Follow
https://nitter.hu/giannis_daras/status/1531693104821985280#m verdict: doesn't feel surprising to me. adversarial examples apparently have some structure, which is (some) evidence against the natural abstraction hypothesis. we will end up having to extrapolate concepts
and we knew already that adversarial examples are often likely features, not bugs: https://arxiv.org/abs/1905.02175