EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

GradKAN: Enhanced Kolmogorov–Arnold Networks With Adaptive Knot Allocation

Kursad Arslan, Emrah Donmez

IEEE Access · 2026 · 14 · 112669-112680

DOI: 10.1109/access.2026.3716130 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Kursad Arslan und 1 weiteren Autor:innen. Die Arbeit nennt 29 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The basic Kolmogorov-Arnold network model uses a uniform knot allocation for B-spline functions. This static knot allocation leads to knot waste and low parameter efficiency, particularly when modeling functions with sharp local variations and high-frequency dynamics. This study proposes a new architecture, the Gradient-Adaptive Kolmogorov-Arnold network model, which adaptively adjusts knot positions based on local error gradients without increasing the total parameter cost. In the proposed method, the magnitudes of the network’s error gradients are first computed and then filtered using an exponential moving average to prevent training instabilities that could arise from dynamic grid updates. These gradients are then converted into a probability density function, and a cumulative distribution function is derived. Using the inverse cumulative distribution function transformation, the knots, which are initially uniformly distributed, are adaptively relocated to regions where the error is concentrated. The performance of the proposed architecture has been comprehensively evaluated on a dataset using 1D high-frequency synthetic functions, 2D complex surface problems, physics-informed scenarios, and Feynman equations. Experiments on these scenarios, which involve various dynamics, have demonstrated that the Gradient-adaptive Kolmogorov-Arnold network improves parameter efficiency compared to the Base Kolmogorov-Arnold network model. It has been observed that the proposed approach overcomes the regional ineffectiveness issue while using the same parameter cost and significantly reduces error rates.

Zitieren

APA

Arslan, K., & Donmez, E. (2026). GradKAN: Enhanced Kolmogorov–Arnold Networks With Adaptive Knot Allocation. IEEE Access, 14, 112669-112680. https://doi.org/10.1109/access.2026.3716130

Vancouver

Arslan K, Donmez E. GradKAN: Enhanced Kolmogorov–Arnold Networks With Adaptive Knot Allocation. IEEE Access. 2026;14:112669-112680. doi:10.1109/access.2026.3716130

BibTeX
@article{Arslan2026,
  author = {Arslan, Kursad and Donmez, Emrah},
  title = {GradKAN: Enhanced Kolmogorov–Arnold Networks With Adaptive Knot Allocation},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {112669--112680},
  doi = {10.1109/access.2026.3716130}
}
Kostenlos starten →