[Colloquium] Guest Speakers @ TTI-C This Week (4/17/06-4/21/06)

[Katherine Cumming]

**********TTI-C Guest Speakers This Week***********
                             April 17 - April 21, 2006
        Presented by:  Toyota Technological Institute at Chicago
 
(1)
 
Speaker:  Alexander Ihler, University of California, Irvine
Speaker's home page:  http://www.ics.uci.edu/~ihler/
 
Date: Monday, April 17, 2006 
Location: TTI-C Conference Room, Part of Theory Seminar
Time: 10:00 am      
Title:  Loopy Belief Propagation: Convergence and Approximations
Abstract:
Graphical models provide a convenient method of describing probabilistic
distributions and statistical dependency in such a way as to enable the
development of relatively simple exact or approximate inference algorithms.
Perhaps the best known of these algorithms is belief propagation (BP), which
formulates an exact procedure for marginalization, in simple (cycle-free, or
tree-structured) systems in such a way that it can be easily applied to more
complex problems.
 
Although this has often resulted in great success in such problem domains as
turbo and LDPC decoding, computer vision, machine learning, and sensor
networks, on these more complex problems so-called "loopy" BP is no longer
guaranteed to converge to a unique solution.
 
However, intuition and empirical findings tell us that loopy BP will be well
behaved on problems which are tree-like, and that there are several factors
which may contribute to this being the case.  By performing a stability
analysis of BP we can quantify many of these notions and derive powerful
sufficient conditions for convergence; these conditions appear to be tight
for at least a subset of problems.
 
Moreover, the framework in which this analysis is performed makes it
possible to draw many additional conclusions, including bounding the number
of iterations required to achieve a particular precision or analyzing the
performance of quantized, censored, or otherwise modified BP-like
algorithms.
 
Alexander Ihler received his B.S. degree in Math and Electrical Engineering
from the California Institute of Technology in 1998, and S.M. and Ph.D.
degrees in Electrical Engineering and Computer Science from the
Massachusetts Institute of Technology in 2000 and 2005 as a member of the
Stochastic Systems Group.  He is currently a Postdoctoral Scholar at the
University of California, Irvine.  His research interests include
statistical signal processing, machine learning, and nonparametric
statistics, with applications in computational biology, atmospheric science,
and sensor networks. 
 
(2) 
 
Speaker: Jie Liang, Department of Bioengineering, UIC
Speaker's home page:
http://www.uic.edu/depts/bioe/faculty/j_liang/index.htm
 
Date: Monday, April 17, 2006
Time: 2:00pm
Location: TTI-C conference room
 
Title:
Predicting protein functions through evolutionary models of geometrically
computed structural binding surfaces
 
Abstract:
 
Predicting biological roles of proteins and classifying them by their
functions are challenging tasks, as global protein sequence similarities are
often unreliable for functional inference.  Protein plays its role by
interacting with other molecules, and protein structures provide direct
useful information.  The three dimensional structures of proteins are often
thought to be tightly packed as in solids, but closer examination using
revealed that they contain interesting geometric features such as voids and
pockets.  We discuss the origin of their existence, their relationship with
folding, and how to distinguish those voids and pockets that are involved in
important biological functions such as binding from other voids and pockets.
In order to discover binding pockets on protein structures by similarity to
known binding site, scoring matrix such as PAM and BLOSUM are not suitable,
because residues on protein functional surfaces experience different
selection pressure than residues in folding core.  We develop methods for
estimating replacement rates of residues based on a continuous time Markov
model using Bayesian Markov chain Monte Carlo.  Combined with geometrically
computed libraries of millions of binding surfaces using alpha shape, we
show our method can predict protein functions from structures with
sensitivity and specificity.  We give examples of predicting functions of
orphan structures from structure genomics project where protein structures
are solved before their biological functions are fully characterized.  We
further discuss how to construct models to account for possible
cross-reactivities of proteins to multiple substrates, and examples that
peptide library can be designed based on structures.  (please visit
http://www.uic.edu/~jliang for further information).
 
(3)
 
Speaker:  Amal Ahmed, Harvard University  
Speaker's home page:  http://www.eecs.harvard.edu/~amal/
 
Date: Wednesday, April 19, 2006 
Location: TTI-C Conference Room 
Time:  10:00 am
 
Title:   TBA
Abstract:
TBA
----------------------------------------------------------------------------
------
If you have questions, or would like to meet the speaker, please contact
Katherine at 773-834-1994 or  <mailto:kcumming at tti-c.org> kcumming at tti-c.org

For information on future TTI-C talks and events, please go to the TTI-C
Events page:  http://www.tti-c.org/events.html.  TTI-C (1427 East 60th
Street, Chicago, IL  60637)
 
 
 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20060417/7d96b391/attachment.htm