Gergely Flamich

Data Compression with Stochastic Codes

This talk’s aim is to answer two basic questions: How can we use rejection sampling (and some more advanced sampling algorithms) for data compression? And why use sampling algorithms for data compression in the first place?
These two questions motivate the development of relative entropy coding, whose foundation is a curious mixture of information theory and computational statistics, and which finds applications in machine learning-based compression, rate-distortion-perception theory, and differentially private compression.
The talk will be a brief tour of the landscape of relative entropy coding. I will begin by defining relative entropy coding and outlining a basic solution. Then, I will discuss some practical applications, after which I will cover more advanced algorithms and (if time allows) recent applications to computational statistics.