Ryan Robinett University of Chicago
The field of Riemannian Optimization generalizes optimization techniques from Euclidean state-spaces to Riemannian manifolds. While the theory of Riemannian optimization is well developed, it is scarcely implemented due to “update-and-project” methods on representations of Riemannian manifolds within a higher-dimensional Euclidean space being costly. Conjugately, while dimensionality reduction techniques allow for insightful, simplifying looks into high-dimensional data, only linear dimensionality reduction techniques succeed in preserving metric information to a degree necessary to perform Riemannian optimization on the reduced space. In this talk, we demonstrate a data structure which allows the user to implement Riemannian optimization algorithms on a pseudo-Riemannian simplicial complex which closely approximates a manifold learned from point cloud data.
Join via Zoom