Coding and Message-Passing for Large-Scale Distributed Storage and Inference

Dr. Alex Dimakis
USC

Monday, March 16, 2009 at 2:00pm
Engr IV Room 57-124

Abstract
Multiple recent advances in technology have catalyzed a paradigm shift away from centralized schemes and in the direction of distributed and cooperative architectures for large-scale systems. In applications like data centers, sensor networks, and peer-to-peer networks, coding is used to introduce redundancy for robustness. I will show that network coding can surprisingly reduce the communication requirements compared to standard Reed-Solomon codes used in current architectures. Further, I will present novel information theoretic performance bounds and explicit network codes that achieve optimal performance.

For the case of large-scale distributed inference, I will present some novel message-passing algorithms and show explicit results on convergence rate. In particular, I will present a gossip algorithm that scales linearly in the number of nodes for a large class of geometric graphs, resolving an open problem in distributed message passing algorithms.

Biography
Alex Dimakis is a CMI postdoctoral scholar at Caltech and will be joining USC as an Assistant Professor this summer. He received his Ph.D. in 2008 from UC Berkeley working with Martin Wainwright and Kannan Ramchandran and his Diploma degree in Electrical and Computer Engineering from the National Technical University of Athens in 2003. His research interests include Communications, Signal Processing, and Networking with applications in large-scale distributed systems. He received two outstanding paper awards, the UC Berkeley Departmental Fellowship in 2003, and the Microsoft Research Fellowship in 2007.