A New Perspective on Building Efficient and Expressive 3D Equivariant Graph Neural Networks

Explore a new perspective on building efficient and expressive 3D equivariant Graph Neural Networks in this 55-minute conference talk by Valence Labs. Dive into the world of geometric deep learning and its application in encoding physical symmetries for 3D object modeling. Learn about the proposed local hierarchy of 3D isomorphism for evaluating equivariant GNNs' expressive power and the process of representing global geometric information from local patches. Discover two crucial modules for designing expressive and efficient geometric GNNs: local substructure encoding (LSE) and frame transition encoding (FTE). Examine the implementation of LEFTNet, which effectively utilizes these modules to achieve state-of-the-art performance in molecular property prediction tasks. Gain insights into the design space for future developments of equivariant graph neural networks through comprehensive discussions on implementing 3D equivariant GNNs, scalarization of frame bundles, equivariant frames, and neural sheaf interpretation. Conclude with an exploration of experiments, applications, and a Q&A session to deepen your understanding of this cutting-edge approach in geometric deep learning.

- Intro and Overview
- Implementing 3D Equivariant GNNs
- Scalarization of Frame Bundles and Equivariant Frames
- Equivariant Frame is a Simple Way
- Towards Expressive 3D Equivariant GNNs
- A Neural Sheaf Interpretation
- LEFTNet Implementation
- Experiments and Applications
- Q+A