Continual generalized category discovery via enhancing representation learning

Weng, Jianqiao, Karakuş, Oktay ORCID: https://orcid.org/0000-0001-8009-9319 and Li, Yuhua ORCID: https://orcid.org/0000-0003-2913-4478 2026. Continual generalized category discovery via enhancing representation learning. Presented at: UKAIRS 2026, Edinburgh, Scotland, UK, 24-25 November 2026.

PDF - Accepted Post-Print Version Download (588kB)

Abstract

Continual generalized category discovery enables machine learning models to efficiently discover and integrate potentially novel categories emerging in evolving data streams, while maintaining robust recognition and classification of known categories. Although prior research has demonstrated the effectiveness of adopting category boundary exemplars in task-incremental learning, the significance of boundary exemplars and strategies for leveraging discriminative characteristics in continual generalized category discovery remain insufficiently explored. This work aims to sample and preserve both central exemplars, which exhibit representative characteristics, and peripheral exemplars, which capture discriminative features. The representation space is regularized and adapted through similarity constraints between central or peripheral exemplars and the corresponding category prototypes. Additionally, this work proposes an adaptive threshold for estimating the novelty of observations from potentially new categories. Preliminary evaluation on the CIFAR-100 dataset demonstrates effective recognition performance on known categories and considerable ability to discover and integrate potentially novel categories.

Item Type:	Conference or Workshop Item - unpublished
Status:	Unpublished
Schools:	Schools > Computational & Mathematical Sciences
Date of First Compliant Deposit:	13 August 2026
Last Modified:	13 Aug 2026 09:45
URI:	https://orca.cardiff.ac.uk/id/eprint/188957

Actions (repository staff only)

Edit Item