[Colloquium] Reminder: today's talk by Erik Learned-Miller

[Margery Ishmael]

DEPARTMENT OF COMPUTER SCIENCE

Date: Wednesday, March 10, 2004
Time: 2:30 p.m.
Place: Ryerson 251

-------------------------------------------

Speaker:  ERIK LEARNED-MILLER, University of California at Berkeley

Url: http://www.eecs.berkeley.edu/~egmil/

Title:   Entropy Minimization Methods for Finding Independent Causes in 
Data

Abstract

When two independent random variables are added, the entropy of the sum 
is
greater than the entropy of either of the original variables. The 
reversal of this
process, finding combinations of dependent random variables with lower 
entropy
than the originals, can be used to develop factored probability models 
and to find
independent "causes" of data.

In this talk, we present three applications of this idea. The first is 
to
the classic Blind Source Separation problem, an application of the 
widely
used Independent Components Analysis (ICA). We present a new ICA 
algorithm
which is computationally efficient and significantly outperforms all 
algorithms against
which it has been tested.

Second, we show how entropy minimization methods can be used to pull 
apart
variability in images due to a certain set of spatial transformations 
(like rotation and
translation). We develop factored models of handwritten digits, and 
show how
these factored models can be used to develop classifiers, do 
unsupervised learning,
and transfer learned concepts from one learning task to another.

Finally, we present recent work in which we use the same ideas to 
remove bias
fields from magnetic resonance images. We use an entropy minimization 
procedure
to separate patients' true anatomy from the independent artifacts due 
to the scanner
apparatus, resulting in improved medical images.

-------------------------------------------

Host: YALI AMIT

*Refreshments will follow the talk in Ryerson 255*

People in need of assistance should call 773-834-8977 in advance.