How can I ameliorate the Neural network implementation?

I used these code lines to built a Neural Network classifier in a face recognition project:

 net = patternnet(10);

 [net,tr] = train(net,feaVectors,labels); % feaVectors=4800*90 and labels=15*90

testY = net(mat_test); % mat_test=4800*75

[c,cm] = confusion(labels_test,testY);

The problem is that I don't get a satisfying result, the true recognition rate is around 50% or less, and that's too low. I don't get why because I think the implementation is correct !! In fact, I use the same inputs with SVM classifier and I get very satisfying results. can you help me in this ?

Answer 1

[Greg Heath]

1. Try not to have have more unknown weights than training equations (OVERFITTING)because there is no guarantee that you will get good performance on nontraining data.
2. Rank multiple designs by the performance on validation data
3. Obtain unbiased performance estiamtes using the performance on test data.
4. It is preferrable to have more training equations than unknown weights. Therefore it is desirable to use as few hidden nodes as possible.
5. There is a default 0.7/0.15/0.15 RANDOM Ntrn/Nval/Ntst data division
6. The default initial weights are RANDOM
7. Given a sufficient number of hidden nodes, you may have to train 10 or 20 nets to obtain one that yields a good set of initial weights.

 [ I N ] = size(input)   % [ 4800 90 ]
 [ O N ] = size(target)  % [ 15 90 ]
 Ntst = round(0.15*N)    % 14
 Nval = Ntst             % 14
 Ntrn = N - Nval-Ntst    % 62
 Ntrneq = Ntrn*O         % 930  training equations
 H = 10                  % default # of training equations
 Nw = (I+1)*H+(H+1)*O = 48175  % Unknown weights

You have more than 50 times as many weights than you have equations (SEVERE OVERFITTING). You should try to use image feature extraction to drastically reduce the number of inputs.

Also, use the training record tr to separate training, validation and testing performance.

Hope this helps.

Thank you for formally accepting my answer

Greg

PS: I have zillions of patternnet posts in the NEWSGROUP and ANSWERS. For initial searches try

 patternnet tutorial

and

 patternnet greg

[Sa rah]

thank you for your answer, but can you please tell me what algorithm do you suggest to use for feature extraction so that the inputs have less dimensions? Do you agree on PCA for example?

[Greg Heath]

In general, PLS is better for classification. However, not sure of the relative difficulty in using both for your problem.