RBF-KAN: Radial Basis Function–Kolmogorov–Arnold Network

This paper proposes a novel neural network architecture RBF-KAN, integrating the theoretical robustness of Kolmogorov-Arnold Networks (KANs) with the localized features of Radial Basis Function (RBF) networks. To address the low computational efficiency and poor interpretability of original KANs caused by B-spline basis functions, RBF-KAN adopts RBF as the core function for activation parameterization, while retaining KAN’s theoretical guarantees. Experiments on function approximation, basic classification datasets (Moon, Concentric circles, etc.), and the real-world large-scale Adult dataset demonstrate that RBF-KAN achieves competitive accuracy with significantly enhanced efficiency compared to KAN and FourierKAN. Specifically, it reduces parameters by 33%–65% and accelerates inference/training by 65%–99% while maintaining comparable accuracy (e.g., 0.8357 vs. KAN’s 0.8157 on Adult dataset). It exhibits excellent stability and generalization, especially in high-dimensional and large-scale scenarios. This work highlights RBF-KAN’s strengths, limitations, and future directions, underscoring its potential in tackling complex function approximation and classification challenges.