Is is possible that although neural nets can overfit as this paper shows, practitioners just stop training early before this happens? And/or they use a validation set? Would that be enough to explain the good generalization despite the huge number of parameters