Advanced Network Colloquium: Pulkit Grover, "Controlling and suppressing noise"
Friday, May 6, 2016
1146 AV Williams Building
The Advanced Networks Colloquium
Strategies for controlling and suppressing noise in error-prone computing
Electrical and Computer Engineering
Center for Neural Basis of Cognition
Carnegie Mellon University
What is the minimum energy required to compute reliably using error-prone gates, given an energy-reliability tradeoff for any single gate? With saturation of Moore's law (and Dennard's scaling), this question is becoming increasingly important as researchers seek alternatives to ultra-reliabe CMOS devices. I’ll describe our novel ENCODED (ENcoded COmputation with DEcoders EmbeddeD) strategy that computes linear transforms using error-prone components. Despite *all* components being error-prone, this strategy succeeds in keeping information intact (error-probability bounded below a constant) while consuming less energy than classical repetition-type (and "uncoded") approaches in order-sense. Our work complements recent work on strong data-processing inequality, which points out the fundamental difficulty in noisy computing: dissipation of information along any computation path. ENCODED addresses information-dissipation on a single path by "coding the computation itself," and then accumulating information from multiple coded paths, all using error-prone components. I will conclude this part by discussing why the problem of information-energy dissipation tradeoff is so interesting and exciting. While Shannon's communication theory can be viewed as the first word on this tradeoff, it barely scratches the surface. I’ll next talk about a parallel tradeoff in a distributed bio potential sensing problem, which is a novel twist on the classical problem of distributed compression of a Markov source. Here, differences of potentials are sensed, and sophisticated (binning-type) strategies are hard to implement. I'll discuss how a novel “hierarchical” architecture that limits error-accumulation turns out to have a substantially improved information-energy dissipation tradeoff than simply "compressing innovations" (a strategy known to be suboptimal from a work of Kim and Berger). This is a part of a larger work on utilizing information theory to motivate and engineer ultra-high-density neural sensing interfaces, as well as provide fundamental limits on their precision and performance.
Pulkit Grover (Ph.D. UC Berkeley'10, B.Tech.'03, M.Tech.'05 IIT Kanpur) is an assistant professor at CMU. Prior to joining CMU in 2013, he was a postdoctoral researcher at Stanford. He is interested in interdisciplinary research directed towards developing a science of information for making decentralized sensing, communication and computing systems (including biomedical systems) energy-efficient and stable.
He is the recipient of the 2010 best student paper award at the IEEE Conference in Decision and Control (CDC); a 2010 best student paper finalist at the IEEE International Symposium on Information Theory (ISIT); the 2011 Eli Jury Award from UC Berkeley; the 2012 Leonard G. Abraham best paper award from the IEEE Communications Society; a 2014 best paper award at the International Symposium on Integrated Circuits (ISIC); and a 2014 NSF CAREER award.
He has served as an editor for two issues of IEEE Journal of Selected Areas in Communications (JSAC) on energy harvesting and green communications (in 2014 and 2015), has been on the TPC of IEEE WiOpt 2015, ICDCS 2015 Energy Management and Green Computing, and IEEE ICC 2013 workshop on green communications, and will be on the TPC of ISIT 2016. He has also co-organized video-abstracts at ITA 2013 and 2014.