blog
spilling high-value tokens
-
Introduction to Triton
GPU basics, kernel fusion, and a step-by-step implementation of Flash Attention in Triton.
-
Erwin Transformer
We can organize irregular data (point clouds, meshes) using ball trees to enable sub-quadratic (sparse) attention. Fast and expressive!
-
Clifford-Steerable CNNs
Using Clifford algebra allows us to generalize E(n)-equivariant CNNs to E(p,q), which now includes isometries of spacetime!
-
Implicit Steerable Kernels
Using implicit parameterization of G-steerable kernels to simplify designing steerable CNNs. Let me show you how.