Performance improvements of a Kohonen self-organizing training algorithm
Е. В. Хворостухина, Alexey Arlenovich L’vov, Sergey P. Ivzhenko · 2017
The self-organizing algorithm of Kohonen is well known for its ability to map an input space. That technique is named as Self-Organizing Map - SOM. A SOM can be trained in a short period of time with a few optimization techniques such as “winning” neurons search scope limit. In this paper we propose alternative options for improving the SOM learning speed. The basic idea of the proposed modification is based on the fact that learning is based on the method of a “winner” quick search until you reach so-called “breakthrough” epoch of learning. When a “breakthrough” epoch is reached it can be assumed that new “winner” neuron is always located at the place of the old one that allows the search to finish. This modification can significantly reduce the neuron-winners search time.