Smooth, Exact Rotational Symmetrization for Deep Learning on Point Clouds

Explore a comprehensive lecture on smooth, exact rotational symmetrization for deep learning on point clouds. Delve into the challenges of applying deep learning models to point cloud representations in chemical and materials modeling, where strict adherence to physical constraints is crucial. Learn about a novel general symmetrization method that adds rotational equivariance to existing models while preserving other requirements. Discover the Point Edge Transformer (PET) architecture, which achieves state-of-the-art performance on molecular and solid benchmark datasets. Examine topics such as equivariant coordinate system ensemble, computational cost, adaptive cutoff, impact of beta, smoothness, and related work. Gain insights into how this approach simplifies the development of improved atomic-scale machine learning schemes by relaxing design space constraints and incorporating effective ideas from other domains.

- Intro + Background
- Equivariant Coordinate System Ensemble
- Computational Cost
- Adaptive Cutoff
- Impact of beta
- Smoothness
- Related Work
- Point Edge Transformer
- Q+A