Improved techniques for deterministic l2 robustness

Training convolutional neural networks (CNNs) with a strict 1-Lipschitz constraint under the l_{2} norm is useful for adversarial robustness, interpretable gradients and stable training.