Given the forecast that “by 2021, users will view 3 trillion internet videos per month, or about a million video minutes every second,” it’s clear that data compression remains critical to the field of computer science. In an effort to develop innovative ways to store and distribute information, Professor Stephan Mandt of UCI’s Donald Bren School of Information and Computer Sciences (ICS) is exploring fundamentally new approaches for compressing video and images to unprecedentedly small file sizes while preserving visual quality.
In support of this work, Mandt has been awarded $425,000 in funding from the National Science Foundation (NSF) for his grant “Deep Variational Data Compression.” The goal is to promote new ideas within data compression using deep neural networks — in particular, so-called variational autoencoders.
“Deep learning has the potential to revolutionize how images and videos are compressed to very small file sizes with minimal quality loss,” explains Mandt. “In particular, improving video compression is very timely, since most web traffic is nowadays due to streaming.”
Mandt will contribute to better compression algorithms by improving video coding, enabling faster data transfer between machine learning systems, and improving the modularity of neural codec design.
“Learned video codecs can be made even more efficient than general-purpose codecs when tuned to particular data modalities,” he explains. “The work will explore fundamentally new ways of compressing videos and images using recent advances in generative modeling and Bayesian deep learning.”
Advances in neural compression could also support data storage and distribution in fields such as astronomy, medical imaging, autonomous driving and particle physics.
— Shani Murray