![[2602.19872] GOAL: Geometrically Optimal Alignment for Continual Generalized Category Discovery](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.19872] GOAL: Geometrically Optimal Alignment for Continual Generalized Category Discovery
Summary
The paper presents GOAL, a framework for Continual Generalized Category Discovery (C-GCD) that enhances class discovery while minimizing forgetting through a fixed geometric structure.
Why It Matters
As machine learning models increasingly face the challenge of learning from dynamic data, the ability to discover new categories without losing knowledge of existing ones is crucial. GOAL offers a significant improvement over previous methods, making it relevant for researchers and practitioners in AI and computer vision.
Key Takeaways
- GOAL introduces a unified framework for continual category discovery.
- It utilizes a fixed Equiangular Tight Frame (ETF) classifier for consistent learning.
- The method reduces forgetting by 16.1% compared to previous approaches.
- It improves novel class discovery by 3.2%, enhancing overall model performance.
- The framework supports both labeled and unlabeled data effectively.
Computer Science > Computer Vision and Pattern Recognition arXiv:2602.19872 (cs) [Submitted on 23 Feb 2026] Title:GOAL: Geometrically Optimal Alignment for Continual Generalized Category Discovery Authors:Jizhou Han, Chenhao Ding, SongLin Dong, Yuhang He, Shaokun Wang, Qiang Wang, Yihong Gong View a PDF of the paper titled GOAL: Geometrically Optimal Alignment for Continual Generalized Category Discovery, by Jizhou Han and 6 other authors View PDF HTML (experimental) Abstract:Continual Generalized Category Discovery (C-GCD) requires identifying novel classes from unlabeled data while retaining knowledge of known classes over time. Existing methods typically update classifier weights dynamically, resulting in forgetting and inconsistent feature alignment. We propose GOAL, a unified framework that introduces a fixed Equiangular Tight Frame (ETF) classifier to impose a consistent geometric structure throughout learning. GOAL conducts supervised alignment for labeled samples and confidence-guided alignment for novel samples, enabling stable integration of new classes without disrupting old ones. Experiments on four benchmarks show that GOAL outperforms the prior method Happy, reducing forgetting by 16.1% and boosting novel class discovery by 3.2%, establishing a strong solution for long-horizon continual discovery. Comments: Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) Cite as: arXiv:2602.19872 [cs.CV] (or arXiv:2602.19872v1 [cs....