Representations and Generalizations in Pattern Recognition

This page introduces my present research interest in pattern recognition.
See also my home page and my projects.

Bob Duin

Ph.D. Students:

Drs. E. Pekalska (NWO)
Ir. P. Juszczak (NWO)
Ir. P. Paclik (STW)

Drs. S. Verzakov

Drs. T.C.W. Landgrebe

Drs. A. Harol

Post-docs:

Dr. D.M.J. Tax (NWO)

Dr. M. Skurichina (STW)

Cooperations

Dr. H.J. Kappen, Medical Physics and Biophysics, University of Nijmegen.
Dr.ir. B.J.A. Kröse, Intelligent Autonomous Systems, Faculty of Scienc, Univ. of Amsterdam.
Drs. M. Loog, Image Sciences Institute, Academic Hospital, Utrecht.
Prof. S. Raudys, Inst. of Math. and Cyb., Vilnius, Lithuanian
Prof. J. Kittler, Centre for Vision, Speech and Signal Processing, University of Surrey, UK.
Dr. L.I. Kuncheva, School of Informatics, University of Wales, Bangor, UK.
Prof. A.K. Jain, Michigan State University, East Lansing, USA.

Prof. H. Bunke, University of Bern, Switzerland

Research Summary

The basic question is: what are good techniques for generalizing measument data?

In this context the traditional tools for statistical pattern recognition are studied, including or in combination with neural networks, self-organizing maps and nonlinear data mapping. It is investigated how these approaches have to be applied depending on the specific research issue and the data characteristics. An important issue is the relation between the complexity of a particular approach and its generalization capability: Simple approaches may not be powerful enough, but complicated techniques with many parameters to optimize may fit the data noise instead of its underlying structure, resulting in overtraining or the classical peaking phenomenon in pattern recognition.

Answers have to be related to the data characteristics. The complexity of the data and in particular its compactness are important. What are good ways to define and measure these? Is it possible to select from these the good analysis techniques?

Finally the data representation is studied: features and distance measures determine the above data complexity and thereby the significance of the results.

Last update: June 2004