Abstract

In recent years there have been significant advances in using machine learning to improve compression algorithms, a field known as learned compression. Learned compression improves on traditional compression methods, by using ML methods to adapt the compression algorithm to the data at hand, often outperforming the best traditional methods. In this talk, we first provide an accessible introduction to traditional compression methods. We then give an overview of three approaches for performing learned compression: quantisation-based approaches, bits-back coding and relative entropy coding. We discuss the strengths of each of these approaches, their limitations and their practicality, and give example applications for them.

Reading list:

Ballé, Johannes, Valero Laparra, and Eero P. Simoncelli. “End-to-end optimized image compression.” arXiv preprint arXiv:1611.01704 (2016).

Townsend, James, Tom Bird, and David Barber. “Practical lossless compression with latent variables using bits back coding.” arXiv preprint arXiv:1901.04866 (2019).

Maddison, Chris J., Daniel Tarlow, and Tom Minka. “A* sampling.” arXiv preprint arXiv:1411.0030 (2014).

Our reading groups are livestreamed via Zoom and recorded for our Youtube channel. The Zoom details are distributed via our weekly mailing list.