March 06, 2012, HFH 4164

In modern systems, information flows not only through the systems that communicate it, but also through the circuits that process it, and through the control agents that use it. In this talk, I will discuss how investigating these less obvious flows of information can enhance our understanding of these systems. I will first consider a computational twist on the classic communication problem. Traditional information theory focuses almost exclusively on transmit power. In modern systems, however, the power required to process the signal at the transmitter and the receiver can be significantly larger than the transmit power (e.g. in communication in data-centers, short-distance wireless, etc.). Just as channel models help us to understand the required transmit power, I will present new models of the circuit implementation of encoding and decoding to understand the required processing power. I will then derive fundamental limits for these models showing that the traditional approach of operating close to the Shannon limit fundamentally requires large encoding and decoding power. The derivation of these limits involves analyzing the information bottlenecks in circuits using tools from information theory, decentralized control, and large-deviations. I will also present our experimental work in jointly designing code and decoding circuits for minimizing total power, which complements our theoretical bounds. In the second part of the talk, I will consider the problem of control of cyber-physical systems (e.g. the power grid, robotic teams, vehicular networks, etc.). In these systems, there is often a possibility for the agents to interact through the plant. This possibility of “implicit” communication makes decentralized control problems nonconvex and extremely hard to optimize. I will provide the first approximately-optimal solutions to the minimalist problem of implicit communication: the Witsenhausen counterexample, that has been open for more than 40 years. These solutions are derived by understanding the information bottlenecks in implicit communication using information theory, deterministic models, large deviations, and control theory. I will then explain how the understanding of implicit communication can extend to larger and more complicated systems.

Pulkit Grover (Ph.D. UC Berkeley) is a postdoc at Stanford University. He is interested in interdisciplinary research directed towards developing a science of information applicable to decentralized systems and low-power communication and circuits. Dr. Grover is the recipient of the best student paper award at the IEEE Conference in Decision and Control (CDC) 2010. For his dissertation research, he received the 2011 Eli Jury Award from the Department of EECS at UC Berkeley.