March 2, 2018, Abbas Rahimi, Postdoctoral Research Fellow, ETH Zürich

Friday, March 2, 2018, 2pm-3pm
8 St. Mary’s Street, PHO 210
Refreshments at 1:45pm

Abbas Rahimi
Postdoctoral Research Fellow, ETH Zürich

Breaking Down the Error Barriers: From Error-Tolerance to Approximate and Brain-Inspired Computing

Scaling model of semiconductors has been immensely successful in providing exponentially increasing computational performance at an ever-reducing cost and energy footprint. Underlying this evolution is a set of well-defined abstraction layers, starting from robust switching devices to scalable and stored program architecture, which is Turing complete. Unfortunately, this abstraction chain is being challenged as scaling continues to nanometer dimensions. Maintaining the current deterministic computational model ultimately puts a lower bound on the energy scaling set in place by uncertainty (arising from process variations, temporal changes and data statistics). On the other hand, the nature of computation itself is changing with data rather than algorithm taking primacy. Both these trends force us to rethink functionality to cope with uncertainty by adopting computational approaches that are inherently robust to uncertainty and “approximate” in nature.

We entail the formulation, analysis, and development of a unified hardware/software environment that addresses the challenge of uncertainty in deeply scaled CMOS processes. We devise methods to predict and prevent, detect and correct, and opportunistically accept impact of uncertainty and the resulting errors at many layers in the system abstraction. This discussion naturally leads to use of these methods into emerging area of approximate computing. The methods are further combined across hardware/software stack to significantly improve cost and scale of error tolerance in massively parallel integrated architectures, accelerators, and FPGAs. Going one step further, we take inspiration from the very size of the brain’s circuits, to compute with points of a high-dimensional (HD) space that thrives on randomness and mediocre components. HD computing provides a novel look at data representations (holographic and pseudorandom HD vectors), associated operations, circuits, and architectures; it overcomes low SNR and large variability in both data and platform to perform robust learning and classification improving energy efficiency. This offers an alternative approach for the next-generation nanofabrics and applications.

Abbas Rahimi received his B.S. in computer engineering from the University of Tehran, Tehran, Iran (2010) and his M.S. and Ph.D. in computer science and engineering from the University of California San Diego, CA, USA (2015), followed by two years postdoctoral research in the Department of Electrical Engineering and Computer Sciences at the University of California Berkeley, Berkeley, CA, USA. Dr. Rahimi has been awarded an ETH Zurich Postdoctoral Fellowship, and subsequently joined the Department of Information Technology and Electrical Engineering in June 2017. He is also affiliated with the Berkeley Wireless Research Center. His research interests include embedded systems and software, brain-inspired computing, approximate computing, and massively parallel integrated architectures with an emphasis on improving energy efficiency and robustness. His doctoral dissertation has received the 2015 Outstanding Dissertation Award in the area of “New Directions in Embedded System Design and Embedded Software” from the European Design and Automation Association (EDAA). He has also received the Best Paper at BICT, 2017 and the Best Paper Candidate at DAC, 2013.

Faculty Host: Ayse Coskun
Student Host: Sean Sanchez