[Colloquium] Talks This Week at TTI-C: Zhu, McAllester, Jansson, Johnson, Shen (4/18-4/22)

[Katherine Cumming]

Guest Speaker (1)
 
Speaker:  Jerry Zhu, Carnegie Mellon University
Speaker's Homepage:  http://www-2.cs.cmu.edu/~zhuxj/
 
Time:  Monday, April 18, 2005 @ 3:00pm
Place:  TTI-C Conference Room
Title:  Semi-Supervised Learning with Graphs
Abstract:
In traditional machine learning approaches to classification, one uses
only a labeled set to train the classifier. Labeled data however are
often difficult, expensive, or time consuming to obtain, as they require
the efforts of experienced human annotators. Meanwhile unlabeled data
may be relatively easy to collect, but there has been few ways to use
them. Semi-supervised learning addresses this problem by using large
amount of unlabeled data, together with the labeled data, to build
better classifiers. Because semi-supervised learning requires less human
effort and gives higher accuracy, it is of great interest both in theory
and in practice.

One can ask many questions: How to use unlabeled data? Is there a
probabilistic framework? What assumptions does it make, and how to meet
the assumptions? Can it work together with my favorite classifiers, e.g.
support vector machines? Will it work on sequential data? Can it handle
large problems? What if we have the freedom to choose which items to
label in the first place? Most importantly, does it work in practice? In
this talk I answer these questions. I present a series of novel
semi-supervised learning methods, and show unlabeled data improve
performance on text categorization, image recognition and other tasks.
The same technologies have great potential in many areas, including
speech recognition and bioinformatics. 
 
Show and Tell  Series (2)
Speaker:  David McAllester, TTI-C
Speaker's Homepage:  http://www.tti-c.org//mcallester.html
 
Time:  Tuesday, April 19th @ 12:15pm
Place:  TTI-C Conference Room
Title:  A* beats Dynamic Programming
Abstract:
This talk will describe a generalization of A* search to a wide class of
dynamic programming optimization algorithms. A dynamic programming
optimization algorithm is defined in general as a set of rules for
deriving table entries from table entries. This general definition can
be made precise using concepts from bottom-up logic programming. We show
that for any dynamic programing algorithm in this class there is also a
Dijkstra lightest derivation algorithm and an A* algorithm. We also show
that, in general, admissible A* heuristics can be constructed in a
mechanical way from abstraction functions on the assertion set (set of
possible table entries). Hierarchies of abstractions can be constructed
where each level of the hierarchy provides a heuristic for computing the
level below. All levels of the abstraction hierarchy can be computed
simultaneously on a single priority queue. These algorithmic
observations lead to an "AI architecture" for the "perception pipeline"
with a particular algorithm for allowing high level (or late) abstract
processing to guide low level concrete processing. Applications in
language and vision will be discussed.

This is joint work with Pedro Felzenszwalb. 
 
Guest Speaker (3)
 
Speaker:  Johan Jansson, Chalmers University of Technology
Speaker's Homepage:  http://www.math.chalmers.se/~johanjan/
 
Time:  Tuesday, April 19th, 2005 @ 3:00pm
Place:  TTI-C Conference Room
Title:  Solving an Elasto-Plastic Model Using DOLFIN
Abstract:
We discuss a model of elasto-plasticity supporting large displacements
based on an updated Lagrangian formulation. The model is discretized
using the finite element method. The computation is done in the FEniCS
framework, which is a collection of free software components for finite
element computation. We describe how the computation is carried out,
mainly using the DOLFIN component, which is a C++ interface to FEniCS.
We also show new developments in FEniCS, where the FEniCS Form Compiler
(FFC) is used to pre-compute much of the assembly work previously done
in DOLFIN, leading to a large speed increase in assembly. 
 
Guest Speaker (4)
 
Speaker:  Claes Johnson, Chalmers University of Technology
Speaker's Homepage:  http://www.math.chalmers.se/~claes/
 
Time:  Thursday, April 21st, 2005 @ 10:00am
Place:  TTI-C Conference Room
Title:  Irreversibility in Reversible Systems
Abstract:
We propose a new foundation of thermodynamics based on reversible
Hamiltonian mechanics in the form of the Euler equations for inviscid
fluid flow, combined with finite precision computation in the form of
General Galerkin G2. We show that the Second Law is a consequence of the
First Law combined with G2. 
 
Guest Speaker (5)
 
Speaker:  Jackie Shen, PACIM University of Minnesota
Speaker's Homepage:  http://www.math.umn.edu/~jhshen/
 
Time:  Thursday, April 21st, 2005 @ 3:00pm
Place:  TTI-C Conference Room
Title:  Variational/PDE Methods in Image and Vision Analysis
Abstract:
By quantum mechanics, all basic particles (photons, electrons, quarks,
and gluons, etc) and their interactions are governed by probabilistic or
stochastic rules. Mother Nature however presents herself with great
regularities at the macro continuum scales, including, for example,
smooth water surface-waves or soap bubbles under surface tensions.

In exactly the same manner, the variational/PDE approach to image and
vision modeling can be considered as a regularized version of the
powerful Bayesian framework in the stochastic setting. Consequently, it
is not immune to the universal balance constraint: the tradeoff between
regularity and fidelity. In this talk, I will mainly focus on several
successful variational/PDE models that I have worked on recently, and
emphasize their geometric, functional, biological, Bayesian, stochastic,
and digital implications. 
 
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 or events, please go to the TTI-C Events
<http://ttic.uchicago.edu/events/events_dyn.php>  page. 
 
 
 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20050418/6171da88/attachment.htm