The SOM will always reflect the classification in terms of the neuron that gets fired. If we find a typical set of neurons with high frequency  that gets fired in an epoch is indication of no of groups. For ex. if neuron 2, 4, 5, 8 fires with high frequency then it is an indication of four groups. It is possible to separate the corresponding record or input in to a group looking in to the neuron that fired for that record.

In regression settings, representativity of model set is crucial for future prediction abilities of a model. In general, training set should incorporate all sources of expected variability. From a geometrical point of view it means that one should include into training set most diverse samples. This objective can be accomplished using uniform subset selection methods like, for instance, the Kenard and Stone algoritm. It belongs to a family of uniform subset selection metods that can assist in the selection of representstive model set scattered uniformly over experimental space.

It is also possible to design a representative test set with a similar priciple by a small modification of the Kenard and Stone algoritm (the so-called Duplex algorithm).

Cluster-based design assumes the selection of representatives from each cluster. In the context of SOM, from each neuron one could select a representative to model set.

For more details see for instance reference:

M. Daszykowski, B. Walczak, D.L. Massart, Representative subset selection, Analytica Chimica Acta, 468 (2002) 91-103

http://dx.doi.org/10.1016/S0003-2670(02)00651-7

@ Jayaram M.A

Thanks for your response. If I want to have three subsets of my input space what's the specific rule to this division. Does it need to have three statistical similar data subsets ?

@ Michal Daszykowski

Thanks Michal. I found your response and the article you sent interesting. But can you explain how I can implement the methods you've just mentioned?