Adversarial Robustness is at Odds with Lazy Training

In this work, we extend this line of work to " lazy training " of neural networks - a dominant model in deep learning theory in which neural networks are provably efficiently learnable.