How to troubleshoot my SOM

Hi,

I've have a training set of 10 different instances. Each of the instances has 245 feature values.

I know that the feature set is of good quality as I trained a Support Vector Machine with this data and got perfect classification results.

But if I train a SOM with the following code it seems that the correct winner can never be determined:


this.input = new double[10][245];
..... filling the array .....

SOMPattern somPattern = new SOMPattern();
somPattern.setInputNeurons(245);
somPattern.setOutputNeurons(10);
this.network = somPattern.generate();
this.network.reset();

NeuralDataSet trainingSet = new BasicNeuralDataSet(this.input, null);
CompetitiveTraining train = new CompetitiveTraining(this.network, 0.25, trainingSet,
new NeighborhoodGaussian( new GaussianFunction(0, 1, 2)));

train.setForceWinner(true);

for(int epoch = 0; epoch < 100; epoch++){
train.iteration();
logger.info("Epoch #" + epoch + " Error:" + Format.formatPercent(train.getError()));
}

this.network = train.getNetwork();

During the iterations I can see that the lowest Errorrate is:
Epoch #11 Error:327,434577%

I've tried within a range from 50 to 5000 iterations.

However, even classifying a training instance for evaluation with the code


NeuralData neuralInput = new BasicNeuralData(features);
final int output = this.network.winner(neuralInput);


never retrieved the correct winner. It seems that my SOM is unable to learn the input data correctly.

So, where do I start to troubleshoot? Is the code correct? Is the error rate way to high?

BTW: If I try to generate a BasicNetwork with 245 - 245 - 1 the error rates are also noted between 1.300% and 120% and the network also never determines the correct output value.

André