BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:A Bit of Network Information Theory - Suhas Diggavi (EPFL)
DTSTART:20100112T150000Z
DTEND:20100112T160000Z
UID:TALK22063@talks.cam.ac.uk
CONTACT:Eiko Yoneki
DESCRIPTION:Network information theory deals with fundamental characteriza
 tions of information flow and/or compression over communication networks. 
 While there are some satisfying characterizations for flows on networks re
 presented as graphs\, there have been few complete characterizations for n
 etworks with more complicated interactions such as those encountered in wi
 reless networks. Given this discouraging state of affairs\, a natural ques
 tion to ask is whether there are methodologies to make progress on these q
 uestions.\n\nIn this talk we posit that one could gain insight into these 
 problems by asking for less\, i.e.\, by looking for (provable) approximate
  characterizations\, with the hope that in terms of engineering solutions\
 , these might actually be enough. Underlying these approximations are exac
 t results for deterministic/lossless network communications.\n\nWe illustr
 ate this approach by examining two problems. First is that of information 
 flow over wireless Gaussian relay networks\, where we show that the achiev
 able rate is within a constant number of bits from the information-theoret
 ic cut-set upper bound on the capacity of these networks. Underlying this 
 result is an exact information-theoretic max-flow min-cut characterization
  of a deterministic relay network. The second is the approximate character
 ization of the K-description Gaussian multiple description source coding p
 roblem.\nHere we show that the rate region can be sandwiched between two p
 olytopes with matching hyperplanes\, separated by at most one bit. Underly
 ing this is an exact characterization of a lossless multi-level source cod
 ing problem.\n\nThese results show that characterizations\, within a const
 ant number of bits approximation\, may be a promising direction to get ins
 ight in network information theory problems. We will conclude with implica
 tions of this methodology by illustrating several applications such as wir
 eless network secrecy.\n\nParts of this work were done in collaboration wi
 th S. Avestimehr (Cornell)\, S. Mohajer (EPFL)\, C. Tian (AT&T) and D. Tse
  (UC\, Berkeley).\n\nBio: Suhas N. Diggavi received the B. Tech. degree in
  electrical engineering from the Indian Institute of Technology\, Delhi\, 
 India\, and the Ph.D. degree in electrical engineering from Stanford Unive
 rsity\, Stanford\, CA\, in December 1998.\n\nAfter completing his Ph.D.\, 
 he joined the Information Sciences Center\, AT&T Shannon Laboratories\, Fl
 orham Park\, NJ\, where he was a Principal Member Technical Staff. He is c
 urrently on the faculty of the School of Computer and Communication Scienc
 es\, EPFL\, where he heads the Laboratory for Information and Communicatio
 n Systems (LICOS). He will join the University of California\, Los Angeles
  as a full professor in 2010. His research interests include information t
 heory\, network communications\, data compression and signal processing. M
 ore information can be found at http://licos.epfl.ch.\n\nHe is a recipient
  of the 2006 IEEE Donald Fink prize paper award\, 2005 IEEE Vehicular Tech
 nology Conference best paper award and the Okawa foundation research award
 . He is currently an Associate Editor for the IEEE/ACM Transactions on Net
 working and the IEEE Transactions on Information Theory. He has served as 
 an associate editor for the IEEE Communication Letters\, a guest editor fo
 r IEEE Selected Topics in Signal Processing and on the technical program c
 ommittee for several conferences including ISIT\, ICC\, ITW\, and Globecom
 . He has 8 issued patents.\n
LOCATION:FW26\, Computer Laboratory\, William Gates Builiding
END:VEVENT
END:VCALENDAR