WHAT MORE IS NECESSARY FOR A GOOD NEURAL NET DESIGN?

    x = input; 
    t = target;

 1. [I  N  ] = size(x)    
    [O  N  ] = size(t) 

 2. The default(AUTOMATIC & RANDOM) design division 
    procedure is

   a. Train (70%)
   b. Validate (15%): occurs during training and stops 
      training if validation error increases for 
      6(default) continuous epochs.
   c. Test (15%)

 3. The default normalization (0 mean/unit variance) is sufficient.

 4. a. The (SUFFICIENT) default configuration only contains 
       ONE hidden layer  with H = 10 nodes and (I+1)*H+(H+1)*O 
       randomly initialized weights.
    b. If H is sufficiently large, ANY "reasonable" 
       input/output transformation can be approximated.
    c. However, if H is too large, the phenomenon of   
       overfitting occurs and special steps have to be taken.
    d. My approach is to find the minimum value of H that 
       yields an acceptable result
    e. Typically my training goal is

       mse(target-output) <= 0.01*var(target,1)

       NOTE: var(target,1) is the error for the naïve guess

             output = mean(target)

 5. Weights are automatically initiallized randomly. So, 
    typically, all you have to do is 
    a. Start using the MATLAB sample code with the default 
       H = 10
    b. Use a do loop to design a number of nets (e.g., 10 or 
       more) to find the best result from the random initial 
       weights 
    c. Search for the smallest value of H that will yield a 
       satisfactory solution.

 6. Any questions?