Technical Perspective: Why Don't Today's Deep Nets Overfit to Their Training Data?