E(3)-Equivariant Neural Networks

Michael Scherbela, 05. Oct 2022

When learning a property from data, it is often useful or required that the model obeys symmetries of the underlying data. E(3)-equivariant networks obey the symmetry of rotation and translation in 3D space. They are therefore a natural choice when modelling data in 3D space, e.g. atomic coordinates of a molecule or point clouds obtained from 3D imaging. In this talk I plan to:

  • Explain the concept of equivariance
  • Describe the basic building blocks of an E(3) equivariant network
  • Briefly discuss an application: Learning properties of molecules