Understanding Adversarial Training: Increasing Local Stability of Neural Nets through Robust Optimization