Convergence of Denoising Diffusion Models Under the Manifold Hypothesis
Offered By: Alan Turing Institute via YouTube
Course Description
Overview
Explore the theoretical foundations of denoising diffusion models in this 46-minute lecture by Valentin de Bortoli from CNRS, France. Delve into the convergence analysis of these state-of-the-art generative models for image and audio synthesis, focusing on scenarios where the target distribution is supported on a lower-dimensional manifold or given by an empirical distribution. Examine quantitative bounds on the Wasserstein distance between the target data distribution and the generative distribution of diffusion models. Gain insights into the theoretical underpinnings of these models, addressing limitations in current approaches that assume target density admits a density with respect to the Lebesgue measure.
Syllabus
Convergence of denoising diffusion models under the manifold hypothesis
Taught by
Alan Turing Institute
Related Courses
Regularization for Optimal Transport and Dynamic Time Warping Distances - Marco CuturiAlan Turing Institute via YouTube Analysis of Mean-Field Games - Lecture 1
International Centre for Theoretical Sciences via YouTube Why Should Q=P in the Wasserstein Distance Between Persistence Diagrams?
Applied Algebraic Topology Network via YouTube Washington Mio - Stable Homology of Metric Measure Spaces
Applied Algebraic Topology Network via YouTube Wasserstein Distributionally Robust Optimization - Theory and Applications in Machine Learning
Institute for Pure & Applied Mathematics (IPAM) via YouTube