[2016-12-30 ]Professor H. T. Kung , Harvard University," How Can End Devices Play a Deeper Role in Deep Learning, Beyond Just Being Dumb Devices?”
Title: How Can End Devices Play a Deeper Role in Deep Learning, Beyond Just Being Dumb Devices?
Date: 2016-12-30 3:40pm-5:00pm
Location: R102, CSIE
Speaker: Professor H. T. Kung , Harvard University
Hosted by: 許永真 教授
At present, due to their limited resources, end devices such as mobiles, wearables and IoT devices send data to the cloud for deep learning processing. Such cloud offloading is unsatisfactory because of concerns related to communication cost, battery consumption, response time and privacy, as well as giving away of valuable labeled data. This practice is also unfortunate for end devices because they are cut off from participating in the exciting era of deep learning, beyond just being dumb devices. In this talk, we will discuss distributed deep neural networks (DDNNs) over the cloud, the edge and end devices, where end devices play a critical role in providing early decisions, improving quality of inference, reducing communication cost, and supporting fault tolerance. We demonstrate DDNN with a multi-view, multi-camera dataset. As a DDNN end device, we illustrate a neural network implementation for the MNIST and CIFAR10 datasets on a microcontroller with only 15KB of usable memory. DDNN is joint work with Harvard graduate students, Bradley McDanel and Surat Teerapittayanon.
Professor H. T. Kung is interested in computing, communications and sensing, with a current focus on embedded deep neural networks, distributed computing for machine learning, multimedia classification, and millimeter wave wireless communications for next-generation mobile networks. Prior to joining Harvard in 1992, he taught at Carnegie Mellon University for 19 years.
Professor Kung has pursued a variety of research interests in his career, including complexity theory, database systems, VLSI design, parallel computing, computer networks, network security, wireless communications, and networking of unmanned aerial systems. He is widely known for his work in systolic arrays, optimistic concurrency control, and theory of I/O complexity. From 1999 to 2006, he co-chaired a joint Ph.D. program with colleagues at the Harvard Business School on information, technology, and management.
To complement his academic activities, Professor Kung maintains a strong link with industry. He has served as a consultant and board member to numerous companies and government agencies. Professor Kung's professional honors include: Member of the National Academy of Engineering in US; Member of the Academia Sinica in Taiwan; Guggenheim Fellowship; Shell Distinguished Chair; recipient of the Inventor of the Year Award by the Pittsburgh Intellectual Property Law Association; and the ACM SIGOPS 2015 Hall of Fame Award (with John Robinson) that recognizes the most influential Operating Systems papers that were published at least ten years in the past. He received his BS from National Tsing Hua University.