Title: You map, but why?
Abstract: In this talk we will look at some ideas behind UMAP (Uniform Manifold Approximation and Projection), which (to me) seems to be a very popular manifold learning algorithm despite its unusually topological/geometric underpinnings. In addition, we will see UMAP isn't just for dimension reduction, and we will sketch the ideas behind variants of UMAP in other parts of unsupervised learning. A portion of the time will be spent introducing the learning problems and common ways of tackling them. If there's any time left we might try to shoehorn in a few words about a parametric UMAP, autoencoders, and umap_pytorch. This will be an introductory talk (taking longer than 1.316 seconds) focusing on the geometric and topological aspects, and in particular we will not assume prior knowledge of manifold learning or really any unsupervised learning in general.