PhD Dissertation Proposal by Huynh Nguyen Loc | Exploiting Approximation, Caching and Specialization to Accelerate Vision Sensing Applications

About The Talk

The advances in deep learning research have revolutionized many important fields such as speech recognition, natural language processing, especially computer vision. Cameras, which are currently deployed on personal smartphones or in public and private spaces, have become ubiquitous and contributed an important role in the success of deep learning. By collecting a huge amount of imagery data from cameras, people can use it to train highly accurate deep learning models. Such models enable applications like Amazon Go [3] that allows users to experience “grab and go” services at retail stores without any lines and checkouts, or local assistant applications that give guidance advices for individuals who suffer from dementia. However, despite its success, deep learning still relies on a massive amount of computational power and poses many new challenges in terms of efficiency, scalability and even a threat to privacy. This thesis shows that it is possible to significantly reduce the memory usage and latency of deep learning pipelines, with minimal loss of accuracy, by utilizing novel system optimizations such as 1) a smart caching algorithm that reuses feature data between multiple video frames, 2) a pruning technique that removes redundant filters of existing models and, 3) a catalog of models that provides shared computations and flexible voting mechanism.

Speaker Biography

Huynh Nguyen Loc is a Ph.D. candidate in the School of Information Systems, Singapore Management University, working in the area of mobile and wearable computing. He received his Bachelor of Engineering and Master of Engineering degree from the Ho Chi Minh University of Technology, Vietnam, in 2011 and 2014 correspondingly. His current research focuses on optimizations for deep learning systems, especially vision sensing systems.