An exploration of KANs and CKANs for more efficient deep learning architecture

Deep learning has revolutionized the field of machine learning with its ability to discern complex patterns from voluminous data. Despite the success of Multi-Layer Perceptrons (MLPs) and Convolutional Neural Networks (CNNs), there is an ongoing quest for architectures that offer higher expressiveness with fewer parameters. This paper focuses on the Kolmogorov-Arnold Networks (KANs) and Convolutional Kolmogorov-Arnold Networks (CKANs), which integrate learnable spline functions for enhanced expressiveness and efficiency. This study designs a range of networks to compare KANs with MLPs and CKANs with classical CNNs on the CIFAR-10 dataset. Moreover, this study evaluates the models based on several metrics, including accuracy, precision, recall, F1 score, and parameter count. Based on the experimental results, networks with KANs and CKANs demonstrated improved accuracy with a reduced parameter footprint, indicating the potential of KAN-based models in capturing complex patterns. In conclusion, integrating KANs into CNNs and MLPs is a promising approach for developing more efficient and interpretable models, offering a path to advance deep learning architectures.