ProtoGCD: Unified and Unbiased Prototype Learning for Generalized Category Discovery
Shijie Ma, Fei Zhu, Xu-Yao Zhang, Cheng-Lin Liu
2025-04-09
Summary
This paper talks about ProtoGCD, an AI tool that helps computers sort data into categories—both known and new ones—by learning from examples and avoiding unfair biases towards old categories.
What's the problem?
Current systems either treat old and new categories separately, causing uneven accuracy, or use methods that create biased results and miss important patterns.
What's the solution?
ProtoGCD uses shared models for old and new categories, adjusts its learning to reduce mistakes, and adds checks to make sure it learns balanced and accurate patterns for all categories.
Why it matters?
This helps AI systems better organize real-world data, like sorting products in stores or medical records, by fairly recognizing both familiar and new types without extra human help.
Abstract
Generalized category discovery (GCD) is a pragmatic but underexplored problem, which requires models to automatically cluster and discover novel categories by leveraging the labeled samples from old classes. The challenge is that unlabeled data contain both old and new classes. Early works leveraging pseudo-labeling with parametric classifiers handle old and new classes separately, which brings about imbalanced accuracy between them. Recent methods employing contrastive learning neglect potential positives and are decoupled from the clustering objective, leading to biased representations and sub-optimal results. To address these issues, we introduce a unified and unbiased prototype learning framework, namely ProtoGCD, wherein old and new classes are modeled with joint prototypes and unified learning objectives, {enabling unified modeling between old and new classes}. Specifically, we propose a dual-level adaptive pseudo-labeling mechanism to mitigate confirmation bias, together with two regularization terms to collectively help learn more suitable representations for GCD. Moreover, for practical considerations, we devise a criterion to estimate the number of new classes. Furthermore, we extend ProtoGCD to detect unseen outliers, achieving task-level unification. Comprehensive experiments show that ProtoGCD achieves state-of-the-art performance on both generic and fine-grained datasets. The code is available at https://github.com/mashijie1028/ProtoGCD.