Dresden 2006 – scientific programme

Parts | Days | Selection | Search | Downloads | Help

DY: Dynamik und Statistische Physik

DY 46: Poster

DY 46.9: Poster

Thursday, March 30, 2006, 16:00–18:00, P1

Learning from examples in Neural Gas and Vector Quantization — •Michael Biehl, Anarta Ghosh, and Aree Witoelaar — Inst. of Mathematics and Computing Science, University Groningen, P.O. Box 800, 9700 AV Groningen, The Netherlands

The dynamics of training a neural gas for vector quantization in high dimensions is studied by means of methods from statistical physics. Prototype vectors for the repesentation of the data are updated either ’off-line’ from an entire set of example data, or ’on-line’ from a sequence of single data. In the first case, learning can be interpreted as to approach an equilibrium state and its typical outcome is studied in terms of macroscopic order parameters. For the on-line learning scenario, a description of the learning dynamics in terms of ordinary differential equations for the order parameters is possible. We explain the methods and present firstresults.