The Lipschitz Constant of Self-Attention

                  Hyunjik Kim 1 George Papamakarios 1 Andriy Mnih 1

              Abstract               constraint for neural networks, to control how much a net-
    Lipschitz constants of neural networks have been     work’s output can change relative to its input. Such Lips-
    explored in various contexts in deep learning,      chitz constraints are useful in several contexts. For example,
    such as provable adversarial robustness, estimat-    L ...