IEEE SiPS 2018 Tutorial: Machine Learning Methods for Power/Performance/Thermal Optimization of Signal Processing Systems

IEEE SiPS 2018 Tutorial: Machine Learning Methods for Power/Performance/Thermal Optimization of Signal Processing Systems

This tutorial on ML for PPT will be held on Sunday, October 21 at IEEE SiPS 2018 in Cape Town, South Africa.

Tutorial Summary

This tutorial will cover machine learning methods for use at both design time and run time. We will cover two major machine learning methods: convolutional neural networks (CNNs) and Markov models. Each has very different uses in signal processing systems, particularly for PPT characteristics.

We will discuss efficient implementations of CNNs, including numerical representation, nonlinear function representations, and memory systems.  Topics include:

  • Introduction to convolutional neural networks: architecture, operation, classification, training.
  • Implementation technologies for CNNs: CPU software, GPU and tensor processor, FPGA, ASIC.
  • Required numerical precision for CNN arithmetic.
  • Memory system requirements for CNNs.

We will discuss the use of Markov chains and Markov decision processes for signal processing systems. We will discuss design-time use of Markov decision processes (MDPs) to build run-time controllers with optimized policies. We will discuss methods for online optimization of Markov decision processes for incremental policy optimization. Topics include:

  • Overview of Markov models and Markov decision processes.
  • Algorithms to solve MDPs: value iteration, policy iteration.
  • Hierarchical Markov decision processes.
  • Using MDPs during design space exploration.
  • Online training refinement for MDPs.


Marilyn Wolf is Farmer Distinguished Chair in Embedded Computing Systems and GRA Eminent Scholar at the Georgia Institute of Technology.  She received her BS, MS, and PhD in electrical engineering from Stanford University in 1980, 1981, and 1984.  She was with AT&T Bell Laboratories from 1984 to 1989 and was on the faculty of Princeton University from 1989 to 2007.  Her research interests include cyber-physical systems, Internet-of-Things, embedded computing, embedded computer vision, and VLSI systems. She has received the ASEE Terman Award and IEEE Circuits and Systems Society Education Award. She is a Fellow of the IEEE and ACM.

Shuvra S. Bhattacharyya is a Professor in the Department of Electrical and Computer Engineering at the University of Maryland, College Park. He holds a joint appointment in the University of Maryland Institute for Advanced Computer Studies (UMIACS), and is a member of the Maryland Cybersecurity Center (MC2), and the University of Maryland Energy Research Center (UMERC). He is also a part time visiting professor in the Department of Pervasive Computing at the Tampere University of Technology, Finland, as part of the Finland Distinguished Professor Programme (FiDiPro). He is an author of six books, and over 250 papers in the areas of signal processing, embedded systems, electronic design automation, wireless communication, and wireless sensor networks. He received the B.S. degree from the University of Wisconsin at Madison, and the Ph.D. degree from the University of California at Berkeley. Dr. Bhattacharyya has previously held industrial positions as a Researcher at the Hitachi America Semiconductor Research Laboratory (San Jose, California), and Compiler Developer at Kuck & Associates (Champaign, Illinois). He has held a visiting research position at the US Air Force Research Laboratory (Rome, New York). Dr. Bhattacharyya is director of the Maryland DSPCAD Research Group. He serves as co-Editor-in-Chief for the Journal of Signal Processing Systems. He has previously served as Associate Editor for the EURASIP Journal on Embedded Systems, Journal of Signal Processing Systems, and IEEE Transactions on Signal Processing Systems, and as Guest Editor for several other journals.