Theory

On the Expressive Power of Geometric Graph Neural Networks
A unified theory of Graph Neural Networks and Transformers for 3D molecular structure. And mechanistic understanding of how architectural decisions, such as invariance vs. equivariance, impact the class of functions that models can express. (ICML 2023 & Oral at NeurIPS 2022 Workshop)
Chaitanya K. Joshi, Cristian Bodnar, Simon V. Mathis, Taco Cohen, Pietro Liò
PDF Cite Code Slides Video NeurIPS Workshop Oral
On the Expressive Power of Geometric Graph Neural Networks