Update: Perl AI::NeuralNet::SOM
The previous version was quite unsmart in the way it would select the next feature vector to test against the current SOM. Worse, under adverse circumstances it would ignore some vectors. So it would converge, but much slower than actually possible. If the inbound vector dimension is small or the SOM itself is not big, this is not a big issue, but it becomes so once one (or both) is bigger.
There is quite some room for improvement, maybe I am forced to invest more time soon.