Geometric Kolmogorov Superposition Representation of group invariant function for computational science

Francesco Alesiani, Takashi Maruyama, Henrik Christiansen, Viktor Zaverkin

Keywords: machine learning, geometry, invariance, equivariance, symmetry, physics

TL;DR: We explore the extension of KAN to model group symmetries, as rotation, reflection, translation, Lorentz action, permuation or generic GL

Abstract:

The Kolmogorov-Arnold Theorem (KAT), or more generally, the Kolmogorov Superposition Theorem (KST), establishes that any non-linear multivariate function can be exactly represented as a finite superposition of non-linear univariate functions. Unlike the universal approximation theorem, which provides only an approximate representation without guaranteeing a fixed network size, KST offers a theoretically exact decomposition. The Kolmogorov-Arnold Network (KAN) was introduced as a trainable model to implement KAT, and recent advancements have adapted KAN using concepts from modern neural networks.

However, KAN struggles to effectively model physical systems that require inherent equivariance or invariance geometric symmetries as
$E(3)$
transformations, a key property for many scientific and engineering applications.
In this work, we propose the Geometric
Kolmogorov Superposition Representation (GKSR), a novel extension of KAT,
and Geometric Kolmogorov Superposition Network (GKSN), its implementation,
which incorporate invariance over various group actions,
including $O(n)$, $O(1,n)$, $S_n$ and general $GL$,
enabling accurate and efficient modeling of these systems.

Our approach provides a unified approach that bridges the gap between mathematical theory and practical architectures for physical systems, expanding the applicability of KAN to a broader class of problems. We provide experimental validation on molecular dynamical systems and particle physics.

Primary Area: learning theory

Submission Number: 24661