Expressive power of invariant and equivariant GNNs
Precise characterization of what symmetry-aware graph networks can represent.
With Marc Lelarge I analyzed the approximation capabilities of invariant and equivariant graph neural networks, resulting in the ICLR 2021 paper Expressive power of invariant and equivariant graph neural networks.
The work clarifies how architectural constraints interact with group symmetries, and when additional structure is required to model combinatorial problems.
We showcased the findings at the MIPT-UGA workshop and the Thoth seminar (slides).