Adaptive Distributed Source
Coding
David Varodayan
Department of Electrical Engineering
Stanford University
Friday, April 30, 2pm
Location: BAB024 (Bahen Centre Basement)
Abstract:
Distributed source
coding is a type of data compression that arises naturally in networks,
ranging from peer-to-peer systems to camera arrays. The principle is to
exploit the correlation of multiple sources of data at a joint decoder, so
that the encoders can compress efficiently without communicating with each
other. This is well-suited
to network settings in which communication among encoders comes at a cost.
In this talk, we introduce novel algorithms to address the unpredictable
statistics of nonergodic image and video signals. At the encoding side, our
invention of rate-adaptive distributed low-density parity-check codes lets the
encoders switch among encoding rates flexibly. When these codes are combined
with feedback from the decoder, the encoders can be agnostic of statistics. At
the decoding side, we propose an iterative unsupervised learning algorithm to
estimate statistical parameters that relate the multiple sources. These
techniques make distributed source coding much more versatile for real
signals. We then apply these algorithms to a variety of networked media
systems: multiview image coding for camera arrays, low-complexity video
encoding, image authentication for insecure distribution networks, and quality
monitoring of streamed video.
Bio:
David
Varodayan received the Ph.D. in Electrical Engineering at Stanford University in
2010. He also holds a B.A.Sc. in Engineering Science from the University of
Toronto, and an M.S. in Electrical Engineering from Stanford. His research
interests are in the areas of data compression and machine learning, applied
particularly to video coding and media security. He has won Best Student Paper
Award on two occasions and the Signal Processing Journal's Most Cited Paper
Award in 2009. In June 2010, he will begin at post-doc at HP Labs in Palo Alto.
Host of Talk:
Ashish Khisti (akhisti@comm.utoronto.ca)