Kolmogorov-Arnold Network analysis and comparison with classical neural networks

Abstract

This thesis analyzes Kolmogorov-Arnold Networks and compares them with classical Multi-Layer Perceptrons. The theoretical part describes the principles of both architectures. The experimental part evaluates KANs and MLPs on binary classification, multiclass classification, and regression tasks using the Two Moons, MNIST, and PJME Hourly Energy Consumption datasets. The comparison considers predictive performance, training time, inference time, memory usage, and the number of parameters. The results show that KANs can outperform MLPs on selected low-dimensional and regression tasks, but they are usually computationally more expensive. Therefore, KANs represent a promising alternative to MLPs, although their practical usefulness depends on the task and computational constraints.