open access

https://hdl.handle.net/2115/100454

Dealing with limited data and uncertainty in deep learning using Kolmogorov–Arnold Neural Network

英語

Uncertainty

2026-07-01

Proceedings of the 10th International Symposium on Reliability Engineering and Risk Management (ISRERM 2026)

Accurate modelling of complex physical systems under limited data availability and inherent uncertainty remains a major challenge in the deep learning field. Conventional deep learning models rely heavily on large datasets and easily overfit when data is limited. Physics-informed neural networks (PINNs) address this issue by incorporating physical rules, but their performance is often limited by rigid loss balancing and is difficult to further improve. Recently, Kolmogorov–Arnold Networks (KANs) have shown improved expressiveness and interpretability. However, the reliance on B-spline bases introduces high computational cost and limits generalization, and uncertainty cannot be quantified. We propose a novel model based on the KAN structure that integrates Bayesian inference, physical constraint regularization, attention mechanism and Chebyshev polynomial-based functional representations. The capability and flexibility of the proposed approach are demonstrated by solving a benchmark problem from manufacture experiment from NASA Prognostic Data Repository. Comparative results with baseline models further demonstrate the advantages. The findings indicate that the proposed approach provides a robust and effective solution for dealing with small dataset in complex engineering systems.

会議発表論文

Version of Record