Wednesday, October 08, 2003
Hewlett Packard Auditorium, 306 Soda Hall
4:00-5:00 p.m.

Professor Michael Gastpar

Electrical Engineering and Computer Sciences Dept.,
UC Berkeley

Scaling Laws for Large Sensor Networks

Abstract:

Real-world signals are often analog, and hence, sensor networks typically involve both a data compression and a data transmission problem. Traditional wisdom for the point-to-point communication scenario is to separately solve the two tasks, but it is known that this is suboptimal for networks. The degree of suboptimality, however, cannot be assessed, since the optimum performance is unknown in general.

In this talk, we consider a typical sensor network situation in which the sensors obtain noisy observations of multiple underlying sources, and the goal is for a central unit to get the best estimate of the underlying sources. For this scenario, we are able to determine the optimum performance in the scaling sense, i.e., as the number of sensors becomes large. Our results reveal that separating compression from transmission is not only suboptimal, it has an entirely different scaling law: For a fixed target distortion and number of sensors, it requires an exponentially larger total power, as compared to the optimum scheme. This illustrates the need and potential for new (joint) source-channel coding techniques for sensor networks. We outline initial results.

Michael Gastpar is currently an Assistant Professor at the University of California, Berkeley. He received the Doctorates Science degree from the Swiss Federal Institute of Technology (EPFL), Lausanne, Switzerland, in 2002, the M. S. degree from the University of Illinois at Urbana-Champaign in 1999, and the Dipl. El.-Ing. degree from the Swiss Federal Institute of Technology (ETH), Zurich, Switzerland, in 1997, all in electrical engineering. From July to September 2001, he was a summer researcher in the Mathematics of Communications group at Bell Labs, Lucent Technologies, Murray Hill, NJ. He was awarded the 2002 EPFL Best Thesis Award. His current research interests are centered around networks and involve methods and questions from signal processing and information theory.