Researchers have explored the universality of Kolmogorov-Arnold Networks (KANs), demonstrating that a single non-affine edge function, combined with affine ones, is sufficient for deep KANs to be universal approximators. Further analysis shows that for KANs with exactly two hidden layers, universality depends on the non-polynomial nature of the edge function. Additionally, a new variant called Partition-of-Unity Gaussian KANs (PU-GKANs) has been introduced, utilizing Gaussian basis functions for improved stability and accuracy compared to spline-based activations. AI
Summary written by gemini-2.5-flash-lite from 2 sources. How we write summaries →
IMPACT New theoretical findings on KAN universality and a novel PU-GKAN variant may lead to more stable and accurate neural network architectures.
RANK_REASON Two arXiv papers published on April 26, 2026, detailing theoretical properties and new variants of Kolmogorov-Arnold Networks.