Very interesting. A related paper from a couple of years ago proposed a similar idea to understand generalization in deep learning:

https://arxiv.org/abs/2203.10036