Comparative study of domain-shift robustness of Kolmogorov-Arnold convolutional networks and classical CNNs

Comparative study of domain-shift robustness of Kolmogorov-Arnold convolutional networks and classical CNNs

Larionov A.I. (UNN, Nizhnij Novgorod, Russia)
Turlapov V.E. (UNN, Nizhnij Novgorod, Russia)

Abstract

An experimental comparison of the robustness to domain shift of classical convolutional neural networks (CNN) and Kolmogorov-Arnold convolutional networks (convKAN) with trainable nonlinearities in the convolution kernels is performed. Activation functions are represented by Gaussian radial basis functions (RBF). The study is performed using the ResNet-18 architecture in a configuration without a decoder. PlantVillage was chosen as a reference dataset as it contains the norm and 26 diseases, including 17 fungal ones, for 14 plant species (38 classes in total). Diseases are represented by 4 classes, of which 2 are the most represented (fungal – 17; bacterial – 4). Domains with a shift are represented by 6 datasets of plant images, in which only a small part of plant types and diseases intersects with the PlantVillage classes. Surveying conditions are represented by laboratory and field ones. To quantitatively assess domain shifts, 3 metrics were investigated: Maximum Mean Discrepancy (MMD); Jensen-Shannon divergence; Wasserstein distance. KAN and CNN models were compared using classical metrics: Accuracy, F1 macro, F1 weighted, Recall, and FAR. To compare the influence of convolution architecture, embeddings extracted by the convolutional part of the KAN and CNN models, which are fed directly to the input of the fully connected classifying "head" of the model, were analyzed. It was experimentally established that convolutional KAN networks (convKANs) demonstrate better results in terms of domain shift resistance by an average of: 10% for Accuracy, 2% for F1-macro, and 6% for F1-weighted for the normal and 4 disease classes. It was shown that the influence of the three domain shift metrics on the results of KAN and CNN is heterogeneous, but ultimately these metrics are able to reveal which domain shifts are strong. The usefulness of using the FAR metric, which showed extreme growth for both KANs and CNNs in fungal diseases, was demonstrated, with the CNN network favoring the latter. The results indicate that strong domain shift causes a significant degradation in classification performance for both CNNs and convKAN models, but KANs are generally significantly more robust to domain shift, with slightly inferior FAR values to CNNs. It was found that after training, unimodal, multimodal, and oscillating activation functions dominate in KANs, and during training, the proportion of unimodal functions increases at the expense of oscillating ones.

Keywords

domain-shift robustness; convolutional neural networks; Kolmogorov-Arnold networks; trainable activation functions; PlantVillage dataset; plant disease classification; feature space analysis; computer vision.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 3, part 3, 2026, pp. 17-26

ISSN 2220-6426 (Online), ISSN 2079-8156 (Print).

DOI: 10.15514/ISPRAS-2026-38(3)-32

For citation

Larionov A.I., Turlapov V.E. Comparative study of domain-shift robustness of Kolmogorov-Arnold convolutional networks and classical CNNs. Proceedings of the Institute for System Programming, vol. 38, issue 3, part 3, 2026, pp. 17-26 DOI: 10.15514/ISPRAS-2026-38(3)-32.

Full text of the paper in pdf (in Russian) Back to the contents of the volume