Self Organizing Maps

Hi,

I use the neural network toolbox of Matlab R2010a (not Kohonen somtoolbox). I use SOM to classify water radiances spectra. I cannot find in the plot tools how to represent each neuron with his reference vector (prototype spectrum), which is a statistical mean of all the spectra captured by the neuron.

Thank you for your help!

Answer 1

[Mark Hudson Beale]

The prototype pattern for each neuron is its weight vector. To see all the neurons' weight vectors:

net.IW

Each row represents the prototype vector for a different neuron.

These row/prototype vectors are what are shown in PLOTSOMPOS graphically, if the network has two or fewer inputs.

Answer 2

[Zied]

Thank you for your answer Mark. The problem is that i want to plot theses weight vectors inside each of their neuron (like the PLOTSOMHITS figure but instead of the sample number, i want the spectrum inside the neuron). When i type net.IW ,it gives me the weight matrix but i cannot find the appropriate command to plot it.

Thank your for your help again

Zièd

Answer 3

[Zied]

I solved the problem, thank you very much for your help Mark!

Answer 4

[Preetisha Kaur]

Hi there...I want to use SOM to classify a 124 X 26 data set into 3 clusters. I am struggling with the same. I want to know how to define the number of neurons for the layers in newsom command, also how can I make sure that the resultant clusters are 3 in numbers? Could you please help...

[Zied]

to define the number of neurons, you have to set the dimensions of your maps in the newsom command (d1 and d2):

net=newsom(input,[d1 d2],topology function, distance function, steps for neighborhood to shrink to 1, initial neighborhood size).

Also, you must not forget that the SOM in an unsepervised classification method. This means that the result of unsupervised learning is used to assign an observation to a subset, regardless of any notion of class. to get your clusters you have to do another classification after getting your results, like ascending hierarchical classification. That's also what i m trying to do right now.

Answer 5

[Zied]

Hi there, i want to plot all the input vectors (spectra) classified by each neuron (not only the prototype spectrum given by net.IW)

Could you please help?

[Zied]

The PLOTSOMHITS command guives you how much spectra each neuron had captured, i would like to plot these spectra for each neuron, please help!!