Federated few-shot class-incremental learning
Files
(Published version)
Date
2025
Authors
Ma'sum, M.A.
Pratama, M.
Liu, L.
Habibullah, H.
Kowalczyk, R.
Editors
Advisors
Journal Title
Journal ISSN
Volume Title
Type:
Conference paper
Citation
The Thirteenth International Conference on Learning Representations (ICLR 2025), 2025, pp.1-30
Statement of Responsibility
Conference Name
The Thirteenth International Conference on Learning Representations (ICLR 2025) (24 Apr 2025 : Singapore)
Abstract
This study proposes a challenging yet practical Federated Few-Shot Class-Incremental Learning (FFSCIL) problem, where clients only hold very few samples for new classes. We develop a novel Unified Optimized Prototype Prompt (UOPP) model to simultaneously handle catastrophic forgetting, over-fitting, and prototype bias in FFSCIL. UOPP utilizes task-wise prompt learning to mitigate task interference and over-fitting, unified static-dynamic prototypes to achieve a stability-plasticity balance, and adaptive dual heads for enhanced inferences. Dynamic prototypes represent new classes in the current few-shot task and are rectified to deal with prototype bias. Our comprehensive experimental results show that UOPP significantly outperforms state-of-the-art (SOTA) methods on three datasets with improvements up to 76% on average accuracy and 90% on harmonic mean accuracy respectively. Our extensive analysis shows UOPP robustness in various numbers of local clients and global rounds, low communication costs, and moderate running time. The source code of UOPP is publicly available at https://github.com/anwarmaxsum/FFSCIL.
School/Discipline
Dissertation Note
Provenance
Description
Access Status
Rights
Copyright 2025 ICLR (https://creativecommons.org/licenses/by/4.0/)