Deep Neural Network Accelerator Designs with Approximate, Stochastic, and Neuromorphic Computing

Publisher:闻天明Release Time:2021-04-26Number of visits:10

Speaker:     Prof. Weikang Qian 
Time:          Apr.28.2021 10:00-12:00 
Location:    SIST 1A 200 
Host:            Prof. Pingqiang Zhou 
  
Abstract: 
Deep neural networks (DNNs) play an increasingly important role in many fields such as computer vision and natural language processing. However, they require a huge amount of arithmetic operations and data movement, causing large energy consumption. Thus, designing energy-efficient DNN accelerator is imperative. In this talk, we will present our recent research in designing energy-efficient DNN accelerator by exploiting emerging computing paradigms including approximate computing, stochastic computing, and neuromorphic computing. The first part of the talk will introduce a logic synthesis method for approximate computing and its integration into the design loop of DNN accelerators. The second part of the talk will introduce a DNN accelerator design related to stochastic computing. We propose a novel parallel bit stream-based DNN accelerator, which uses sorting network to do accumulation. The third part of the talk will focus on resistive random access memory (RRAM) crossbar-based neuromorphic DNN accelerator design. The imperfect fabrication process combined with stochastic filament-based switching leads to resistance variations, which can significantly degrade the accuracy of DNNs. We propose two novel solutions to address this issue. The first one applies unary coding for weight representation to reduce the device-to-device variation, while the second one introduce digital offsets into the crossbar to reduce the cycle-to-cycle variation.

  
Bio: 
Weikang Qian is an associate professor in the University of Michigan-Shanghai Jiao Tong University Joint Institute at Shanghai Jiao Tong University. He received his Ph.D. degree in Electrical Engineering at the University of Minnesota in 2011 and his B.Eng. degree in Automation at Tsinghua University in 2006. His main research interests include electronic design automation and digital design for emerging computing paradigms. His research works were nominated for the Best Paper Awards at 2009 and 2020 International Conference on Computer-Aided Design (ICCAD), at 2020 Design, Automation, and Test in Europe Conference (DATE), and at 2016 International Workshop on Logic and Synthesis (IWLS).

SIST Seminar 202107