From: lynch.268@osu.edu   
      
   dhirendra wrote:   
   > Hi All,   
   >   
   >   I m Having a problem with NN training & testing for Face   
   > Recognition.   
   >   here i want to share with u my input & output pattern for NN & want   
   > suggestion of modifications...   
   >   
   >  I m having 100 images in which set of 10 images belong to a   
   > particular person,so 10 person are there.   
   >  If i process these 100 images with pca & give input to NN then input   
   > pattern goes columnwise-   
   >  1-1 1-2 1-3......after 10 images of first person...2-1 2-2 2-3.....   
   >  & output is like   
   > 1 1 1...after 10 values...2 2 2... likewise   
   >   
   > but the problem is that the NN is getting trained but testing fails.   
   >   
   > whts the solution of it?   
   > should i train Setwise(set of 10 images for each person) & save   
   > weights then test one by one with each save weights.   
   > plz help guys.   
   >   
   > Regards   
   > Dhirendra   
   >   
      
   So let  me get this straight. You have 100 inputs in the training set.   
   The inputs consist of 10 pictures for each of ten people. You are   
   inputing these ten images one at a time. First the 10 pictures of person   
   one, then the ten pictures of person two, etc.   
      
   Then when you test it the NN guests wrong, guessing that the first 10   
   images are persons one, etc.   
      
   What is the question being asked of the neural network? Is the question   
   "Is this a picture of person one?" or is the question "This a picture of   
   which person?" or is the question "Is this a picture of one of the ten   
   people you know?", different questions, different error rates. Some   
   questions are easier than others.   
      
   Second are you only feeding the NN each picture but once and always in   
   the same order? I have never tried to train a NN with such limited data.   
     We always use 10's of thousands of training input and ran the training   
   input through repeatedly until the error rate dropped to the cutoff point.   
      
   BTW You procedure should have 3 sets of data: Training data, Tuning data   
   and Test data. The tuning data is used to test you NN after training   
   while you are refining the process.   
      
   I would feed the training data through multiple times until the error   
   rate fell below 10 or 5%. I would also feed them in random order by   
   shuffling them at the start of each run.   
      
   tom   
      
   [ comp.ai is moderated ... your article may take a while to appear. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)