How much do self organising maps suffer from local minima problems? I assume that the original orientation of the grid can have major orientation on the resulting shape of the grid, but does this have large impacts on the fit (e.g. mean square difference from data to assigned grid points), or are the differences fairly small?


If your data is high-dimensional, there will be many local minima, and the optimization problem is almost combinatorial. Then it is likely, but not guaranteed, that you will have a large number of optima that are almost as good as the global optimum—which you will never find.

That is, there is no simple answer. It definitely is prone to local minima. Try different initializations and choose the result with the smallest MSE.

  • $\begingroup$ As an extension to the problem, do you think that a k-dimensional grid in a k-dimensional data space would also suffer from local minima problems? $\endgroup$ – naught101 Dec 23 '13 at 23:12
  • $\begingroup$ It depends on the effective dimensionality of the data and on its distribution. There is still the K-means quantisation, which can always cause local minima, but of course they are in some sense less pathological. $\endgroup$ – scellus Dec 24 '13 at 12:46
  • $\begingroup$ what do you mean by pathological? $\endgroup$ – naught101 Jan 7 '14 at 3:42
  • $\begingroup$ I mean the overall topography on the map may still be similar even if you have K-means local minima. $\endgroup$ – scellus Jan 7 '14 at 8:22

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.