Beyond Classification: Taxonomic Discovery of Novel Categories in Open-World Settings

Yong Wu, Yang Wang, Sanqing Qu, Fan Lu, Junqiao Zhao, Zhongcong Xu, Ya Wu, Ziqiao Wang, Guang Chen

Keywords: Generalized Category Discovery, Taxonomic Classification, Meta-Learning, Open-World Setting

Abstract: Generalized Category Discovery (GCD) has emerged as an important open-world learning problem that aims to automatically cluster partially labeled data. While significant progress has been made in GCD, existing methods focus solely on discovering new categories without capturing their semantic meaning that is a crucial capability for real-world applications. To address this limitation, we propose a novel setting called Taxonomic Discovery of Novel Categories in Open-World Setting (Taxo-GCD) that incorporates hierarchical classification into the GCD framework. Our approach not only identifies novel categories but also determines their taxonomic attributes by learning to recognize concepts at different hierarchical levels. Unlike traditional GCD methods, we develop a parametric classification approach that combines a hierarchical classifier with a unified learner network, further enhanced through a meta-learning architecture. This architecture employs a bi-level optimization scheme with inner update using self-supervised loss for fast test-time adaptation and outer updates for overall optimization. Extensive experiments demonstrate that Taxo-GCD achieves state-of-the-art performance on three standard visual recognition benchmarks, advancing open-world recognition by enabling semantic understanding of novel categories through their taxonomic relationships.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 4378