implement 2D Self-Organized-Maps output wrong answer in matlab

I have read the SOM algorithm form here http://www.ai-junkie.com/ann/som/som2.html and I tried using the matlab to implement it.

What I have encountered problem is that the cluster output for all nodes is the same.

It should output 9 cluster in the form of 9 distinct 2D point.

=====Input layer=====

And I use generate 9 type data from normal distribution means: form [1,1], [2,2] ~[9,9].

Wach type has 9 samples.

I use cell array which each element store one 2D point.

=====Output layer=====

I use the 3 X 3 map that will output 9 cluster.

I use cell array which each element store one 2D point.

=====Problem encountered=====

I have checked the algorithm for many times and cannot find the error.

Please point out the error, I cannot figure it out by myself and I have spent time for debugging about all two days.

Or you can give me some suggestion or advice to get a better approach, please.

Thank for your help in advance.

=====Code implementation=====

% ==========initialization part==========

    % data
    % how many cluster(output node)
    clusterSize = 9;

    % how many data(input number)
    dataSize = 81;

    % data dimension(input node)    
    dataDimension = sqrt(dataSize);

    % cluster dimension(map width or map length, I use square map)
    clusterDimension = sqrt(clusterSize);

    % parameter
    % neighbor radius 
    iniRadius = clusterDimension / 2;
    curRadius = iniRadius;

    % learning rate
    iniRate = 0.1;
    curRate = iniRate;

    % how many iteration time shoud do 
    curEpoch = 1;
    maxEpoch = 2000;

    % used to update radius by the formula
    timeConstant = 1000 / log(iniRadius);

    % stroe the distance between one input vector to all the output node
    distanceCluster = zeros(clusterDimension, clusterDimension);

    % store the distance between BMU and one outputNode
    distanceBMU = zeros(clusterDimension, clusterDimension);

    % used to update weight by the formula
    influence = 0;

    % used to display how many progress finished
    progress = 0;

    % generate data
    % generate data by normal distribution from [1,1] [2,2] .... [9.9]
    % data{1, j}, data{2,j} ........... are the same type if the j is the same
    data = cell(dataDimension, dataDimension);
    for i = 1:dataDimension
      for j = 1:dataDimension  
        data{i, j} = normrnd([j, j], 0.1, 1, 2);
      end
    end

    % random weight
    % using random number to initial weight
    cluster = cell(clusterDimension, clusterDimension);
    for i = 1:clusterDimension
      for j = 1:clusterDimension
        cluster{i, j} = rand(1,2); 
      end
    end

    % ==========training part==========
    for t = 1:maxEpoch
      % diplay progress
        % display the progress the program has finished
      if(mod(t, maxEpoch / 10) == 0)
        progress = progress + 10;
        disp(['progress has finished: ' num2str(progress) '% ......']);
      end

      for i = 1:dataDimension
        for j = 1:dataDimension
                % given one input vector
          tempData = cell2mat(data(i, j));

          % calculate distance between data and cluster
                % iterate over all the output node for one input vector
          for k = 1:clusterDimension
            for l = 1:clusterDimension
                        % given one output node
              tempCluster = cell2mat(cluster(k, l));

                        % calculate the distacne between one input vector to one output node
              distanceCluster(k, l) = sqrt(sum((tempData - tempCluster).^2));

            end
          end

          % BestMatchUnit(BMU)
                % get the min distance by sort the distance 
          [V I] = min(distanceCluster(:));

                % get the index of them 
          [I1,I2] = ind2sub(size(distanceCluster),I);

                % store the BMU vector
          BMU = cell2mat(cluster(I1, I2));

          % update weight
          % calculate distance between cluster and BMU
                % iterate over all the output node for update weight
          for k = 1:clusterDimension
            for l = 1:clusterDimension

                        % given one output node
              tempCluster = cell2mat(cluster(k, l));

                        % following are by the formula
                        % calculate the distance between BMU and one output node
              distanceBMU = sqrt(sum((BMU - tempCluster).^2));

                        % calculate the influence by the formula
              influence = exp((- distanceBMU * distanceBMU)   / (2 * curRadius * curRadius));

                        % update the weight
              cluster{k, l} = cluster{k, l} + influence * curRate * (tempData - cluster{k, l});

            end
          end
        end
      end

        % update the neighbor radius and learning rate
      curRadius = iniRadius * exp(-t / timeConstant);
      curRate = iniRate * exp(-t / 1000);
    end

[Greg Heath]

Do you think it is disrespectful to ask someone to try, without pay, to understand uncommented code that is that lengthy?

[LiangYu]

Sorry about that, I will add more comment.

Answer 1

[Image Analyst]

Follow this link and it will most likely lead you to the solution: http://blogs.mathworks.com/videos/2012/07/03/debugging-in-matlab/

[Greg Heath]

The program runs without error. What is the resulting answer and why is it wrong?