Dynamic Sparse Tracking with Applications in Acoustic Communications

---------------------------------------------------------------------
                       AI Seminar
----------------------------------------------------------------------
Speaker:        Dr. Weichang LI
                ExxonMobil Corporate Strategic Research

Title:          "Dynamic Sparse Tracking with Applications in Acoustic
                Communications"

Date:           Thursday, 17 May 2012

Time:           11:00am - 12 noon

Venue:          Rm4504 (via lifts 25/26), HKUST

Abstract:

In many engineering applications, the assumed model not only has
structures such as sparsity, but also can be highly dynamic. Model
dynamics induce variations in parameter values and can also destablize its
structure. Direct application of many of the recently developed sparse
learning algorithms and analysis without recognizing model dynamic leads
to structure smearing, and therefore poor performance. On the other hand,
an intrinsically sparse model leads to convergence issues in joint state
estimation and model identification, due to the fact that the model
parameters associated with close-to-zero state components are ill-defined
or insufficiently excited. This talk presents two sets of algorithms:
sparse EKF and sparse EM algorithms, to overcome these issues and
dynamically track potentially highly sparse systems. Both algorithms may
be viewed as extension of sparse regression algorithms, such as lasso,
elastic net and group lasso, to a dynamic model setting. Results from
channel estimation and equalization in broadband acoustic communications
will be presented to illustrate the performance of these algorithms.

Note: this material is based on part of the author's thesis work at MIT,
jointly with Dr. Jim Preisig.

*****************
Biography:

Weichang Li is currently with ExxonMobil Corporate Strategic Research,
working in the areas of machine learning, signal processing and acoustic
sensing. He obtained his Ph.D. in Electrical and Oceanographic Engineering
from MIT in 2006, followed by an ONR postdoctoral fellowship at Woods Hole
Oceanographic Institution before his current position. His past work
involves statistical signal processing, underwater acoustic
communications, and 3D bioimaging via computational holography.