Equivariant Multilayer Perceptrons

[PhilipVinc]

@femtobit , @chrisrothUT Maybe you saw already but there is this interesting article by @ mfinzi about implementing equivariant perceptrons in jax.
Maybe there could be ideas we can steal.

Their approach is also quite general on the groups they can represent...
They also have a type system for irreducible representations, apparently
A Practical Method for Constructing Equivariant Multilayer Perceptrons for Arbitrary Matrix Groups

mfinzi/equivariant-MLP

[chrisrothUT]

Thanks! I actually had not seen this. It’s not the lightest read but I definitely want to understand it. On first glance it looks like a different model than the G-CNN where they somehow make the representation space more compact.