NeurIPS 2020

SE(3)-Transformers: 3D Roto-Translation Equivariant Attention Networks

Meta Review

All reviewers recommend acceptance. The main concern, shared by all reviewers, is the limited novelty of the work relative to Tensor Field Networks. Other concerns included some missing related work, and some issues with the empirical evaluation. Although the discussion did not really change anyone's view of this work in the end, the fact that it was solidly executed plays in its favour. While the combination of ideas from transformers / self-attention and tensor field networks is a logical next step following on from the latter work, I concur with R3 that the concretisation of this idea is not necessarily obvious and worth sharing with the community. I would like to join the reviewers in recommending that the authors try to emphasize the potential impact of the ability to generate spherical harmonics online in the revised version of their manuscript.