SCIEN 2017: Professor Kayvon Fatahalian

Towards an Exapixel Per Second: Enabling Efficient Visual Data Analysis at Scale

Abstract: In the next decade the proliferation of cheap image sensors, connectivity, and new image analysis algorithms poses a great visual computing systems challenge: efficiently analyzing the exapixel-per-second data stream that will be produced worldwide by these cameras.  This rapidly growing workload will be central to new forms of entertainment (VR video, live video streaming), autonomous robotics, retail, urban management, medicine, and basic survey science. To achieve desired scales, these workloads require efficient algorithms, specialized, heterogeneous processing, and datacenter-scale storage and computation — effectively blending systems technologies from real-time graphics, supercomputing, and cloud-scale computing.

This talk will describe systems challenges of conducting visual data analysis at scale, and discuss initial experiences with emerging applications (ranging from VR video processing to understanding bias thousands of hours of TV news video).  I will also present work on new programming systems for large-scale video analysis that allow pixel processing pipelines to be executed and efficiently at cloud scale on large video datasets, utilizing advanced processing hardware such as GPUs, ASICs, and FPGAs.  These systems have allowed researchers and data analysts to perform video analysis tasks that previously took days or weeks, in only a few hours on large clusters, significantly expanding the set of questions that can be asked about large video datasets.

Biography: Kayvon Fatahalian is an Assistant Professor in the Computer Science Department at Stanford University. He earned a Ph.D. from Stanford in 2011, and previously was an Assistant Professor at CMU. He is the co-PI of the Intel Science and Technology Center for Visual Cloud Systems located at CMU and Stanford.
His research focuses on the design of high-performance systems for real-time graphics and for enabling analysis and data mining of image and videos collections at scale.

Link to video recording of this talk is here