DISTRIBUTED COMPRESSION: THEORY AND PRACTICE

Kannan Ramchandran
EECS Department, UC Berkeley


Networking of a multitude of wireless distributed sensors is increasingly
becoming important in present day applications.  The architecture and design
of the physical layers of these evolving distributed networks have generated
much interest recently.   In this talk, we consider an important component
in this system: distributed source coding.  This refers to a class of
compression strategies for correlated information sources to be transmitted
over rate-constrained channels to a central decoder which can reliably
recover any/all of the sources of interest. The caveat however is that
these distributed correlated sources cannot directly collaborate to reduce
their common redundancy. We tackle this problem by proposing a practical
and constructive approach that is inspired by information theory and referred
to as DIstributed Source Coding Using Syndromes (DISCUS).  Our computationally
efficient algorithmic framework is built on group-codes.  We study and
analyze the construction of trellis and lattice codes for this problem as
special cases of generalized coset codes.  Some of the techniques developed
here are already being tested in a wireless-sensor testbed at Berkeley related
to the Sensor-Webs project.  Other applications include digital upgrade of
noisy analog transmission of multimedia data as well as interesting "dual"
links to multimedia data hiding.