High-Speed Complex Systems: Harnessing Core Structures and Randomness

Friday, April 24, 2009 - 3:33pm

9:00 – 10:00am
Electrical and Computer Engineering Conference Room, Harold Frank Hall Rm. 4164

HOST: Computer Engineering Program

Massachusetts Institute of Technology

Title: High-Speed Complex Systems: Harnessing Core Structures and Randomness

We are ushering into the era of ubiquitous large-scale complex systems that will
offer high-speed data communication, reliable low-power data storage and
efficient data inference — all occurring on a massive scale. An indispensable
step towards a systematic innovation in such high-performance systems is the
identification of the appropriate performance metric, and the subsequent
development of a methodology for efficient system evaluation based on this
metric. I will put forth an approach based on sophisticated statistical methods
that harness core system structures and the underlying randomness to produce
fast and accurate system evaluation. I will discuss this approach in the
context of graph-based codes that are becoming the error correcting codes of
choice for most high-speed communication systems. The performance of such
systems is determined by the probability of error in decoding. The key
challenge in efficient evaluation and better system design arises from the
dependency on the underlying iterative decoding algorithms that are practical
but not well understood. I will introduce the concept of an absorbing set as
the fundamental combinatorial structure for identifying dominant decoding
failures: this structure redefines the conventional performance metric. A new
theoretical framework based on the absorbing sets leads to a highly efficient
and accurate importance sampling evaluation of graph-based codes, and moreover
enables a systematic improvement of practical communication systems. By
generalizing the traditional domain of communication systems to the realm of
delay-sensitive complex systems, I will discuss how the proposed approach based
on the information theoretic ideas of rare events coupled with the suitable fast
statistical algorithms, can be successfully applied for evaluating the yield
(proportion of functional devices) of nano-scale circuit systems, as well as
for efficient inference in increasingly popular large-scale social networks.

Lara Dolecek is a post-doctoral researcher with the Massachusetts Institute of
Technology. She holds a B.S, M.S. and Ph.D. degrees in Electrical Engineering
and Computer Sciences, as well as an M.A. degree in Statistics, all from the
University of California, Berkeley. For her dissertation she received the 2007
David J. Sakrison Memorial Prize for the most outstanding doctoral research in
the Department of Electrical Engineering and Computer Sciences at UC Berkeley.

She also received several UC Berkeley-wide awards for her graduate research
including the multi-year Eugene Cota-Robles Fellowship and the Dissertation
Year Fellowship. Her research interests span information and probability
theory, graphical models, combinatorics, statistical algorithms and
computational methods with applications to high-performance complex systems for
data processing, communication, and storage.