Abstract:

The present contribution discusses the benefits of topography in neural maps. In data representation and processing by self-organizing maps (SOM) incompatibilities between the data manifold and the output configuration may occur, which are due to discontinuities in the data or to dimensional mismatches. In classification problems the chance for discontinuities to occur is highest at the class boundaries. This gave rise to a supervised version of SOMs, where the data are represented by independent class-specific maps. As an illustration three standard classification tasks have been considered. Using this supervised version of SOM for two of the data sets that are effectively low-dimensional, a considerable increase in classification accuracy was achieved in comparison to other methods, which was possible by avoiding to represent in the map discontinuities occurring at class boundaries.