万博manbetx网页版在线登录 _ 2022welcome

2019-04-15

美国密歇根大学安娜堡分校John P. Hayes教授到访微纳电子学系

2019411日,来自美国密歇根大学安娜堡分校的香农讲席教授(Claude E. Shannon Chair of Engineering ScienceJohn P. Hayes教授应邀来访微纳电子学系。

Hayes教授首先于上午10点在微纳电子大厦103报告厅做了题为《Computing with Randomness: The Stochastic Circuit Approach》的学术报告。微纳电子学系黄如院士等多位老师以及20余位研究生、本科生参加了报告会,并与报告人进行了热烈的讨论。下午,Hayes教授在205会议室进一步与相关师生展开了讨论。会上,王润声老师与相关学生介绍了微纳电子系在随机计算领域的研究进展,并与Hayes教授以及来访的上海交通大学钱炜慷老师展开了深入的讨论。

                                                      

 

Computing with Randomness: The Stochastic Circuit Approach

Almost all modern computers are deterministic and exact; randomness plays no significant role in their operation. Yet randomness has advantages, as suggested by its widespread occurrence in nature, ranging from quantum mechanical systems to the human brain. Stochastic computing (SC) is an emerging computing technique that processes data defined by pseudo-random bit-streams. It mimics aspects of the nervous system, and enables complex arithmetic operations to be performed using extremely small, low-power, and error-tolerant circuitry. SC has applications in several important areas such as image processing, complex coding techniques, and the design of artificial neural networks.  However, achieving acceptably accurate results is difficult and SC tends to require very long bit-streams and run-times.  This talk reviews the underlying concepts of SC and its applications, and discusses recent research results that focus on the accuracy issue.  Among the major sources of inaccuracy are random fluctuations in individual bit-streams, correlations between bit-streams, and inadequate randomness sources.  For example, input bit-streams denoting constant stochastic numbers play an essential role in SC, but are a significant source of random fluctuation errors. We show that it is possible to completely remove all error-inducing constant inputs from stochastic circuits by resorting to a new class of highly accurate sequential designs. We also discuss the potential benefits of SC in the design of neural networks.

 


返回