Proposing a differentiable regularizer for deep classifiers to increase input margin and developing a scalable method for Lipschitz constant estimation.