[Colloquium] Guest Speakers @ TTI-C Next Week (2/27/06-3/3/06)

[Katherine Cumming]

**********TTI-C Guest Speakers (4) Next Week***********
                      February 27 - March 3, 2006
        Presented by:  Toyota Technological Institute at Chicago
 
(1)
 
TTI-C Show and Tell Series
Speaker: Cristian Sminchisescu, TTI-C
Speaker's home page: http://www.tti-c.org/sminchisescu.html
 
Date: Monday, February 27, 2006 
Location: TTI-C Conference Room
Time:  12:00 pm (Lunch)
           12:15pm (Talk)
 
Title:  Generative and Conditional Models for Visual Inference
 
Abstract:
 
In this talk, I will discuss ongoing work on using conditional models of
various kinds for visual reconstruction and recognition, and I will show
applications to 3d human motion analysis in monocular video.  I will
illustrate the key ingredients that make the computation efficient,
including the use of observation context, sparsity, low- dimensionality and
mixture modeling.  I will also describe some of the limitations of
conditional models, and show how more reliable estimates can be obtained by
jointly learning bidirectional, conditional-generative tandems.
 
 
(2)
 
Speaker:  Ryan O'Donnell, Microsoft Research Theory Group
Speaker's home page: http://research.microsoft.com/~odonnell/
 
Date: Tuesday, February 28, 2006 
Location: TTI-C Conference Room
Time:  10:00 am
 
Title: Constraint Satisfaction and Property Testing
 
Abstract: 
 
For most constraint satisfaction problems such as finding the maximum cut in
a graph, or trying to solve an over constrained system of equations, it is
hard to find an optimal solution.  Nevertheless, one still wants to find
algorithms that come as close to the optimum as they can.  Unfortunately,
the best-known algorithms and the best-known hardness results do not yet
match for many basic problems. 
In this talk, I will talk about recent attempts to find the sharp boundaries
between P and NP for approximating constraint satisfaction problems.
Interestingly, progress on proving hardness results mostly comes by finding
more efficient algorithms in the area of "property testing''.  I will
describe new such algorithms and mention how their analysis is connected to
such diverse areas as voting theory and the "double bubble'' problem.  As an
example consequence of these connections, we get complexity-theoretic
evidence that there is no efficient algorithm for finding the maximum cut in
a graph with guarantee better than Goemans-Williamson's: $87.8567 \dots$\%. 
 
 
(3)
 
Speaker:  Kristen Grauman, MIT-CSAIL
Speaker's home page: http://people.csail.mit.edu/kgrauman/
 
Date: Thursday, March 2, 2006 
Location: TTI-C Conference Room
Time:  10:00am
 
Title:  Efficient Matching for Recognition and Retrieval
 
Abstract:
 
Local image features have emerged as a powerful way to describe images of
objects and scenes.  Their stability under variable image conditions is
critical for success in a wide range of recognition and retrieval
applications.  However, comparing images represented by their collections of
local features is challenging, since each set may vary in cardinality and
its elements lack a meaningful ordering.  Existing methods compare feature
sets by searching for explicit correspondences between their elements, which
is too computationally expensive in many realistic settings.  I will present
the pyramid match, which efficiently forms an implicit partial matching
between two sets of feature vectors.  The matching has linear time
complexity, naturally forms a Mercer kernel, and is robust to clutter or
outlier features, a critical advantage for handling images with variable
backgrounds, occlusions, and viewpoint changes.  I will show how this
dramatic increase in performance enables accurate and flexible image
comparisons to be made on large-scale data sets, and removes the need to
artificially limit the size of images' local descriptions.  As a result, we
can now access a new class of applications that relies on the analysis of
rich visual data, such as place or object recognition and meta-data
labeling.  I will provide results on several important vision tasks,
including our algorithm's state-of-the-art recognition performance on a
challenging data set of object categories.
 
 
(4)
 
Speaker: Tong Zhang, Yahoo
Speaker's home page: http://www-cs-students.stanford.edu/~tzhang/
 
Date: Thursday, March 2, 2006 
Location: TTI-C Conference Room
Time:  3:00pm
Title:  Learning with Structured Inputs
Abstract:
I will present a novel approach to semi-supervised learning that employs a
method, which we refer to as structural learning (aka multi-task learning).

The idea is to learn predictive structures from many auxiliary problems that
are created from the unlabeled data (and are related to the target problem),
and then transfer the learned structure to the supervised target problem.
In the first part, I will give a high-level description of the general
approach, a specific bi-linear structure model we use, and then explain how
we generate auxiliary problems that are related to the target task.  I will
then show some empirical results.  For example, this method produces
performance higher than the previous best results on some standard NLP
benchmark tests.  It is also highly effective for some other problems such
as image classification and even information retrieval.  In particular, I
will describe our successful participation in 2005's TREC genomics ad hoc
retrieval task using this idea.

In the second part, I will present a general framework for learning
structures including its learning theoretical analysis.  Under this
framework, I will investigate the theoretical justification of the bi-linear
structure model used in our experiments, and show that the resulting
formulation can be solved by an iterative SVD procedure.  The relationship
of this framework and transfer learning, and Bayesian hierarchical models
will be discussed.
----------------------------------------------------------------------------
------
If you have questions, or would like to meet the speaker, please contact
Katherine at 773-834-1994 or 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/20060224/1a51a617/attachment.htm