB. Our code is available at：https://anonymous.4open.science/r/DebiasGCD-34F0.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

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: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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.

Submission Number: 3711