The first Dagstuhl Seminar on Theory and Praxis of Machine Learning was organized by Thomas G. Dietterich (Oregon State University), Wolfgang Maass (Technical University in Graz), Hans U. Simon (Universit¨at Dortmund), and Manfred K. Warmuth (University of California at Santa Cruz). The 34 participants came from 13 countries, 25 came from europe, 5 from north america, 2 from Israel, 1 from the GUS, and 1 from Japan.

32 lectures were given, 8 of them were related to the pac-learning model of Leslie Valiant, 6 to the on-line learning model of Nick Littlestone (which is roughly equivalent to the EQU-query model of Dana Angluin), 5 to neural nets, 4 to inductive logic programming, 4 to direct learning applications, and 5 covered still other topics in machine learning.

Two introductory tutorials were given on neural nets (on monday morning by Thomas G. Dietterich and Andreas Weigend) and on the pac-learning model (on thursday morning by Shai Ben-David and Nick Littlestone). The tutorials were adressed to participants who are not experts in these fields (which was quite helpful because the seminar brought together experts from different areas in machine learning). A special tutorial was held by Manfred K. Warmuth on monday night. He compared additive and multiplicative updating schemes for agnostic on-line learning of an unknown target function by linear functions. He presented quite new results concerning this topic, thereby probably opening a new area of research.

On Tuesday night, a panel discussion about the proper choice of models for learning and opportunities for cross-fertilization between theory and practice of machine learning took place. The discussion shed light on the gap between the needs of practically inclined people and that what can be done by theoreticians. Wednesday night, an open problem session was held.

The seminar was intended to provide a meeting place for computer scientists who explore from various points of view the possibility for computing machinery to ‘learn’. ‘Machine learning’ is a fast growing research area that attracts researchers from Theoretical Computer Science (‘Computational Learning Theory’), from Artificial Intelligence, and from various other areas such as Pattern Recognition, Neural Networks, and Statistical Physics. Unfortunately, these areas tend to organize separate conferences, and opportunities where researchers in Machine Learning from different ares can meet are very rare. One main goal of the seminar was to fill this gap.

Second, it should be pointed out that Computational Learning Theory has become by now a well-established discipline of Theoretical Computer Science within the USA, with a very succesful annual conference (COLT). On the other hand, the active researchers in Computational Learning Theory in Europe have no regular common meetings, and they rarely collaborate across borderlines. The second main goal of the seminar was to offer for European researchers an opportunity to get to know each other and to exchange ideas. This may facilitate the creation a more permanent organisation for this research community in Europe.

The above-mentioned classification of the 34 talks shows that the seminar was indeed a meeting place for experts from different areas although there was a bias towards the theory side. The gap between these areas was certainly not filled. However, the seminar helped to understand more thoroughly ‘how people from the other area are thinking’. A major problem for future meetings of this kind will be to build upon earlier meetings and, simultaneously, to enable new researchers ‘to jump on the train’. For instance, for this meeting the introductory tutorials were quite helpful. Presenting them again at a second meeting will be helpful for new participants, but probably boring for the other-ones. This problem will be taken into account within an application for a second meeting.

Creative Commons BY 3.0 Unported license