About Neural Network..... How to train Neural Network?

I am working on Pedestrian classification. I have data set having positive and negative images. I am using cohog as a feature extraction method. After Extracting features from both image sets I need to give them to Neural Network. Now here I have a confusion whether I have to mix the features of both pos and neg image set. Or train the NN individually. Kindly help me

Answer 1

[Greg Heath]

The I input features of the N = N0+N1 point class mixture are put in the same input matrix

 [ I N ] = size(input)

For c =2 classes it is sufficient to use a binary 0/1 target vector

 [ O N ] = size(target)  % O = c-1 = 1

For a pattern recognition net with I-H-O node topology, there are

 Ntrneq = Ntrn*O  % e.g., Ntrn = N - 2*round(0.15*N) ~ 0.7*N

training equations to design

 net = patternnet(H);

using

 [ net tr ] = train(net,input,target);

The resulting net has

 Nw = (I+1)*H+(H+1)*O

unknown weights. Although it is desirable that

 H < -1 + ceil((Ntrneq-O)/(I+O+1)

so that Nw does not exceed Ntrneq, it is not necessary if validation set stopping (or regularization using msereg) is used.

The resulting net outputs

 y = net(input);

are interpreted as estimates of the class 1 posterior probabilty conditional on the input. The corresponding assigned class indices and error rates are given by

 class = round(y);
 Nerr1 = sum(class ~= 1)
 Nerr0 = sum(class ~= 0)
 PCTerr1 = 100*Nerr1/N1
 PCTerr0 = 100*Nerr0/N0
 PCTerr = 100*(Nerr1+Nerr0)/(N1+N0)

Hope this helps.

Thank you for formally accepting my answer

Greg

Answer 2

[Mohan]

Use the train function after declaring your net as follows :

net=newff(minmax(in'),[20,1],{'tansig','purelin'},'train');

[net,tr]=train(net,trainInput',trainOutput');

where trainInput would be the "input vector" which has the extracted feature values.

trainOuput would be the corresponding output value - say a "1" for a positive image and a "0" for a negative image