Status : Verified
| Personal Name | Guevarra, Arjeus A. |
|---|---|
| Resource Title | Continual Learning in Kolmogorov-Arnold Networks using Automatic Grow-and-Prune Approach |
| Date Issued | 13 June 2025 |
| Abstract | Kolmogorov-Arnold Networks (KANs) replace fixed activation functions with learnable spline-based functions on network edges, enabling flexible, interpretable function approximation inspired by the Kolmogorov-Arnold representation theorem. This work explores KANs for class-incremental learning (CIL), where models learn new classes sequentially without access to prior data. We propose a dynamic KAN architecture that grows and prunes both network structure and spline grids to optimize generalization across incremental tasks. To en hance stability, we introduce magnitude-based regularization that limits large changes in spline coefficients, and apply a gradual early-stopping strategy to reduce overfit ting. Experiments show our dynamic KAN performs comparably to exemplar-based state-of-the-art methods, and outperforms others in exemplar-free settings. An abla tion study confirms the effectiveness of the grow-and-prune method, which supports knowledge retention via subnetwork freezing. While magnitude-based regularization can further improve results, it is sensitive to hyperparameter settings. Results also confirm a strong correlation between exemplar quantity and model accuracy, underscoring the challenge of catastrophic forgetting with limited exem plars. Overall, KANs demonstrate strong potential as a flexible, adaptive framework for continual learning. |
| Degree Course | Master of Science in Computer Science |
| Language | English |
| Keyword | Continual Learning; Deep Learning; Machine Learning; Kolmogorov-Arnold Networks |
| Material Type | Thesis/Dissertation |
Preliminary Pages
1,021.45 Kb
Category : F - Regular work, i.e., it has no patentable invention or creation, the author does not wish for personal publication, there is no confidential information.
Access Permission : Open Access