Hacker NewsJan 21

Batmobile: 10-20x Faster CUDA Kernels for Equivariant Graph Neural Networks

https://elliotarledge.com/blog/batmobile

#HackerNews #Batmobile #Equivariant #Graph #Neural #Networks #CUDA #Kernels #MachineLearning #AI

Elliot Arledge

Systems engineer and educator. Building and teaching GPU programming, CUDA, and low-level ML systems.

Show threadGregor KasieczkaJan 20, 2023

3/ Our approach is similar: the #equivariant #PointCloud layer (aka EPiC layer ๐Ÿ˜Ž) does exactly that; but with the extra twist that we also have global attributes that exchange information with the points. This turned out to be crucial!

These EPiC layers can then be combined into a standard #GAN architecture (in fact we use LSGAN, 1611.04076) and trained the usual way