Neural Network tool box

my pattern P size is 180*15 and target T size is 15*15.

   Create a network : mynet=newff(P,T,[80],{'logsig' 'logsig'},'traingdx'); 
  train it : mynet=train(mynet,P,T);

However I get different result at different time of executing the script. is it the nature of neural network???

Answer 1

[Greg Heath]

Default trn/val/tst data division and initial weights are random. To duplicate runs you have to initialize the RNG to the same initial state as before

 help rng
 doc rng

WARNING: YOUR PROBLEM IS VERY, VERY ILL-POSED

 [ I N ] = size(P )            % [ 180 15 ]
 [ O N ] = size(T)             % [ 15 15  ]
 Ntrn    = N -2*round(0.15*N)  % DEFAULT: 11 training (2 validation and 2 test) examples
 Ntrneq  = Ntrn*O              % 165 training equations
 H       = 80                  % 80 hidden nodes
 Nw      = (I+1)*H+(H+1)*O     % 181*80+81*15 = 15,695 unknown weights

1. You are trying to define an I= 180 dimensional space with only Ntrn = 11 training examples. 11 examples define at most a 10 dimensional space. Even if you used all 15 examples for training, you could only define, at most, a 14 dimensional space.

2. It is desirable to have MANY MORE training examples than twice the dimensionality of the input space. SUGGESTION: Get more examples and/or reduce the dimensionality of your input by projecting your inputs onto a smaller dimensional basis.

3. You are seeking accurate estimates of 15,695 unknown weights with only 165 training equations.

4. Ideally

 a. N >> I,O
 b. Ntrneq >> Nw  <==>  H << Hub = -1+ceil((Ntrneq-O)/(I+O+1))

If you explain in more detail P,I, T and O, perhaps we can give you better advice and relevant references.

Hope this helps.

Thank you for formally accepting my answer

Greg

[mamata]

I have a pattern recognition problem. I use 15 different character of font size 12. for each generate a matrix of 280 * 1. So for 15 number of characters, Training file contain matrix of P 280*15. The target is an identity matrix of 15*15. Try to test the network by taking same fonts of size 12, 16 and 24, so it became 280*45 (Ptest). Q1. Is my approach is correct Create a network : mynet=newff(P,T,[80],{'logsig' 'logsig'},'traingdx'); train it : mynet=train(mynet,P,T); test it :output=mynet(Ptest); Q2. How to generate the confusion matrix

Please replay

[Greg Heath]

Now you are trying to define a 280-dimensional space with 15 vectors.

Find a source on image feature extraction (e.g., search NEWSGROUP, ANSWERS and Google)

You are training with fontsize 12 and testing with fontsizes 16 and 24

Why use the obsolete newff instead of patternnet?

 help patternnet

Why use logsig and traingdx instead of defaults?

 help confusion
 doc confusion

Hope this helps.

Greg