International Journal of Computer Vision Компьютерное зрение и распознавание образов

Saving for the Future: Enhancing Generalization via Partial Logic Regularization

Tan, Zhaorui, Hu, Yijie, Yang, Xi, Wang, Qiufeng, Nguyen, Anh, Huang, Kaizhu
Springer Nature · 2026-11-02 · DOI 10.1007/s11263-026-03105-6

Аннотация

Generalization remains a significant challenge in visual classification tasks, particularly in handling unknown classes in real-world applications. Existing research focuses on the class discovery paradigm, which tends to favor known classes, and the incremental learning paradigm, which suffers from catastrophic forgetting. Recent approaches, such as the L-Reg technique, employ logic-based regularization to enhance generalization but are bound by the necessity of fully defined logical formulas, limiting flexibility for unknown classes. This paper introduces PL-Reg, a novel partial-logic regularization term that allows models to reserve space for undefined logic formulas, improving adaptability to unknown classes. Specifically, it is demonstrated that tasks involving unknown classes can be effectively explained using partial logic. An idealized analysis is also presented, motivating why partial logic may help preserve generalization capacity for unknown classes. We evaluate PL-Reg through extensive experiments on Generalized Category Discovery, Multi-Domain Generalized Category Discovery, long-tailed Class Incremental Learning, and Continual Category Discovery, and observe improvements across the evaluated settings, with gains of varying magnitudes. Our results highlight the effectiveness of partial logic in tackling challenges related to unknown classes.

Читать у издателя →