Kolmogorov-Arnold Networks
KAN: Kolmogorov-Arnold Networks · Also known as: KAN, Kolmogorov-Arnold
Kolmogorov-Arnold Networks (KAN) is a neural network architecture introduced by Liu et al. in 2024 that replaces linear transformations with learned univariate functions on edges. Inspired by the Kolmogorov-Arnold representation theorem, KAN achieves superior function approximation with fewer parameters than traditional MLPs, offering potential efficiency gains and improved interpretability.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
KAN is particularly valuable for scientific computing and symbolic regression where interpretability and parameter efficiency matter. It excels on low-dimensional problems where univariate function composition is natural. Use traditional MLPs for high-dimensional problems (images, language) where edge-based learning may be less natural. KAN is especially promising for physics-informed neural networks and scientific machine learning.
Strengths & limitations
- Achieves superior parameter efficiency compared to MLPs on many function approximation tasks
- Improved interpretability through direct visualization of learned univariate functions on edges
- Theoretically grounded in approximation theory through Kolmogorov-Arnold representation theorem
- Particularly effective for low-to-moderate dimensional problems and scientific computing
- May underperform MLPs on high-dimensional problems like vision and language tasks
- Spline computations add computational overhead per forward pass compared to simple linear transformations
- Limited applications demonstrated on large-scale benchmarks; effectiveness on diverse domains remains unclear
Frequently asked
What is the Kolmogorov-Arnold representation theorem and how does KAN relate to it?
The theorem states any continuous multivariate function on a compact domain can be represented as a superposition of univariate functions and addition. KAN implements this by placing learnable univariate functions (splines) on network edges. This provides theoretical justification for the architecture and suggests it should efficiently approximate multivariate functions through composition of simpler univariate components.
Why use splines instead of other learnable functions?
Splines (piecewise polynomials) are computationally efficient to evaluate and have strong approximation properties. They can represent arbitrary smooth functions through local polynomial pieces. Splines also enable easy visualization of learned univariate functions, supporting interpretability. Control point optimization is straightforward compared to other learnable function classes.
How does KAN compare to MLPs on typical benchmark tasks?
On standard vision and language benchmarks, KAN typically underperforms MLPs due to overhead from spline computations. However, on function approximation and symbolic regression, KAN achieves better parameter efficiency. KAN is most advantageous on low-to-moderate dimensional problems where univariate function composition is natural and parameter efficiency matters.
Sources
- Liu, Z., Wang, Y., Vaidya, S., Ruehle, F., Halverson, J., Soljačić, M., Hou, T. Y., & Tegmark, M. (2024). KAN: Kolmogorov-Arnold Networks. arXiv preprint arXiv:2404.19756. link ↗
How to cite this page
ScholarGate. (2026, June 3). KAN: Kolmogorov-Arnold Networks. ScholarGate. https://scholargate.app/en/deep-learning/kolmogorov-arnold-networks
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Mamba (State Space Model)Deep learning↔ compare
- Masked AutoencodersDeep learning↔ compare
- Neural Radiance Fields (NeRF)Deep learning↔ compare
- Vision TransformerDeep learning↔ compare