Description: Error bounded lossy compression (EBLC) presents an opportunity to accelerate I/O for traditional and emerging AI-based scientific workflows, but for this to occur users need tools to understand it's impact on scientific results. I will present two works related to the topic of accelerating AI for science through lossy compression. The first work OptZ-config provides a technique to bound the quality of a user's analysis using black-box optimization techniques. Using this technique we are able to bound new classes of metrics previously not preserved by EBLC while improving performance against specialized compressors, analytical methods, and prior black-box methods on previously supported classes. The second work studies the use of EBLC to reduce the volume of training data for ML/AI. We find that EBLC is both safe and more effective on AI training data than previous methods and identify a path forward for compressor design for this area.
Join on your computer or mobile app
Or call in (audio only)
+1 630-556-7958,,403701240# United States, Big Rock
Phone Conference ID: 403 701 240#