Get to Know Duin

As an undergraduate student, I joined the Delft Pattern Recognition Group in The Netherlands before I had any idea what this field was about. In my study of applied physics at the Delft University of Technology, I attended a course by Prof. C.J.D.M. Verhagen on physical measurements. He explained that there are two fundamentally different ways of measuring physical properties: comparing and counting. Comparing is for human beings only, as we have to find relations with other observations on earth, e.g. the standard meter stick. It is relative. In counting (e.g. the number of pendulum swings) however, he explained with a wink, we share the angels. It is absolute. Any other creature in the universe will find the same value. Although he was partially joking, this observation impressed me deeply and stayed with me for the rest of my life.

Verhagen was an impressive person, an inspiring teacher, a great organiser, and a leader who was able to get the best out of his team. As his Ph.D. student I could observe how he contributed to the organisation of the field: he organised the 2nd ICPR in Copenhagen in 1974 and participated in the founding of the IAPR in 1976, together with Herbert Freeman and King Sun Fu. I had, thereby, the privilege of meeting these early workers in the PR field in Delft in the 70’s. Other interesting researchers visited Delft as well in that period, e.g. David Cooper, Laveen Kanal, Anil Jain, Godfried Toussaint, Pierre Devijver, Josef Kittler, and Ted Young. With the majority of them, I have had regular contact for many years. I stayed with Josef Kittler and Anil Jain during my sabbatical leaves. Ted Young later became the successor of Verhagen after his retirement. It felt very good to work in a group with worldwide connections.

I gradually found out what pattern recognition was about: a challenging starting area with great questions. It could be phrased as a natural science, as it tries to understand and simulate the pattern recognition systems around us (human experts). Statistics is a powerful tool for describing how they learn pattern classes. Bayes’ rule is very helpful, but at the end the curse of dimensionality is against us and causes a strange paradox. This was the topic of my thesis and I was rather depressed by the conclusion: unrealistic demands with respect to the size of the training sets in comparison with human experts.

After finishing my thesis I left the area of statistical pattern recognition for about 10 years and worked on parallel architectures for image processing. These were needed for pattern recognition in order to obtain large amounts of data fast. Intellectually and technically this topic was interesting, and I still find it a valuable experience that I once debugged computers on the level of micro-code and hardware. From a scientific point however it didn’t interest me sufficiently. So, after a 10-year break, it was a time to return to the field.

Around 1990, the upcoming area of neural networks caused a crisis in pattern recognition. Its claims contradicted the established principles in the pattern recognition community with respect to the complexity of learning systems and the customary size of the training sets. I was shocked and wanted to understand what was going on. Although many details in the papers I studied appeared to be wrong, I also discovered that I had to change my mind about the possibilities to train systems with many parameters or to tackle problems represented in high-dimensional spaces. My strong cooperation with Sarunas Raudys was very helpful to making progress.

The publication of the support vector classifier in 1995 gave a beautiful framework for understanding why this is possible. What disturbed me however was the Mercer condition for the kernel: the basis has to live in a Euclidean space. This was a severe restriction. I had the feeling that a much more general concept was needed for real world pattern recognition problems and started to develop the dissimilarity representation together with Elzbieta Pekalska.

It is a generalisation of the Mercer kernel and avoids the underlying feature representation. The resulting dissimilarity space equipped with a support vector classifier became my main toolset in pattern recognition. Horst Bunke made clear to me how this representation is capable of bridging the statistical and structural approaches to pattern recognition. Together with my gradually growing understanding of the insights that Lev Goldfarb already published years before, this encouraged me to make my final steps in approaching pattern recognition problems: it is non-Euclidean and in essence classes don’t overlap as they are based on different concepts. In practice they often do, of course, but by replacing the feature representation by a dissimilarity representation this can be postponed or sometimes entirely avoided.

This brought me from Bayes to Occam: the observable statistics as the fundamental generalisation principle for the design of pattern recognition systems should be replaced by human defined distances (or dissimilarities). I started to exploit this in various fields of application like computer vision, medical diagnosis, hyperspectral imaging, seismics and chemometrics.

The study of pattern recognition triggered my interest in consciousness. How do we find good features or an appropriate dissimilarity measure? By asking an expert or by introspection. The errors in the resulting classifier may show that the original answers are wrong or insufficient: we have to do better. The same holds for the more general questions: how do we learn and recognize anyway? Do we really collect and store features? Or do we directly observe object differences? Do we apply some measure for that? And how do we transform and use object differences to discriminate between classes, patterns, or concepts? How is daily consciousness organised? What levels can be distinguished? This is still a large domain waiting for exploration.

My work brought me in contact with Ph.D. students, visitors and colleagues from all continents. I have met many interesting people of whom I have named here just a few. Cooperations brought me to interesting places like the Australian desert, the top of a Colombian volcano, the Verona Arena, and the rural regions of the Korean peninsula. Next to the scientific and technical aspects, this social side of the work has brought me great satisfaction.