Kohonen's self-organizing maps with a number of distinguishing features: (1) A very efficient, single threaded, stochastic training algorithm based on ideas from tensor algebra. Up to 60x faster than traditional single-threaded training algorithms. No special accelerator hardware required. (2) Automatic centroid detection and visualization using starbursts. (3) Two models of the data: (a) a self-organizing map model, (b) a centroid based clustering model. (4) A number of easily accessible quality metrics for the self-organizing map and the centroid based cluster model.
R package for self-organizing maps contains state of the art learning algorithms, visualizations, and evaluation functions.