Ziming Liu

Hi, I found this linear projection trick from GraphKAN useful: https://github.com/WillHua127/GraphKAN-Graph-Kolmogorov-Arnold-Networks. In short, it's better to implement KAN in latent space so it would be nice to first linear map...

please try remove set default torch dtype in [kan init file](https://github.com/KindXiaoming/pykan/blob/master/kan/__init__.py). The latest branch has done that, or you can comment that line locally.

yeah I think KANs, as they are right now, cannot handle convolution. It seems reasonable to define`ConvKAN` layers. Given the current implementation, the only thing you can do with vision...

As a quick cute example, you may try play with KAN as if playing with an MLP for MNIST. Please make sure input data have shape `[data size, indim]`, indim=784....

Nice! yes, there's still some issue with GPU training. Looking forward to your new development :-)

My experience with MNIST is that a 2-Layer KAN with an extremely small (say 5 or 10) hidden neurons is enough to train MNIST (but maybe my impression was from...

Hi, I agree with @guyko81, that's basically what I want to say. From my experience (which is also quite limited tbh), it would be great to start from a small...

Hi, it's likely that my random seed is not set properly or the notebook tutorials are a bit lagged behind the code. Does random seed work for you (with the...

Sorry, $\ell=CN^{-\alpha}$ we meant the scaling exponent $\alpha$ is independent of dimemsion, but C is depedent.

大家好，感谢大家的讨论。KAN（对于所有模型都是这样）最终是否有用取决于你的应用是什么。可解释性是一个挺主观的东西，不同领域的人对什么东西可解释不一样。比如科学领域，函数就是语言，KAN是有帮助的。对自然语言，人类的概念确实很难用函数去表达，KAN就不行。另外符号回归到底是离散（Search）还是连续（训练神经网络）做，现在还没有定论，只是我个人的偏好，很难说谁好谁坏（还是看应用场景）。