The recent buzz surrounding Kolmogorov-Arnold networks (KANs) has sparked interest in their potential to outperform traditional neural networks. While KANs show promise in achieving similar performance with fewer parameters, they require extensive tuning and complex implementation. B-Splines, the preferred activation function in KANs, offer robust customizations and differentiability crucial for machine learning. However, PyKAN, a KAN implementation, introduces additional techniques like bias vectors, spline weights, and base functions to enhance performance but at the cost of increased complexity. Despite experimenting with various adaptations, the simplicity and efficiency of traditional neural networks outshined the elaborate KAN models in the end.
https://cprimozic.net/blog/trying-out-kans/