[ENH] Kolmogorov Arnold Block for NBeats network

[benHeid]

Is your feature request related to a problem? Please describe.
The following paper implements a KAN block in NBEats.

https://arxiv.org/pdf/2412.17853

Describe the solution you'd like
Implement exchangeable blocks for nbeats.

Check if the same can also be used for nhits.

[Sohaib-Ahmed21]

@benHeid @fkiraly I want to work on this issue. Kindly assign me.

[Sohaib-Ahmed21]

https://arxiv.org/pdf/2412.17853

@benHeid The paper utilizes Domain-Adversarial Training of Neural Networks (DANN) to learn domain-invariant features. Should DANN be incorporated into the implementation, or is it sufficient to simply switch between MLP and KAN blocks? Since DANN seems to operate at a higher level and can be integrated into any model, it might be better suited for the base design.

Currently, my approach involves using a flag (param) in NBeats: when the flag is set to True, KAN blocks are used; otherwise, MLP blocks are used. This enables an interchangeable implementation.

Would appreciate your thoughts or suggestions on this approach.

[benHeid]

I would focus only on implementing KAN blocks.

The paper seems to have two contributions, combining NBEATS with KAN blocks and then also the DANN for zero-shot forecasting, if I understand it correctly.

[Sohaib-Ahmed21]

I would focus only on implementing KAN blocks.

The paper seems to have two contributions, combining NBEATS with KAN blocks and then also the DANN for zero-shot forecasting, if I understand it correctly.

Thanks for your input.

Btw what are your thoughts about having DANN in sktime in future for zero-shot forecasting tasks where we have primary and secondary/multiple domains in training?

[benHeid]

To be honest, I am not really familiar with DANN and currently, I am not that convinced by DANN especially if you consider that the foundation models achieve something similar without doing DANN. Furthermore, I fear that this might be more complicated to implement it in a proper way. However, if you would like to implement it and you have an idea on how this would look like from an architectural perspective you are welcome to do it. But before starting to implement it, I suppose that we need to discuss API design for that.

What are your thought with regard to DANN?