There is no hard and fast answer to this question. Given N nodes and M data points, if NM, that is, you have an emergent self-organizing map, each data point will have its own neighborhood, and the clustering structure can be read out from the topology of the map.

In either case, you can run a clustering algorithm on the weight vectors to get a class assignment.

Thank you, it really helps. @Peter Wittek

Hello,

we are working on it too. We decided we have to select a node number by ecological sense. We are triying to reduce to the max the number of nodes, in order to simplify our datasets, but we have to take care and not to reduce too much, so we can lose out data variability.

regards

According to Jing Tian et al. (find paper attached) you can use the following equation:

M ≈ 5 sqrt(N)

where M is the number of neurons, which is an integer close to the result of the right hand side of the equation, and N is the number of observations.

Note, that M is the total number of neurons. If your map is quadratic, then use the square root of M to obtain the number of neurons in one row.

Hello,

There are two ways to determine the number of nodes needed, the first one is to take 

M ≈5√N (as said by Alexander Maier) or take the number of classes as the number of nodes see paper  attached

you can try structure adapting (growing) SOMs