Impressive work! Their approach makes a ton of sense: "by not directly optimizing for the benchmark, we show that it becomes much more representative" and is e

Impressive work! Their approach makes a ton of sense: "by not directly optimizing for the benchmark, we show that it becomes much more representative" and is evident in the generalization and robustness: ~74% higher on the adversarial benchmark and similar/superior results on the standard Imagenet