This was one of the things Andrew Ng really hammers on in the coursera course. This alongside separating out the training set from the cross validation set for

This was one of the things Andrew Ng really hammers on in the coursera course. This alongside separating out the training set from the cross validation set for tuning parameters went a long way to dispel some of the "magic" in how you iterate towards a sensible model.