Notes on Geometry & Machine Learning

Using Heat Diffusion as a Learning Operator on Manifolds

Fair warning: This isn't the same as Diffusion Models in ML.

I recently got a chance to read DiffusionNet: Discretization Agnostic Learning on Surfaces. There are two versions of this publication, one presented at Siggraph 2022, and another available at arXiv. While the former is written with Computer Graphics audience in mind, the latter one is aimed more towards Machine Learning community. While both of them are nearly identical, the network descriptions (and a few mathematical notations) are slightly different. 

The paper has a very interesting take on replacing Convolution and Pooling operators with Learned Diffusion Layers.