NN trn/val/tst data division terminology:
data = design + test
design = train + validation
train:
1. Estimate weight and bias values for multiple designs
2. Estimate biased and degree-of-freedom-adjusted training performances
validation:
1. Stops training when validation error minimizes
2. Used to rank designs
a. Weed out failed designs as outliers
b. Obtain resulting summary performance stats(min/median/mean/std/max)
test:
1. Uses the "best" design ( e.g.,min(mseval) ) to obtain an UNBIASED estimate of unseen data (generalization).
2. Uses all acceptable designs to estimate UNBIASED summary statistics of performance on unseen data
