MTMC: Generalized Category Discovery via Maximum Token Manifold Capacity

Luyao Tang, Kunze Huang, Xinghao Ding, Xiaotong Tu, Chaoqi Chen, Yue Huang

Keywords: generalized category discovery, deep cluster, manifold capacity

Abstract: Identifying previously unseen data is crucial for enhancing the robustness of deep learning models in the open world. Generalized category discovery (GCD) is a representative problem that requires clustering unlabeled data that includes known and novel categories. Current GCD methods mostly focus on minimizing intra-cluster variations, often at the cost of manifold capacity, thus limiting the richness of within-class representations. In this paper, we introduce a novel GCD approach that emphasizes maximizing the token manifold capacity (MTMC) within class tokens, thereby preserving the diversity and complexity of the data's intrinsic structure. Specifically, MTMC's efficacy is fundamentally rooted in its ability to leverage the nuclear norm of the singular values as a quantitative measure of the manifold capacity. MTMC enforces a richer and more informative representation within the manifolds of different patches constituting the same sample. MTMC ensures that, for each cluster, the representations of different patches of the same sample are compact and lie in a low-dimensional space, thereby enhancing discriminability. By doing so, the model could capture each class's nuanced semantic details and prevent the loss of critical information during the clustering process. MTMC promotes a comprehensive, non-collapsed representation that improves inter-class separability without adding excessive complexity.

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.

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: 3499