KAN Convolution for Image Classification

This paper introduces a convolutional method called KAN convolution, designed to replace traditional convolutional layers in the ResNet architecture. The KAN convolution is inspired by Kolmogorov-Arnold Networks (KANs), which leverage B-spline functions to model univariate functions, allowing for more effective feature extraction and better representation of complex data patterns. Our approach aims to address the limitations of conventional convolutional layers in capturing hierarchical and contextual relationships within the data. By integrating KAN convolution into ResNet, the modified model demonstrates improved performance in image classification tasks across various benchmark datasets. The proposed method is thoroughly evaluated on three widely used datasets: CIFAR10, CIFAR-100-20, and STL-10. The evaluation metrics include accuracy, precision, recall, and F1 score, providing a comprehensive assessment of classification performance. Experimental results show that ResNet models with KAN convolution consistently outperform their traditional counterparts on complex datasets, particularly CIFAR-100-20 and STL-10. Moreover, the introduction of KAN convolution leads to more stable training and enhanced generalization capabilities.