What SOM does sounds good and very useful on a paper (putting similar individuals close together, nonlinear version of PCA) for visualization and for dimensional reduction. Also there are whole chapters about it in ML books (e.g. Marsland), but I newer saw it being used in any serious way, also it was not merged in scikit-learn for the same reason.
If it sounds so good, why it is not? And if it so terrible, why is it being mentioned in ML courses, instead of more useful approaches?