Technical Program

Paper Detail

Paper Title Learned Data Compression
Paper IdentifierWE2.R1.1
Authors Johannes Ballé, Google, United States
Session Deep Learning for Compression
Location Le Théatre (Parterre), Level -1
Session Time Wednesday, 10 July, 11:40 - 13:20
Presentation Time Wednesday, 10 July, 11:40 - 12:00
Abstract Within the last three years, deep learning methods have begun to revitalize the field of data compression. This has been taking place not only in traditional forums – there is also an interest in data compression from the machine learning community due to its connections to representation learning and generative modeling. In this talk, I’ll summarize our efforts to develop TensorFlow-compression, an open source toolkit for learned data compression, as well as how we use it in end-to-end optimized image compression algorithms, which have quickly caught up with (and surpassed) the compression efficiency of established standards such as HEVC. I’ll proceed to highlight a number of current research topics that are of theoretical and practical interest.