
Principal Manifolds 2006 

August 2426, 2006,
Lecture room 119
You are welcome!
Registration
and coffee 24.08.2006, at 9:30,
Opening
24.08.2006, at 10:00
International
Organising Committee
Speakers
that already confirmed their interest
The problems of Large Data Sets analysis and visualisation, model reduction and the struggle with complexity of data sets are important for many areas of human activity. There exist many scientific and engineering communities that attack these problems from their own sides, and now special efforts are needed to organize communication between these groups, to support exchange of ideas and technology transfer among them. Heuristic algorithms and seminal ideas come from all application fields and from mathematics also, and mathematics has a special responsibility to find a solid basis for heuristics, to transform ideas into exact knowledge, and to transfer the resulting ideal technology to all the participants of the struggle with complexity.
The workshop “Principal manifolds for data cartography and dimension reduction,Â” will be focused on modern theory and methodology of geometric data analysis and model reduction. Mathematicians, statisticians, engineers, software developers and advanced users form different areas of applications will attend this workshop.
During the Workshop the following activities will take place:
Â·Â Â Â Â Â Â Â Talks of internationally renowned experts;
Â·Â Â Â Â Â Â Â Demonstrations of the most uptodate software;
Â·Â Â Â Â Â Â Â Demonstration of mapping and spatial analysis software and other GIS products;
Â·Â Â Â Â Â Â Â Active seminar in which different groups are given the same data to analyse and then they speak about their analysis of the data.
Before coming to the workshop we propose to have a look at three datasets containing results of a highthroughput experimental technology application in molecular biology (microarray data). Principal component analysis and principal manifolds are highly demanded methods for analysis of this kind of data. It would be very interesting if you could apply your own methods to all or some of these datasets and present the results at the workshop such that we could compare them together.
The task could be free exploratory data analysis: any kind of message about the structure of the point distribution and relation of this structure to the proposed ab initio gene and sample classifications is interesting.
The multidisciplinary character of
the workshop and the broad theme of discussion aims at eliminating some of the
traditional language barriers
that, unnecessarily sometimes, impede scientific cooperation and interaction of
researchers across disciplines.
International Organising Committee
Prof Chris Brunsdon, Department of Geography,
Prof. JeanClaude Golinval,
Prof. Alexander Gorban, Centre for Mathematical Modelling,
Dr. Uwe Kruger, Intelligent Systems and Control Research
Group, Queen's
Prof Jeremy Levesley, Department of Mathematics,
Prof Steven Damelin, IMA,
Prof. Donald Wunsch, Distinguished Professor of Computer
Engineering at the
Dr. Andrey Zinovyev, Bioinformatics service, Institute
Curie,
Speakers that already confirmed their interest
Prof. Donald Wunsch,
Distinguished Professor of Computer Engineering at the
Dr. A. Zinovyev,
Bioinformatics service, Institute Curie,
Prof. Alexander
Gorban, Centre for Mathematical Modelling, University of Leicester,
Dr. Hujun Yin,
School of Electrical and Electronic Engineering, The
Dr. StÃ©phane Girard,
INRIA
Prof Steven Damelin,
IMA,
Prof Boris
Mirkin Birkbeck College,
Dr. Jochen Einbeck,
Department of Mathematics, National
Prof. Joachim Selbig,
Dr. Uwe Kruger, Intelligent
Systems and Control Research Group, Queen's
Dr. Andrew R. Webb, QinetiQ,
Malvern,
Prof. Colin Fyfe, Department of Computing and Information Systems,
Dr. David A. Elizondo, School of
Computing, De Montfort University,
Prof. Rodolphe Sepulchre, Department
of Electrical Engineering and Computer Science,
Prof. Hans Troger, Institut fÃ¼r Mechanik Technische UniversitÃ¤t
Wien
Prof. Balazs Kegl, Dept. of Computer Science,
Dr. LizÂ Stuart, School of Computing, Communications and
Electronics (SCCE),
D. Wunsch will review the cluster analysis technique and its connection to dimension reduction approaches with application to homeland security.
A. Zinovyev will demonstrate a principal manifold based technology for Microarray visualization and analysis.
A. Gorban will present connections between nonlinear Galerkin approximations in physical applications and principal manifolds methods for data analysis, with applications for mesoscale modeling.
Hujun Yin will discuss adaptive data projection and visualisation methods, and connections between principal manifolds and SOM for various applications.
Steven. B. Damelin is expected to give talk on dimension reduction and its connections to Paley Weiner theorems in scattering and imaging.
Boris Mirkin will give
a talk
Â“Principal cluster analysis methods in different settings.Â”
S. Girard will present a talk about autoassociative models and generalized principal component analysis.
J. Einbeck will present a recent development of principal curves algorithms, a talk Â“Local principal curves.Â”
J. Selbig will present a talk
"Integrated analysis of metabolite, gene expression, and other profile
data: Missing value estimation, dimension reduction, clustering, and classification".
U. Kruger will present principal manifolds and nonlinear principal component based methods for multivariate statistical process control with example of application for diagnosis of faults on internal combustion engines.
Colin Fyfe will present a talk Â“A family of topology preserving mappings for data visualisationÂ”.
D.A. Elizondo will give a talk Â“Geometrical approaches for
Artificial Neural NetworksÂ”
Rodolphe Sepulchre will present optimizationbased numerical
algorithms on manifolds with applications to PCA and
Balazs Kegl will give a general overview of manifold
learning, presenting the achievements of the last ten years and the challenges
of future research.
A. Steindl and H. Troger, Â“Invariant manifolds
in the dimension reduction for fluid conveying tubesÂ”
LizÂ Stuart will
present a talk Â“An Overview of Visualization Techniques for the
Analysis of Large DatasetsÂ”
Ludger Evers (Department of Statistics,
The workshop will deal with the subjects of
principal manifolds, dimension reduction, data visualization and their applications.
One of the goals is to investigate new and innovative concepts of defining
principal curves/surfaces and to discuss new GISinspired technologies of data
representation on these surfaces.
Principal manifolds were
introduced by Hastie and Stueltze in 1984, 1989 as lines or surfaces passing
through ``the middle'' of the data distribution [1]. This intuitive definition
was supported by mathematical notion of selfconsistency: every point of the
principal manifold is a conditional mean of all points that are projected into
this point. In the case of datasets only one or zero data points are projected
in a typical point of the principal manifold, thus, one has to introduce
smoothers that become an essential part of the principal manifold construction
algorithms.
Since the pioneering work of
Hastie, many modifications and alternative definitions of principal manifolds
have appeared in the literature. Many alternative definitions were introduced
(see, for example, [2]) inÂ order to
improve the situation and to allow the construction of principal curves
(manifolds) for a distribution of points with several finite first moments. A
promising approach is based on analogy of principal manifold and elastic
membrane. The idea of using the elastic energy functional for principal
manifold construction in the context of neural network methodology was proposed
in mid 1990s (see [3] and bibliography there). Another computationally
effective and robust algorithmic kernel for principal curve construction,
called the polygonal algorithm, was proposed by Kegl et al [2]. A variant of
this strategy for constructing principal graphs was also formulated in the
context of the skeletonization of handwritten digits. An interesting approach
we would also like to mention is the construction of principal manifolds in a
piecewise manner by fitting unconnected line segments [4].
Perhaps most scientific and
industrial applications of principal manifold methodology were implemented
using the SOM (selforganizing map) approach, coming from the theory of neural
networks [5]. These applications are too numerous to be mentioned here. We only
mention that SOMs, indeed, can provide principal manifold approximations (for
example, see [6]) and are computationally effective. The disadvantage of this
approach is that it is entirely based on heuristics; also it was shown that in
the SOM strategy there does not exist any objective function that is minimized
by the training process [7].
In applications of principal
manifolds to 3Dsurface modeling, one can find similar Â“physicsbasedÂ” new
methods in surface modeling in computer graphics (see, for example [8], [9]). A parametric model based on the generative topographic mapping, and the expectationminimization
algorithm to optimize the parameters was analyzed in [10].
One important
application of principal manifolds is dimension reduction and data
visualization. In this field they compete with multidimensional scaling methods
and the recently introduced advanced algorithms of dimension reduction, such as
locally linear embedding (LLE) [11] and ISOMAP [12] algorithms. The difference
between the two approaches is that the later ones seek new point coordinates
directly and do not use any intermediate geometrical objects. This has several
advantages, in particular that a) there is a unique solution to the problem
(the methods are not iterative in their nature, there is no problem of grid
initialization) and b) there is no problem of choosing a good way to project
points onto a nonlinear manifold. Another advantage is that the methods are
not limited by several first dimensions in dimension reduction (it is difficult
in practice to manipulate nonlinear manifolds of dimension more than three).
Using principal manifolds as a nonlinear lowdimensional screen to
project data gives other benefits. First, the manifold approximates data and
can be used itself, without applying projection, to visualize different
functions defined in data space (for example, density estimation). Also the
manifold as an intermediate object for Â“fixingÂ” the structure of a learning
dataset, can be used in visualization of data points that were not used in the
learning process, for example, for visualization of dataflow Â“on the flyÂ”. Here
we can find many links to modern GIS technologies.
LLE and
ISOMAP methods are more suitable if the lowdimensional structure in
multidimensional data space is complicated but is expected to exist, and if the
data points are situated rather tightly on it. Principal manifolds are more
applicable for the visualization of reallife noisy observations, appearing in
economics, biology, medicine and other sciences, and for constructing data
screens showing not only the data but different related functions defined in
data space.
Â“The data mining community has tended to focus on the algorithmic aspects
of pattern discovery, and has not developed a general underlying theoretical
base. However, such a base is important for any technology: it helps to steer
the direction in which the technology develops, as well as serving to provide a
basis from which algorithms can be compared, and to indicate which problems are
the important ones waiting to be solved.Â” [14] Principal manifold theory is an
important element in this base.
During the
workshop modern development of this direction will be present, as well as many
recent applications, from engineering to bioinformatics.
1.Â Â Â Â Â T. Hastie, W. Stuetzle, Principal curves, Journal of the American Statistical Association 84 (406) (1989), 502516.
2.Â Â Â Â Â B. Kegl, A. Krzyzak, T. Linder, K. Zeger, A polygonal line algorithm for constructing principal curves, Neural Information Processing Systems 1998, MIT Press, 1999,Â 501507.
3.Â Â Â Â Â Gorban, A. Zinovyev, Elastic Principal Graphs and Manifolds and their Practical Applications, Computing 75 (2005), 359Â–379. (See also)
4.Â Â Â Â Â J.J. Verbeek, N. Vlassis, B. Krose, A ksegments algorithm for finding principal curves, Technical report, 2000. Online: http://citeseer.nj.nec.com/article/verbeek00ksegments.html.
5.Â Â Â Â Â T. Kohonen: Selforganized formation of topologically correct feature maps, Biological Cybernetics, 43 (1982), 5969.
6.Â Â Â Â Â F. Mulier, V. Cherkassky: Selforganization as an iterative kernel smoothing process,Â Neural Computation 7 (1995), 11651177.
7.Â Â Â Â Â
E. Erwin,
K. Obermayer, K. Schulten, Selforganizing maps: ordering, convergence
properties and energy functions, Biological Cybernetics 1992 (67), 4755.
8.Â Â Â Â Â Chang K, Ghosh J, A unified model for probabilistic principal surfaces, IEEE Transactions on Pattern Analysis and Machine Intelligence 23 (1): 2241.
9.Â Â Â Â Â Mandal C., Qin H., Vemuri B.C. A novel FEMbased dynamic framework for subdivision surfaces. 2000. Comp.Aided Design 32.Â p.479497.
10.Â Xie H., Qin H. A Physicsbased
Framework for Subdivision Surface Design with Automatic Rules Control. In Proceedings
of the Tenth Pacific Conference on Computer Graphics and Applications
(Pacific Graphics 2002), IEEE Press, p.304315.
11.Â Roweis S. and Saul L. K. Nonlinear
dimensionality reduction by locally linear embedding. 2000. Science 290. p. 23232326.
12.Â Tenenbaum J. B., De Silva V. and Langford J. C. A global geometric framework for nonlinear dimensionality reduction. 2000. Science 290. rp.23192323.
13.Â V. Pestov, On the geometry of similarity search: dimensionality curse and concentration of measure, Information Processing Letters 73 (2000), 4751.
14.Â Hand, D. J. and R. J. Bolton (2004) Pattern Discovery and Detection: A Unified Statistical Methodology, Journal of Applied Statistics, 31(8), 885924.
Contact:
Prof. Alexander Gorban Telephone  +44 (0) 116
223 14 33 
Postal: Department of
Mathematics,
