ODSL Forum: Scalable equivariance with Geometric Algebra Transformers (Johann Brehmer)

Friday 24 May 2024, 14:00 → 15:00 Europe/Berlin

Basement Seminar Room (Origins Building)

Abstract: Should we use neural architectures with strong inductive biases for each problem, or should we focus on scalability? We believe the answer to this hotly debated question may be: both! We present the Geometric Algebra Transformer (GATr 🐊), a versatile architecture for geometric problems. We will explain the geometric algebra representations GATr uses, how the architecture is made equivariant with respect to E(3) or the Lorentz symmetry, and why its transformer architecture makes it scalable. Finally, we demonstrate GATr on problems in biology, wireless communication, and particle physics.

 

 

Zoom: https://tum-conf.zoom-x.de/j/63748396131?pwd=Mk1TaUUrT2V0Wko1eEp6aFFVYTJidz09

Meeting ID: 637 4839 6131

Passcode: 344956