[T16 · Regularization] Spectral norm regularization (SN-GAN) constrains the spectral norm (largest singular value) of weight matrices in neural networks. Contrast this with L2 regularization and explain why spectral norm is important for training Generative Adversarial Networks.