Kolmogorov–Arnold Neural Networks (KANs): Theoretical Foundations, Advanced Architectures, and Critical Evaluation

Kolmogorov–Arnold Neural Networks (KANs): Theoretical Foundations, Advanced Architectures, and Critical Evaluation

Abbas Neekabadi1

1) Department of computer engineering and information technology, Payame Noor University, Tehran, Iran

IOI Code : AAHD-ACZZKD
Publication : National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
    Conventional neural network architectures, such as multilayer perceptrons (MLPs) and convolutional neural networks (CNNs), continue to face challenges including substantial data requirements, high computational costs, limited interpretability, and, in some settings, unstable convergence. In response, recent research has explored alternative neural network architectures in which mathematical structure and theoretical foundations play a more central role. Among these approaches are Kolmogorov–Arnold Networks (KANs), which are inspired by the Kolmogorov–Arnold representation theorem for multivariate functions. Rather than relying on conventional scalar-valued weights between layers, KANs employ learnable univariate functions, such as B-splines and Fourier-based functions, to represent nonlinear relationships. This functional formulation provides greater flexibility in approximating complex functions and can, in appropriate settings, achieve competitive performance with fewer parameters while offering substantially improved interpretability of the learned representations. This paper presents a review of influential and representative developments in KANs, covering the original KAN architecture as well as temporal, convolutional, gradient-free, KAN-based autoencoder, and Fourier-based variants. Although these architectures have been developed for different application domains and problem settings, they share the underlying principle of representing nonlinear transformations through learnable univariate functions associated with network connections. For each architecture, its technical structure, key innovations, reported empirical performance, strengths, and limitations are critically examined based on the corresponding studies. Finally, the paper outlines promising directions for future research on KANs, including improved training strategies, alternative functional bases, enhanced robustness, and broader applications in scientific and engineering domains.

Keywords : kolmogorov arnold network kan splines interpretability