Tweet by Liv:

17/ If this is right, it reframes the robustness problem: adversarial vulnerability might be the price of neural networks' incredible efficiency. They're vulnerable precisely because they're doing something clever with their representations.