Neural Network sim(net, input) gives crazy results

Once I have trained my network I use the sim(net, input) function to get the results The results are incredible, and they are diffrent than the result obtained by manual matrix calculation using net.IW{1} etc. I obtain result around 10^4 with the sim, whereas result are around 1 with the matrix calculation !

Here is a zip file of the workspace and the code that provide what I am talking about

http://dl.dropbox.com/u/26971639/MatlabBug.zip

I also c opy paste the code here in case you want have a quick look

%   in the workspace 'imp' is the data imput (7 variables for each input) and
  % 'targ' is the data target
  sz = size (imp);
  % numbers of the coloumns for separation between train set, validation set, test set.
  d1=round(sz(2)/2); % the half of the dataset
  d2=round((sz(2)-d1)/2)+d1; % the half of the remaining part i.e. the quart
  d3=sz(2);  % the last quart
  % network with 1 input layer (size of imput is 7) , one hidden layer of 5
  % neurons, and one unique neurone in the output layer
  net = newff(imp,targ,5); 
  % actual separation for train test and validation set
  imp1=imp(:,1:d1); % imput for training
  targ1=targ(:,1:d1); % target for training
  VV.P=imp(:,d1+1:d2); % validation set
  VV.T=targ(:,d1+1:d2);
  VT.P=imp(:,d2+1:d3); % test set
  VT.T=targ(:,d2+1:d3);
  net.inputweights{1,1}.initfcn = 'rands';
  net.layers{1}.transferFcn = 'tansig';
  net.layers{2}.transferFcn = 'purelin';
  net = init(net);
  train(net,imp1,targ1,[],[],VV,VT);
  % simulation on the full dataset
  y1 = sim(net,imp);
  % bias of layers addapted for direct calculation with the size of the dataset 
  B1 = net.b{1}*ones(1,size(imp,2)); % all the coloumns are identical, and equal to net.b{1}
  B2 = net.b{2}*ones(1,size(imp,2));
  OutLayer1 = tansig(net.IW{1}*imp+B1); % output from the layer 1 (the hidden layer)
  OutLayer2 = purelin(net.LW{2}*OutLayer1+B2); % output from the layer 2 which is the output layer
  y2 = OutLayer2; % just to give an easy name
  % now you can compare y1 and y2
  plot(1:d3,y1,'o',1:d3,y2,'x');
  % NOTE THE *10^4 in the Y axis

Answer 1

[Lucas García]

As surprising as it can be, the output of the network is correct. So, how come such huge differences from your correct math interpretation of the network?

Well, the network normalizes the input data before processing it through the network and then transforms the output back. This is done by the network with the function mapmimax. You can find it in your network and if you make the network not to use them, then you will obtain the same results as your math:

net.inputs{1}.processFcns 
net.outputs{2}.processFcns

However, I don't recommend this. It is a good idea to normalize your data before you present it to the network or your weights could get too big.

In order to follow the math of the network, you can do the following:

imp2 = mapminmax('apply',imp,net.inputs{1}.processSettings{3});
OutLayer1 = tansig(net.IW{1}*imp2+B1); 
OutLayer2 = purelin(net.LW{2}*OutLayer1+B2); 
y2 = mapminmax('reverse',OutLayer2,net.outputs{2}.processSettings{2});

Now your plot of y1 and y2 should be the same.

[Flo Trentini]

Oh my god you're a life saver, it works ! I knew there was somthg like this hidden that I didn't understand and you found it. I'm really thankful for this answer because I did a research on the net and I didn't find anything

I have to say that I still don't understand how come the result come out so bad and wrong with the process setting, I mean If i took the MEAN (the mean of the target) as my forecast, I would do better mse (the performance function is mse) than the network !

Answer 2

[Elaheh]

Dear Locus,

I have done the same normalization as you said. The results of my sim function for "training dataset" is the same as MATLAB sim function. However, for "test dataset" my results is totally different than MATLAB sim function! I know, it is so strange! Could you please help me?

Thanks, Elaheh