Comparing the PLSOM and the SOM over normal distributed input spaces

The Self-Organizing Map (SOM) [11, 15] is a method for mapping data relationships and distributions in high dimensions to lower dimensions, where they are easier to visualize and process. One of the problems with the SOM is that it needs an externally applied annealing scheme to learn mappings. There is no firm theoretical guidance for selecting annealing schemes and their parameters, something that must therefore be done by trial-and-error which can consume considerable effort, as any of the many parameters that govern the annealing scheme are dependent on the application and can have large influences on the outcome [19]. In addition the SOM has problems with non-uniformly distributed input spaces, as we will demonstrate in this paper. The PLSOM solves some of the problems associated with applying SOMs to non-uniformly distributed input spaces and eliminates the need for an externally enforced annealing scheme. This paper explores the performance of the Parameter-Less SOM (PLSOM) when mapping a normal distributed input space, as compared to two common versions of the ordinary SOM.