[WIP] Equivariant NN feedback

[kjappelbaum]

• [ ] "geometric deep learning, which is learning with non-Euclidean (x,y,z) spatial data". I think the 3G people around Bronstein would see it a bit broader and say that geometric deep learning is a way that unifies many of the ML paradigms (convolutions on grids, graphs, ...)

and i think, as you said, it needs some rewriting to reflect the new results, e.g., with Nequip.

Also, it might be worthwhile to try to look a bit into the similarities of what those Euclidean neural networks and Ceriotti et al. are learning. In the end, they both do some Clebsch-Gordan manipulations and it might be worth looking deeper into the similarities / dissimilarities

[whitead]

Addressed for now in 5bff3a9 - will re open if there is additional feedback