Inquire about more observation cases using LeanKAN.

[linjing-lab]

The Schrodinger_LeanKAN.jl and loss_LeanKAN was really impressive, and I knew the KAN (Kolmogorov–Arnold Networks) structure is a neural network architecture that represents functions through learnable univariate transformations composed in a structured, grid-like manner, offering improved interpretability and potentially better generalization compared to traditional MLPs. I'm inquiring more observation cases about LeanKAN, and in my experience, using CUDA accelerated efficient-kan instead of MLPs will maintain training speed. The development of LeanKAN is may due to lack of CUDA environment in Julia programming, or need to make progress on ODEs within cross-platform packages.

[B-Koenig]

Thanks for the discussion. As you mentioned, CUDA implementation is a potential pathway to accelerate the training speed of LeanKAN layers. While we did not study CUDA implementation or iteration speed in the LeanKAN paper in Neural Networks and are not currently looking into this, Julia does have CUDA capabilities that might be of interest for future efforts (e.g., CUDA.jl). Other languages with CUDA support are also potentially promising, especially for cases that do not involve Neural ODE/KAN-ODE integration.