Self-organized patterns in the SOM network

This paper reports the discovery of certain self-organized patterns that develop automatically and unexpectedly during the training of a typical self-organizing map (SOM) network. These highly structured patterns emerge and evolve gradually from the random initial state as the training progresses. The web-like patterns are characterized by some line features at different scales, which tend to intersect at some common positions, and they form a highly organized hierarchical structure. The properties and variations of these patterns are affected by the parameters used in the training process. The specific mechanism of the formation of such self-organized patterns is still mostly unknown and currently under investigation. As a preliminary effort to understand the phenomenon, this paper also speculates and hypothesizes the possible mechanism of the phenomenon based on some qualitative and heuristic studies.