國立清華大學資訊工程學系

Department of Computer Science

National Tsing Hua University

專題演講

SEMINAR

主講人： HT Kung   孔祥重 教授

SPEAKER  Harvard University

題　目：Accelerating Learning and Inference

TOPIC     

時  間：110年10月8日(五)上午10點至12點

DATE　　

地　點：線上會議室

PLACE

(meeting link: https://teams.microsoft.com/l/meetup-join/19%3afkE_SDD1vdIkP4vY8kqmtlL5o3uDwVlMfQxiWsjw_UI1%40thread.tacv2/1632317678550?context=%7b%22Tid%22%3a%226c3bc511-43c7-4596-baeb-2335c69c41f1%22%2c%22Oid%22%3a%22821516e4-93a6-42fb-bdf6-31b22f1d8c2c%22%7d)

Abstract:

We are in an exciting era where we can build special-purpose AI accelerators for general-purpose use. For instance, deep learning accelerators apply to a variety of applications. Technology companies now have a fantastic lineup of all sorts of attractive AI chips. Many of us have waited decades for this opportunity.

In this presentation, we will first overview approaches that enable efficient deep learning training and inference. Then we will present three example techniques under study in our lab at Harvard:

1. Apply stochastic rounding in ultra-low-precision block floating points to maintain the gradient descent direction in training.

2. Shape computation blocks executed from a local memory for increased computation throughput without increasing external memory bandwidth.

3. Use signed-digit representations instead of conventional binary representations for improved arithmetic efficiency that exploits bit-level sparsity.

Finally, we will mention future research directions, such as in-memory processing for deep learning computation. The presentation is based on our recent publications in conferences such as ASPLOS, ICS, and SC.

About the speaker:

HT Kung is William H. Gates Professor of Computer Science and Electrical Engineering at Harvard. As a volunteer, he also serves as the President of Taiwan AI Academy. Professor Kung has pursued various research interests in his career, including those related to this presentation, such as VLSI design, systolic arrays, parallel computing, computer architecture, embedded deep learning, and distributed computing. Professor Kung's academic honors include Member of the National Academy of Engineering (US), Member of Academia Sinica (Taiwan), Guggenheim Fellowship, and the ACM SIGOPS 2015 Hall of Fame Award.

Host: 郭昱廷