Abstract: The rate-distortion function R(D) occupies a central place in the theory of lossy compression, and characterizes the fundamental compressibility of a given data source of interest. Here, I consider the problem of estimating R(D) from data samples. Starting with the classic Blahut–Arimoto algorithm, I develop new methods for estimating R(D) that can scale to high-dimensional real-world datasets such as images and speech data. I discuss the fundamental connections to statistical estimation and optimal transport behind this work, which give us new insight into the solution of the rate-distortion problem. The talk is based on the following papers: https://arxiv.org/abs/2111.12166 and https://yiboyang.com/files/neurips….

Bio: Yibo Yang is a fifth-year PhD student in computer science at UC Irvine advised by Stephan Mandt. He is broadly interested in applications of probability, statistics, and information theory to challenging real-world problems. His PhD research has focused on the theory and practice of lossy data compression with machine learning.