Speaker: Manfred Opper (U Southampton)

Title:  Tractable Inference for Probabilistic Models
 
Abstract: 
Probabilistic models (aka graphical models, Bayes belief networks) provide
a powerful framework for modeling statistical dependencies between data.
Their applications range from medical expert systems over encoders in
telecommunication systems to spatial probabilistic models in
meteorology. The price that a modeler has to pay for the high degree of flexibility
of these models is the vast increase in computational complexity for
adapting the models to data and for predicting unobserved causes (hidden nodes)
from observations when the number of nodes in the network is large.
Hence, the development of tractable but approximate inference
techniques for Bayes networks is an important area of resarch.
My talk will discuss ideas for approximate inference algorithms
and their applications.

Bio: 
Manfred Opper obtained a PhD in physics in 1987 in Giessen,
Germany. After that he began to work on the application of statistical
physics to the theory of neural networks.  Following postdoctoral
visits at the Ecole Normale Superieure in Paris (1989) and at the
University of California at Santa Cruz (1990) he received the
habilitation degree in theoretical physics in 1991.  In 1992 he was
awarded the Physics Prize of the German Physical Society for work on
the dynamics of learning in neural networks. He was awarded a 3 year
Heisenberg fellowship in 1994 enabling him to work at the machine
learning group of UC Santa Cruz and at the Department of Complex
systems of the Weizmann Institute in Israel. He joined the Neural
Computing Research Group
at Aston University as a Reader in 1997. In 2004 he moved to
Southampton University where he is a reader at the ISIS research
group. His research interests are in the cross
connections between statistical physics, computational learning
theory, information theory and mathematical statistics.  He is
currently working on the theory and applications of Gaussian processes
and on the development of efficient inference techniques for complex
probabilistic models.