Unsupervised learning is currently one of the most active research areas of Neural Computation. Although unsupervised learning algorithms like K-means clustering or principal component analysis have been known since at least fourty years major theoretical questions like generalization in unsupervised learning and the problem of data representation have found only a partial answer. New unsupervised learning algorithms as selforganizing networks, independent component analysis or deterministic annealing algorithms for vision and pattern recognition problems have been invented mostly without a proper learning theoretical framework.

With the rapid and still accelerating proliferation of huge collections of computer-accessable data, the need for a better theoretical understanding has increased significantly. Large part of the value of these data is hidden in patterns and structures that are not known a-priori. Supervised learning, the most successful theory of learning so far, is not applicable due to the lack of labeled exemplars and a new approach has to be developed for the detection of these structures.

Unsupervised learning algorithms will become a key technology to automate the detection of such implicit structure or, on a more modest scale, to assist human users in finding hidden structure in data interactively. We are witnessing a rapid surge of work in data mining, where there is a strong need for unsupervised learning techniques, but were most currently employed approaches are still in their infancy and are using rather ad-hoc techniques instead of scientifically well grounded and theoretically understood algorithms. We also expect the topic of the seminar to be useful and significant for researchers interested in basic issues of learning in the nervous system.

Scope: The proposed Dagstuhl Seminar will focus on some key issues in unsupervised learning that are directly relevant to the above needs, both from the application and the theory side:

1. Stability and Generalisation: How stable is an unsupervised model w.r.t. to fluctuations in the data and how reliable are detected structures in this respect? What is the proper way to address the issue of generalisation in unsupervised learning?
2. Bias and structure detection in reinforcement learning: since reinforcement learning is based on less information than supervised learning, the bias-variance dilemma is even more acute here. What are good and generally applicable, principled methods to bias reinforcement learning towards a detection of "reasonable" structures (e.g., characterised by certain topologies)?
3. Active Data Selection: How can a learner maximise its information gain by an active selection of samples? How can this speed up unsupervised learning and what are the implications for the creation of additional model bias?
4. Collective learning and adaptive environment: While unsupervised learning usually is considered in a fixed environment, in many realistic cases (such as in economic situations) there is a population of learners whose decisions influence both what information is offered and how the costs of decisions are set. How should learning algorithms take such situations into account and what can we benefit from neighbouring disciplines, such as game theory?
5. Performance and complexity issues: Can one find theoretical analogies for unsupervised learning to existing results in supervised learning for learning complexity, scaling properties and generalisation performance?

Focus topics: Each day will have a "focus topic" from the above list. This focus topic will serve as a departure and reference point for the discussions of the day.

Tutorials: We plan to begin each day with a tutorial to provide a condensed review of the state of the art in the focus topic of the day.

Brief presentations: These should focus on a selected issue, such as a report of a particular application, the brief exposition of a particular algorithm or methodology, or the presentation of some particular theoretical point.

Open questions: In the late afternoon, there will be a one-hour session where the participants have occasion to present their views what they feel to be major open questions in connection with the day's

focus topic.

Evening discussions: These should give an occasion to bring together the main points of the day and to follow up on the open questions raised in the last afternoon session. We will also solicit from the participants in advance demonstrations with the computer equipment available at Dagstuhl.

Goal of the workshop is to bring together practicioners and theoreticians for discussing the above issues in a way that is informed from both theory and practical case studies or application examples. We expect such information exchange to foster progress in the application and development of unsupervised learning algorithms and to open up new views that may lead to new research directions or collaborations.