Change trainingmode of Self-Organizing Maps

I want to sort rotordata, which can be shifted, with a SOM. Till now I shifted the data by using some specialpoints in the dataset, but this is not perfect. My solution is to change the SOM-algorithm, so it will not only check all euklidian distances of the outputneurons, but also trys all rotations of the weightvectors.

Where can I change the code? I cant find it in trainbu Thank you!

Answer 1

[Nick Hobbs]

In the documentation for trainbu, under the heading 'Network Use', you will see you can assign a learning function to a network using 'NET.inputWeights{i,j}.learnFcn'. If you do this for each weight, you will have effectively changed how the network learns. If we look at the learning function for a standard self-organizing map

net = selforgmap([8 8]);
>> net.inputWeights{1,1}

ans =

    Neural Network Weight

            delays: 0
           initFcn: 'initsompc'
      initSettings: .inputSize, .sampleSize, .numElements
             learn: true
          learnFcn: 'learnsomb'
        learnParam: .init_neighborhood, .steps
              size: [64 0]
         weightFcn: 'negdist'
       weightParam: (none)
          userdata: (your custom info)

From this output, 'learnFcn' is 'learnsomb'. You cannot edit this file, however you can use it as a template for your own learning function. If you create a new function 'customLearningFunction' that meets the requirements for a learning function and is in your path, then you can set each weight in your network to 'customLearningFunction' and use this instead of the default learning function.