In a recently published paper on Kolmogorov Arnold Networks, the code performs a least squares minimization for fitting the B-spline using the function torch.linalg.lstsq.

I am wondering wether I could introduce a separate regularization term in this least squares expression itself but there is no implementation for this in the pytorch repo.

Moreover, the function is implemented in C and not python so customizing it poses a significant task.

I tried adding the regularization term to the loss function but I would like a more local control over the spline function.