Dynamic Fission and Fusion of Neural Network Weights for Class Incremental Learning With Adaptive Regularization Strength
IEEE Transactions on Artificial Intelligence · 2026-11 · 7 (11) · 6385-6398
DOI: 10.1109/tai.2026.3731365 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Transactions on Artificial Intelligence, 2026-11. Verfasst von Jingang Lai und 2 weiteren Autor:innen. Die Arbeit nennt 33 Literaturangaben. Gefördert von: National Science and Technology Major Project of China, National Natural Science Foundation of China, Key Research and Development Project of Hubei Province of China.
Abstract (Originalsprache)
Continual learning (CL) methods for mitigating catastrophic forgetting in long-sequence task learning face challenges such as limited storage capacity, parameter explosion, and rigid regularization mechanisms. Meanwhile, class incremental learning (CIL) is considered to be the most challenging scenario in CL. To address these issues, this article proposes a novel regularization-based CIL approach that can further enhance the model’s knowledge retention and learning efficiency without increasing storage overhead. First, we design an adaptive weight fission-and-fusion strategy for the neural network output layer, which decomposes knowledge into multiple independent weight matrices to achieve distributed storage and improve the model’s adaptability to long-sequence tasks. Second, a posterior optimization objective for CIL is derived based on Bayesian theories. Theoretical analysis of the upper bound on forgetting is also provided to enhance the reliability and efficiency of the proposed method. Furthermore, an adaptive regularization mechanism driven by task similarity is proposed, which can dynamically adjust the regularization strength for each old task that has been learned previously, achieving fine-grained balance between model stability and plasticity. Experimental results demonstrate that the proposed method in this article outperforms many existing approaches across multiple benchmark datasets.
Förderung
- National Science and Technology Major Project of China
- National Natural Science Foundation of China
- Key Research and Development Project of Hubei Province of China
Zitieren
APA
Lai, J., Zhou, W., & Zeng, Z. (2026). Dynamic Fission and Fusion of Neural Network Weights for Class Incremental Learning With Adaptive Regularization Strength. IEEE Transactions on Artificial Intelligence, 7(11), 6385-6398. https://doi.org/10.1109/tai.2026.3731365
Vancouver
Lai J, Zhou W, Zeng Z. Dynamic Fission and Fusion of Neural Network Weights for Class Incremental Learning With Adaptive Regularization Strength. IEEE Transactions on Artificial Intelligence. 2026;7(11):6385-6398. doi:10.1109/tai.2026.3731365
BibTeX
@article{Lai2026,
author = {Lai, Jingang and Zhou, Wei and Zeng, Zhigang},
title = {Dynamic Fission and Fusion of Neural Network Weights for Class Incremental Learning With Adaptive Regularization Strength},
journal = {IEEE Transactions on Artificial Intelligence},
year = {2026},
volume = {7},
number = {11},
pages = {6385--6398},
doi = {10.1109/tai.2026.3731365}
}
EUVIMED