Data = Designdata + Testdata
Designdata = Trainingdata + Validationdata % Performance estimates are BIASED
If the data set is sufficiently large, a classifier should be chosen on the basis of the BIASED (but non-training) Validationdata error rate (%E). An un-biased evaluation can then be obtained on the basis of the UNBIASED non-design Testdata error rate. The uncertainty of error rate estimates for a sufficiently large data set is inversely proportional to it's size. If the data set is small, repeat the design multiple times so that the number of hidden nodes is minimized and each input vector has many chances to be in each of the three data subsets. It may also be wise to add a little random error to each vector.
A data set size of N = 6 does not provide enough information for a real world problem. You might want to use cross validation with the 6*5/2 =15 combinations of val/test data combinations and/or add noisy duplicates to improve the ability of the nets to perform on unseen data.
This is best done using a loop in a command line program.
Hope this helps.
Thank you for formally accepting my answer
Greg
