MATLAB: I am using a trainbr function to train a NN on 120X134 matrix. It trains very well even with three nodes in hidden layer but does not give satisfactory results on a test matrix. If this is a case of overfitting, how should I avoid it

MATLABneural network

The network is being used to classify into 4 classes

Reduce the number of inputs.

 help plsregress
 doc  plsregress
 type plsregress

Hope this helps.

Greg

The TRAINBR code covering indexing and ratios is BUGGY. I couldn't figure it out when I looked at it back in June 2013.

Now that I'm older, I might have better luck.

1. Complain to MATLAB, directly, with specific, repeatable, examples on MATLAB data

 help nndata
 doc nndata

2. Experiment some more to see if you can impose something other than the 80/0/20 default.

3. Another possibility is to use TRAIN with msereg

Hope this helps.

Thank you for formally accepting my answer

Greg

P.S. Please report any findings

Backpropagation is an optimization technique used in many types of algorithms.

trainlm and trainbr are just two of many algorithms that use BP. However both of these use the BP technique with LM.

At least ten others that use BP are listed in the MATLAB documentation.

Hope this helps.

Thank you for formally accepting my answer

Greg