Explore score estimation techniques using infinite-dimensional exponential families in this 54-minute lecture by Dougal Sutherland from UCL, presented at the Alan Turing Institute. Delve into the mathematical foundations of approximating high-dimensional functions from limited data, addressing the curse of dimensionality and modern approaches to overcome it. Learn about reproducing kernel Hilbert spaces, maximum likelihood, score matching, and Nyström approximation. Examine theoretical aspects, competitor approaches, and experimental results. Gain insights into minimizing scores, loss functions, and practical applications in science and engineering for reconstructing complex processes with numerous parameters.
Score Estimation with Infinite-Dimensional Exponential Families – Dougal Sutherland, UCL
Alan Turing Institute via YouTube
Overview
Syllabus
Introduction
Reproducing kernel Hilbert space
Maximum likelihood
Score matching
Nystrom approximation
Theory overview
Proof
Competitor approach
Experiments
Recap
Minimize the score
Loss function
Experimental results
Summary
Taught by
Alan Turing Institute
