Parallel self-organizing maps for actual applications

In this paper, we look on parallelization of Kohonen´s self-organizing maps (SOM) for the real applications of this neural network model. Node parallelism, the parallelization strategy used by most implementors of the SOM model, is shown to give poor results when the networks are small. A presentation of the problem sizes for actual applications is also given, showing that the problems often are in the range where the node parallel algorithms perform poorly. More attention should be paid to speeding up the smaller problems. Our own implementations suggest that a combination of node parallelism and training example parallelism should be used in order to reduce execution times for the smaller applications of the SOM model