YOUR USE OF THE WORD VALIDATION IS NOT STANDARD
In the field of neural networks the original data is traditionally divided
into three subsets: training, validation and testing
data = design + test
design = training + validation
data = training + validation + test
TRAINING SUBSET:
Explicitly used to change weights (includes bias weights) with most other
training parameters held constant. The resulting error estimate can be
highly biased if the number of training equations is not significantly
higher than the number of weights to be estimated.
VALIDATION SUBSET:
a. Explicitly used to stop training if the val subset error rate
continually increases for a fixed number (MATLAB default
= 6) of epochs. This event is called "VALIDATION STOPPING".
b. Sometimes it is useful to compare valstopped designs those
that are stopped otherwise.
c. Used to rank multiple trained nets. Since the val subset
is only used indirectly during weight estimation, the val error rate
estimate is considered to be "only weakly biased".
test subset:
Nondesign and Nontraining subset used to obtain relatively unbiased
error estimations.
It is not clear how you are using the term validation.
Please elucidate.
Greg
