The Learned Data Compression group is committed to revolutionizing the fundamental architecture of visual data representation. Unlike traditional compression methods that rely on manually crafted pipelines, this group aims to achieve a paradigm shift towards end-to-end optimized systems where every component is learned from data.
A key aspect of this research is Nonlinear Transform Coding – a pioneering methodology developed by the group’s leadership. This technique replaces traditional linear transforms (such as the Discrete Cosine Transform or wavelets) with nonlinear, end-to-end trained neural networks. This innovative approach laid the foundation for the JPEG AI standard, which was finalized in 2025. Building upon this success, the group explores improvements, generalizations, and alternative approaches to established compression tools like quantization and entropy coding, challenging the constraints set forth by Shannon’s seminal work on source coding.
The second pillar of the group’s research is the understanding of human vision. Machine learning also offers the potential to build more sophisticated models of visual perception – by imitating aspects of the way humans learn to see. The group uses these models to better predict how humans experience the quality of rendered imagery. These improved quality measures in turn lead to significantly improved algorithms for visual media compression.