Efficient Convolutional Kolmogorov–Arnold Networks with ReLU Basis Functions for Image Classification
Year: 2026
Type: conference-paper
Abstract: Traditional Convolutional Neural Networks (CNNs) rely on fixed linear weights coupled with node-wise activation functions. In contrast, Kolmogorov-Arnold Networks (KANs) adopt learnable nonlinear functions on edges. While promising, recent convolutional adaptations, specifically spline-based KANC, use B-splines for parameterization, which can be computationally expensive and memory-intensive. In this work, we propose Convolutional ReLU-KAN (ReLU-KANC). We integrate efficient squared-ReLU basis functions into the convolutional kernel, replacing the more complex B-spline parameterization. This approach retains the theoretical advantages of KAN convolutions while significantly reducing computational overhead. Experiments on Fashion-MNIST demonstrate that ReLU-KANC achieves ~3× faster training and 0.21 ms inference time (vs. 0.72 ms for spline-based KANC) with superior accuracy (89.48%), validating it as an efficient and expressive alternative for image classification.
Authors: Feng-Cheng Lin, Hsing-Hao Yeh, Chun-Yu Hung
Institution: Feng Chia University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204087147