See the notes here

I’ve recently been preparing to give a lecture to the first-year Intelligent Earth CDT students at Oxford on the topic of geometric deep learning. However, with my changing research interests, I’ve decided to take the lecture in a somewhat different direction—looking at how deep learning itself can be studied geometrically to complement standard topics in GDL.

I have shared the slides here. Inside, you can find a presentation on foundational topics in Riemannian geometry (with a particular focus on working with local coordinates—something I feel is often overlooked in standard presentations of the subject) and Riemannian diffusion models, as well as a discussion and a few examples on applying geometric tools to neural networks, in particular, architecture design and optimisation.