different result neural network classification

hi i have a dataset consist of 200 image , each 20 image printed with one printer (so classification is 10 class) extract 176 feature from each paper.

so for input i build a 176x200 matrix, 1 to 20th column (sample) belongs to 1th printer, 21 to 40th column belong to 2th printer and so on

i make a 10x200 target matrix , i set elements of 1-20 column form first row to "1" and other elements in first row to "0", for second row set 21-40 column to "1" and other to "0" and so on i use patternnet and plot confusion matrix but at every run, result is very Variable, 30 to 80% have any idea?

Answer 1

[Greg Heath]

Yes. The default trn/val/tst data-division and weight initializations are all random.

For reproducibility, initialize the RNG to the same state before TRAIN is called.

Over the years I have used

 rng(0) and rng(4151941) (a famous birth date).

Now I tend to use

 rng('default')

To search for a good design, I usually use a double loop search over number of hidden nodes and Ntrials random generator states per hidden node choice:

 rng('default')
 j=0
 for h = Hmin:dH:Hmax
     j=j+1
     net = fitnet(h);
     blah; blah; blah;

     for i = 1: Ntrials
        net = configure(net,x,t);
        blah; blah; blah;
        NMSE(i,j) = mse(t-y)/mean(var(t',1));
      end
end

Over the years I have posted zillions of examples in BOTH the NEWSGROUP and ANSWERS.

Hope this helps.

Thank you for formally accepting my answer

Greg

[milad ghanbari]

thank you Greg, that rng worked so target matrix is correct? considering i,m new to neural network , whats easiest way to calculate number of hidden layer? here is my code

inputs = input;
targets = target;

% Create a Pattern Recognition Network
hiddenLayerSize = 10;
net = patternnet(hiddenLayerSize);

% Setup Division of Data for Training, Validation, Testing
net.divideParam.trainRatio = 70/100;
net.divideParam.valRatio = 15/100;
net.divideParam.testRatio = 15/100;

% Train the Network
[net,tr] = train(net,inputs,targets);

% Test the Network
outputs = net(inputs);
errors = gsubtract(targets,outputs);
performance = perform(net,targets,outputs)

% View the Network
view(net)

% Plots
figure, plotconfusion(targets,outputs)

[Greg Heath]

I always use one hidden layer.

Since you have designed a classifier, you are not finished until you have determined the classification error rates for all of the classes in each of the trn/val/tst data divisions.