ИСТИНА

Kolmogorov-Arnold neural networks (KAN) are an alternative approach to standard neural network models. In contrast to the search for matrix coefficients, the KAN searches for continuous activation functions [1]. This allows KAN to show greater non-linearity, and makes the result of its prediction easier to interpret. In this study, we test KAN against usual multilayer perceptron type neural networks on solving the inverse problem of exploration geophysics. Exploration geophysics requires solving specific inverse problems—reconstructing the spatial distribution of the medium properties in the thickness of the earth from the geophysical fields measured on its surface [2]. Dependence of the target variables on a large number of interconnected input features and non-linear interaction among the latter make neural networks one of efficient method of solving such problems. Here we demonstrate that for parameterization schemes with a relatively small number of input features KAN outperform multi-layer perceptron in respect to regression error, at the expense of higher computational cost. The study was carried out at the expense of the grant No. 24-11-00266 from the Russian Science Foundation, https://rscf.ru/en/project/24-11-00266/.