image classifier using neural network

I've implemented a neural network for the pattern recognition. From the two classes of images(I've 70 samples for each class) I've used SIFT feature + BOvW to make the image descriptor (number of clusters is 10). After training and validation I've got confusion matrix with accuracy 80% overall. here I am having two doubts (I've done it in MATLAB GUI)

1)When ever I am training the network it gives different values (confusion matrix) in that case how can I validate the network? (MATLAB documentation says its due to different initial conditions how can I fix it for the good result?).Different values like when ever I've finished training+validation (MATLAB) confusion matrix gives overall accuracy in the range 50%-80%(several times). But mostly more than 70%.

2)How can I use this network to tell the category of the image if I input one, after training and validation?

Answer 1

[Greg Heath]

I assume you are using patternnet.

Initial weights and data division depend on the current state of the RNG which changes each time it is used. Therefore, intitialize the RNG before the outside loop over hidden nodes and store the state everytime a new net is created.

For example

 rng(4151941)
 j=0
 for h = Hmin:dH:Hmax
 j=j+1

 .. define net and parameters

     for i=1:Ntrials
        state(i,j) = rng; 
        net = configure(net,x,t);
        ...

     end
  end

I have posted zillions of examples. Search in NEWSGROUP and ANSWERS using

 greg  Hmin:dH:Hmax  Ntrials

Hope this helps.

Thank you for formally accepting my answer

Greg