CNN with Kolmogorov-Arnold Networks

In our investigation of CNN architectures, we integrated Kolmogorov-Arnold Networks (KANs) to compare against traditional fully connected (FC) layers. We found that KANs, with their non-linear spline-based transformations, can capture complex patterns more efficiently than FC layers, potentially reducing the need for deeper or more complex network structures. Despite KANs typically having a larger parameter count due to their intricate spline functions, they offer a significant advantage in modeling capabilities. This study highlighted the potential of KANs to outperform standard FC layers in tasks requiring high levels of data interpretation and complexity. Our findings suggest a promising avenue for future research in neural network design, focusing on optimizing KAN configurations to balance parameter efficiency with computational performance.