IMPROVING SUPPORT VECTOR LEARNING FOR PATTERN RECOGNITION WITH IRRELEVANT DIMENSIONS
Mr Danny Roobaert, Marie Curie research fellow, CVAP, Numerical Analysis and Computing Science, Royal Institute of Technology, Stockholm, Sweden

Abstract: Three methods for improving the generalisation of linear Support Vector learning algorithms are proposed for the task of high-dimensional pattern recognition with a number of irrelevant dimensions. In contrast to other methods, the improvement in generalisation is not obtained by incorporating domain knowledge in the system, but solely by selecting or generating limited amounts of appropriate training data. The first method guarantees faultless performance, while the two other methods are empirically evaluated by applying it to the task of 3D object recognition with a variety of cluttered backgrounds. We are able to report very good results with one of the approaches and since the method is also practically realisable, effectively a method for 3D object recognition without object segmentation is proposed.

Details: Monday 11 October 1999, 4 pm, Room C355, McCrea Building, Royal Holloway, University of London (see campus map). Tea is served afterwards in the Journals Room of the Computer Science Department, McCrea Building - all welcome.

back