Is there a principled way to not overfit? Or is the best we can do not overtraining or reducing the precision of the calculations?