Neural nets, how to access wrongly classified instances?

Common scenario --> Use the same data set to train a feedforward 3-layer neural network with backprop for classification purposes - keep all things equal (apart from initial randomised weights) only altering number of neurons in the single hidden layer. The trial-and-error search for optimal number of neurons in the hidden layer.

Hypothetical Network A with 6 neurons in hidden layer achieves 95.6% correctness on the test data Hypothetical Network B with 17 neurons in hidden layer achieves 97.3% correct.

I would like to know which specific instances Network A misclassified and which were misclassified by Network B.

I am a newcomer to Matlab, but expect it is possible to access this data.

My question is how might I go about finding out which instances were wrongly classified?

Thanks, Amanda

Answer 1

[Greg Heath]

Although your net probably has three layers of nodes, it is a 2-layer network because it only has 2 layers of weights....I know.

You didn't specify how you designed your c-class classifier.

If you used the standard method of using columns of eye(c) in the target matrix:

 trueclass = vec2ind(target)
 target = ind2vec(trueclass)
 .....
 output = net(input);
 assignedclass = vec2ind(output);

errors = (assignedclass ~= truclass)

Hope this helps.

Thank you for formally accepting my answer

Greg