Motivation and Technical Description
The field of meta-learning has as its primary goal the understanding of the interaction between the mechanism of learning and the concrete contexts in which that mechanism is applicable. Meta-learning differs from base-learning in the scope of the level of adaptation. Whereas learning at the base-level focuses on accumulating experience on a specific learning task (e.g., credit rating, medical diagnosis, mine-rock discrimination, fraud detection, etc.), learning at the meta-level is concerned with accumulating experience on the performance of multiple applications of a learning system.
A practical goal of meta-learning is to allow machine learning and data mining algorithms be more widely used outside research labs; our community must be able to design robust systems that offer support to practitioners (e.g., by matching tasks or domains with proper learning algorithms).
The aim of this workshop is to offer the international community a forum to assess the state-of-the-art, exchange experience, knowledge and perspectives in meta-learning, and provide impetus to the field.
This workshop continues in the tradition of previous related workshops (e.g., ECML98, ICML99, ECML00). It will allow participants a chance to better understand the various facets of the topic, and its relevance both to the theory of machine learning and its applications. As the community continues its efforts to broaden the applicability of machine learning, meta-learning will be instrumental in expanding the usability of current learning tools. |
Back to top
|
Call for papers
Contributions (from all main sub-fields of ML) describing work in progress as well as position papers are invited. All contributions must focus on the automation of machine learning and meta-learning. Of particular interest are methods and proposals that address the following issues:
- Given a task (i.e., a dataset and a learning/mining objective, and possibly a particular expert user), which are the learning algorithms best suited for such task?
- What contribution may be made by natural and physical sciences (e.g., observe each algorithm as an entity and characterize its functioning modes), as a complementary approach to statistical learning?
- What can be learned or transferred from the stochastic complexity analysis developed in the neighbor community of constraint satisfaction problem, known as the phase transition paradigm?
- What criteria and metrics can be used for evaluating and automating model selection in classification and regression? What is the cost of these metrics? - How can expert knowledge be integrated with meta-learning?
- What are the requirements for a dynamic, incremental meta-learning system (e.g, multi-agent-based architectures)?
- What theoretical results are known regarding the feasibility of meta-learning?
- How do we move from single function approximation to learning families of functions (e.g., knowledge or inductive transfer, multi-task learning)?
- What new results exist in the field of learning-to-learn?
Presentations of beta versions of meta-learning tools for automated (or guided) use of methods or algorithms with respect to performance or run stability are also welcome.
Papers must be submitted electronically, preferably in PDF or postscript (paper length is between 4-8 pages using the conference format), to one of the organizers (see email addresses below). Submitted papers will be reviewed by at least two independent referees from the Program Committee. Accepted papers will be published in the Workshop Notes and will be allocated about 30 minutes for oral presentation.
** NEW EXTENDED DEADLINE FOR SUBMISSION **: April 8, 2005
Notification of acceptance: April 22, 2005
Final camera-ready paper: May 13, 2005 |
Back to top
|
Organization
Christophe Giraud-Carrier, Brigham Young University
cgc@cs.byu.edu
Ricardo Vilalta, University of Houston
vilalta@cs.uh.edu
Pavel Brazdil, LIACC, University of Porto
pbrazdil@liacc.up.pt |
Back to top
|
Program Committee:
Jonathan Baxter, Panscient Technologies, USA
Abraham Bernstein, University of Zurich, Switzerland
Juan Botia, Universidad de Murcia, Spain
Pavel Brazdil, LIACC, University of Porto, Portugal
Philip Chan, Florida Institute of Technology, USA
Christophe Giraud-Carrier, Brigham Young University, USA
Melanie Hilario, Univeristy of Geneva, Switzerland
Dan Oblinger, IBM TJ Watson, USA
Bernhard Pfahringer, Austrian Research Institute for AI, Austria
Michele Sebag, LRI, University of Paris-sud, France
Maarten van Someren, University of Amsterdam, The Netherlands
Ricardo Vilalta, University of Houston, USA
Gerhard Widmer, Johannes Kepler University Linz, Austria |
|
Back to top
Final Programe:
8:45 - Welcome.
9:00 - Paper Session 1: Model Selection and Combination.
* Predicting Relative Performance of Classifiers from Samples Using Learning Curves and Adaptation.
Rui Leite and Pavel Brazdil (University of Porto, Portugal).
* A Proposed Meta-learning Framework for Algorithm Selection Utilising Regression-based Landmarkers.
Daren Ler, Irena Koprinska and Sanjay Chawla (University of Sydney, Australia).
* Estimating The Potential for Combining Learning Models.
Adam Peterson and Tony Martinez (Brigham Young University, USA)
10:30 - Break.
11:00 - Invited Contribution and Extended Discussion.
* Toward a Justification of Meta-learning: Is the No Free Lunch Theorem a Show-stopper?
Foster Provost (Stern School of Business, NYU, USA).
12:30 - Lunch.
14:00 - Paper Session 2: Continuous Learning and Transfer.
* Multi-Task Learning and Transfer: The Effect of Algorithm Representation.
Peter R.C. Lane and Fernand Gobet (University of Hertfordshire and University of Brunel, UK).
* Learning Reductions Automate Learning.
John Langford (Toyota Technical Institute at Chicago, USA).
* Continued Practice and Consolidation of a Learning Task.
Robert J. O'Quinn, Daniel L. Silver and Ryan Poirier (Acadia University Wolfville, Canada).
15:30 - Break.
16:00 - Paper Session 3: Other Topics.
* Low Replicability of Machine Learning Experiments is not a Small Data Set Phenomenon.
Remco Bouckaert (University of Waikato, New Zealand).
* Efficient Feature Construction by Meta Learning - Guiding the Search in Meta Hypothesis Space.
Ingo Mierswa and Michael Wurst (University of Dortmund, Germany).
* Comparing learning approaches to coreference resolution. There is more to it than 'bias'.
Veronique Hoste and Walter Daelemans (Universiteit Antwerpen, Belgium).
17:30 - Concluding Remarks.
ICML 2005 Main Conference
|
|
