Lipschitz bound regularization

[DanyaLearning]

Dear friends,

since we need to integrate everything into sQUlearn from the AutoQML project, I was wondering if the Lipschitz bounds regularization stuff I worked on can be integrated and how I can help to realize this.

Here is the reference https://arxiv.org/abs/2311.11871 and you can find an example implementation here aswell https://github.com/daniel-fink-de/training-robust-and-generalizable-quantum-models. to my knowledge Frederic was able to use it in his research.

Best, Danya

[David-Kreplin]

Probably it's best to ask Christian first if it is really necessary before you start all the work.

In principle, it should not be too difficult since you can implement it as a loss, for which a class exists. It's currently a work in progress, but the Loss class is going to contain the QNN and therefore all the data you need for the Lipschitz bounds. In the Loss class, you also have access to the parameter values.

It's best to take a look at ./src/squlearn/qnn/loss.py.

[DanyaLearning]

Actually that is why I raised the feature request, because Christian asked me to.

But if there is this loss class already, I see no issue as we just calculate a norm over the weigths (those that multiply the data).