Tutorial Speakers
http://www.first.gmd.de/persons/Mueller.Klaus-Robert.html
Towards Brain-Computer Interfacing
Abstract: Brain Computer
Interfacing (BCI) aims at making use of brain signals for example the control
of objects, spelling, gaming and so on.
This tutorial will first provide a brief overview of the current BCI
research activities and provide details in recent developments on both invasive
and non-invasive BCI systems. In a second part -- taking a physiologist’s point
of view -- the necessary neurological/neurophysical background is provided and
medical applications are discussed. The third part -- now from a machine learning
and signal processing perspective – shows the wealth, the complexity and the
difficulties of the data available, a truely enormous challenge. In real-time a
multi-variate very noise contaminated data stream is to be processed and
classified.
Finally,
I report in more detail about the Berlin Brain Computer (BBCI) Interface that
is based on EEG signals and take the audience all the way from the measured
signal, the preprocessing and filtering, the classification to the respective
application. BCI communication is
discussed in a clincial setting and for gaming (e.g. pacman).
Bio: Klaus-Robert Mueller received the Masters
Degree in Mathematical Physics 1989 and the PhD in theoretical computer science
in 1992, both from University of Karlsruhe, Germany. From 1992 to 1994 he
worked as a Postdoctoral fellow at GMD FIRST, in Berlin where he started to
built up the Intelligent Data Analysis (IDA) Group. From 1994 to 1995 he was a
European Community STP Research Fellow at University of Tokyo in Prof.~Amari's
Lab. From 1995 on he has been
department head of the IDA group at Fraunhofer FIRST (since 2001 GMD became
Fraunhofer) in Berlin and since 1999 he has has held a joint Professor position
of Fraunhofer and University of Potsdam in Neuroinformatics. He has been lecturing at Humboldt
University, Technical University Berlin and University of Potsdam. In 1999 he received the Annual National
Prize for Pattern Recognition (Olympus Prize) awarded by the German Pattern
Recognition Society DAGM. He serves on
the editorial board of Computational Statistics, IEEE Transactions on
Biomedical Engineering and on program and organization committees of various
international conferences. His research
areas include statistical learning theory, neural networks, kernel-based and
ensemble learning techniques, time-series analysis, blind source
separation. His present interests are
expanded to statistical denoising methods for the analysis of biomedical data,
Brain-Computer Interfacing and most recently to Gene-finding.
Stephen Boyd
http://www.stanford.edu/~boyd/
Convex Optimization and Applications
Abstract: In this talk I will give an overview of some major developments in convex optimization that have emerged over the last ten years or so, and briefly describe several typical applications. The basic idea is that convex problems are fundamentally tractable, in theory and in practice. The polynomial worst-case complexity results of linear programming have been extended to nonlinear convex optimization, and interior-point methods for nonlinear convex optimization achieve efficiencies approaching that of modern linear programming solvers. Several new classes of standard convex optimization problems have emerged, including semidefinite programming, determinant maximization, second-order cone programming, and geometric programming. Like linear and quadratic programming, we have a fairly complete duality theory, and very effective numerical methods for these problem classes.
There
has been a steadily expanding list of new applications of convex optimization,
in areas such as circuit design, signal processing, statistics, communications,
control, and other fields including machine learning. Convex optimization is also emerging as an important tool for
hard, non-convex problems. Convex
relaxations of hard problems provide a general approach for handling hard
optimization problems, with applications in combinatorial optimization and
robust optimization.
This
is joint work with Lieven Vandenberghe.
Bio:
Stephen Boyd is the
Samsung Professor of Engineering and Director of the Information Systems
Laboratory, in Stanford's Electrical Engineering Department. He received the A.B. degree in Mathematics
from Harvard University in 1980, and the Ph.D. in Electrical Engineering and
Computer Science from the University of California, Berkeley, in 1985, and then
joined the faculty at Stanford. His
interests include computer-aided control system design, and convex programming
applications in control, signal processing, and circuits.
Zach Mainen
http://www.cshl.org/public/SCIENCE/mainen.html
Neural Coding and the Olfactory System
Abstract: The sense of smell may be
more tractable than vision, but the capabilities of natural olfactory systems
greatly exceed those of artificial chemical detectors and the fundamental
nature of the olfactory neural code remains deeply controversial. The olfactory
bulb receives input from about 1000 receptor types arrayed in a stereotyped
map. Thus, odors are represented by a
distributed population code that is difficult to study using traditional
single-electrode recordings and has only recently begun to be unraveled using
molecular techniques, imaging and multi-electrode recordings. The features that are encoded in this map
and the function of lateral inhibitory connections in the bulb remain
unclear. It is commonly believed that
olfaction is a slow sense, and the role of time in olfactory coding and
computations has long been mysterious and intriguing. The complexity of spatiotemporal patterns in the olfactory bulb
led Walter Freeman to conclude that olfactory dynamics were inherently chaotic. Evidence for odor coding by temporal synchrony
as well as dynamically-evolving representations undergoing of temporal
decorrelation has been obtained by Gilles Laurent and his collaborators in the
locust, but it is not known whether similar observations hold in mammals. John Hopfield has developed a theory of
olfaction that uses phase coding to explain concentration-invariant odor
recognition and odor segmentation, but this theory also remains untested. The
motor act of sniffing is integral to the olfactory perception, and
physiological evidence suggests that the respiration (theta) cycle is also
relevant to coding and sensorimotor integration. Although the olfactory bulb has been likened to the retina, its
activity reflects not only sensory input but hunger and reward expectation,
reflecting massive reciprocal connections with neocortical and subcortical
structures. Thus, an understanding of
the olfactory system will lead to understanding of the parts of the brain with
which it interacts, including the hippocampus, orbitofrontal cortex, amygdala,
basal ganglia, and brain stem.
This
tutorial will present an overview of elements of contemporary theory and
experimental data pertinent to olfactory coding and computation.
Bio: Zach Mainen is
Assistant Professor at Cold Spring Harbor laboratory. He received his PhD from the University of California, San Diego
in 1995 for work with Terry Sejnowski at The Salk Institute on spike generation
in neocortical neurons. He subsequently
worked at Cold Spring Harbor Laboratory as a postdoctoral fellow with Roberto
Malinow and Karel Svoboda and joined the faculty in 1999. His research interests include neural
coding, neurotransmitter function, and the neural correlates of behavior. His laboratory focuses on the rodent
olfactory system. He is the recipient
of a Burroughs-Wellcome Career Develpment award and is currently a Searle
Scholar.
Real-time Object Recognition using Invariant Local Image Features
Abstract: Invariant local features provide a powerful new approach to image
matching and recognition. For the first
time, they allow for robust real-time recognition with no prior object
segmentation while allowing for high levels of occlusion. New methods have been developed for rapidly
detecting large numbers of features that are invariant to image scale,
orientation, and location, and that also carry enough information to select
potential matches in a large database of previously seen features. Reliable recognition is achieved by
identifying consistent clusters of features followed by top-down model
fitting. Probabilistic models are used
to evaluate feature matches and verify interpretations.
This
tutorial will provide an introduction to invariant local features and will also
include a real-time demonstration of a system for object recognition. Topics will include interest point
detectors, scale invariance, distinctive local descriptors, affine invariance,
illumination invariance, the Hough transform, probabilistic models, learning
with local features, model fitting, and applications to recognition, image
matching, robot localization, and motion tracking.
Bio: David Lowe is a professor of computer science
at the University of British Columbia.
He received his PhD in computer science from Stanford University in
1984. From 1984 to 1987 he was an
assistant professor at the Courant Institute of Mathematical Sciences at New
York University. From 1987 to 1995 he
was a scholar of the Canadian Institute for Advanced Research. His research interests include object
recognition, local invariant features for image matching, robot localization,
and models of human visual recognition. He is on the Editorial Board of the
International Journal of Computer Vision.
He was co-chair of the International Conference on Computer Vision (ICCV
2001 in Vancouver, Canada).
Daniel D. Lee
http://www.seas.upenn.edu/~ddlee/
Learning in
Sensorimotor Systems
Abstract: Many algorithms in machine
learning involve changing the underlying dimensionality of the data set.
Unsupervised learning techniques such as principal components analysis
typically involve dimensionality reduction, whereas supervised learning
techniques such as support vector machines can be understood as mapping the
data to a higher dimensional space. After reviewing recent machine learning
algorithms that utilize changes in dimensionality, I will show how equivalent
problems emerge in artificial sensorimotor systems. Sensory processing typically involves mapping high-dimensional
sensory inputs onto a smaller number of perceptually-relevant features, whereas
motor learning involves driving a large number of actuator parameters with a
smaller number of control variables. I
will illustrate how dimensionality plays an important role in sensorimotor
learning with demonstrations on some prototypical robotic systems.
Bio: .Daniel Lee received his bachelor's degree in
physics from Harvard in 1990 and his Ph.D. in physics from MIT in 1995. Afterwards, he was a member of the technical
staff at Bell Laboratories, Lucent Technologies, in the Biological Computation
and Theoretical Physics Departments until 2001. He is now an Assistant Professor in the Electrical Engineering
and Bioengineering departments at the University of Pennsylvania. His research focuses on trying to understand
the general principles that biological systems use to process and organize
information. He works on applying that
knowledge to building better artificial systems for vision, speech, language,
and data communications.
David Karger, Laboratory for Computer Science, EECS, MIT
http://theory.lcs.mit.edu/~karger/