I have some training data and some test data. Can anyone tell me is training data is the input data or it is desired data in wavelet neural network?

I have EEG signals obtained from MIT_BIH. I want to use some of it for training data for seizure detection and some for testing purposes. I have the following code for training in wavelet neural network. What is desired output and input for my case? clear all; clc; close all; %initiate of data P=3 %number of sample m=1%number of input node n=20%number of hidden node N=1%number of ouptut node % %a(n) b(n) scale and shifting parameter matrix %x(P,m) input matrix of P sample %net(P,n) ouput of hidden node %y(P,N) output of network %d(P,N) ideal output of network % phi(P,n) ouput of hidden node wavelet funciton %W(N,n)weight value between ouput and hidden %WW(n,m) weight value between hidden and input node x=[4;5;6] d=[1.3;3.6;6.7] W=rand(N,n) WW=rand(n,m) a=ones(1,n) for j=1:n b(j)=j*P/n; end %%%%%%%%%%%%%%%%%% %EW(N,n) gradient of W %EWW(n,m) gradient of WW %Ea(n) gradient of a %Eb(n) gradient of b %%%%%%%%%%%%%%] epoch=1; epo=100;%???Purpose of it??? error=0.05; err=0.01; delta =1; lin=0.5; while (error>=err && epoch<=epo)

     u=0;%u is the middle variant
     %caculation of net input 
     for p=1:P
         for j=1:n
             u=0;
              for k=1:m
                  u=u+WW(j,k)*x(p,k);%%%probably vj(n)
              end
              net(p,j)=u;

         end
     end
     %calculation of morlet 0r mexican wavelet output
          for p=1:P
              for j=1:n
                  u=net(p,j);
                  u=(u-b(j))/a(j);
                  phi(p,j)=cos(1.75*u)*exp(-u*u/2); %morlet wavelet
                  %phi(p,j)=(1-u^2)*exp(-u*u/2); %mexican hat wavelet
              end
          end
      %calculation of output of network

         for p=1:P
             for i=1:N
                 u=0;
                 for j=1:n
                     u=u+W(i,j)*phi(p,j);
                 end
                 y(p,i)=delta*abs(u);
             end
         end
       %calculation of error of output
       u=0;
     for p=1:P
         for i=1:N
            %u=u+abs(d(p,i)*log(y(p,i))+(1-d(p,i)*log(1-y(p,i))));
            u=u+(d(p,i)-y(p,i))^2;
         end
     end
     %u=u/2
     error=u;
     %calculate of gradient of network
     for i=1:N
         for j=1:n
             u=0;
             for p=1:P
                 u=u+(d(p,i)-y(p,i))*phi(p,j);
             end
             EW(i,j)=u;
             %EW(i,j)=-u;%the resule would be wrong
         end
     end
     for j=1:n
         for k=1:m
             u=0;
             for p=1:P
                 for i=1:N
                    u=u+(d(p,i)-y(p,i))*W(i,j)*phi(p,j)*x(p,k)/a(j) ;
                 end
             end
             EWW(j,k)=u;
             %EWW(j,k)=u the result would be wrong
         end 
     end 
     for j=1:n
             u=0;
             for p=1:P
                 for i=1:N
                    u=u+(d(p,i)-y(p,i))*W(i,j)*phi(p,j)/a(j) ;
                 end
             end
             Eb(j)=u;
     end 
     for j=1:n
             u=0;
             for p=1:P
                 for i=1:N
                    u=u+(d(p,i)-y(p,i))*W(i,j)*phi(p,j)*((net(p,j)-b(j))/b(j))/a(j) ;
                 end
             end
             Ea(j)=u;
     end 
    %adjust of weight value
    WW=WW-lin*EWW;
    W=W-lin*EW;
    a=a-lin*Ea;
    b=b-lin*Eb;
     %number of epoch increase by 1
     epoch=epoch+1;

end plot(x,d,x,y,'--')

Answer 1

[Greg Heath]

 Your description should be reformatted to prevent confusion.

 data   = design + test
 design = training + validation 
 data   = training + validation + test

EACH OF THESE 3 SUBSETS CONSIST OF 2 PARTS

DATASUBSET = INPUT + TARGET

 default MATLAB division ratios = 0.7/0.15/0.15

 HOWEVER, YOU CAN MODIFY THESE RATIOS

 nontraining = validation + test

 TRAINING:    used to calculate weights and biases
 VALIDATION:  used DURING training to help make sure net works 
              well on nontraining data. NOT directly involved in 
              calculating weight values
 TEST:        used to obtain UNBIASED estimate of performance on 
              ALL (i.e., TEST + UNKNOWN) data not involved in 
              training or validation

Hope this helps.

Thank you for formally accepting my answer

Greg

[Naznin Sultana]

Thanks for answering. however i am looking at a WNN training code where input x and output y (called desired value) was used. My question was if i have EEG features as training set, what i will set as desired output?

Answer 2

[Greg Heath]

Correct notation:

 x input
 t target = desired output
 y output = net(x)
 e error  = t - y

% Reference output and MeanSquareError

 tref     = mean(t',2)   
 MSEref   = mean((t-tref).^2)
          = mean(var(t'),1) 

 [ net tr y e ] = train(net,x,t);
% y = net(x); e = t-y;
 MSE     = mse(e)
 NMSE    = MSE/NMSEref
 Rsquare = 1 - NMSE   % See Wikipedia

% For training details

 tr = tr

Hope this helps

Thank you for formally accepting my answer

Greg