On the Expressive Power of Geometric Graph Neural Networks

Explore the expressive power of Geometric Graph Neural Networks in this comprehensive conference talk. Delve into the proposed Geometric Weisfeiler-Leman (GWL) test for discriminating geometric graphs while respecting underlying physical symmetries. Examine how key design choices influence geometric GNN expressivity, including the limitations of invariant layers and the advantages of equivariant layers. Learn about synthetic experiments that supplement the theoretical framework and highlight the need for higher-order tensors and scalarisation in geometric GNNs. Gain insights from speakers Chaitanya K. Joshi and Simon V. Mathis as they discuss types of GNNs, provide key takeaways, and explore the background of GNNs for geometric graphs. Conclude with a summary and participate in a Q&A session to deepen your understanding of this cutting-edge topic in graph neural networks and geometry.

- Intro
- Types of GNN
- Key Takeaways
- Background: GNNs for Geometric Graphs
- Geometric Weisfeiler-Leman Test
- Synthetic Experiments on Geometric GNN Espressivity
- Conclusion and Summary
- Q+A