Error Correcting Output Codes Improve Probability Estimation and Adversarial Robustness of Deep Neural Networks

From a scientific perspective, the existence of adversarial examples demonstrates that machine learning models that achieve superhuman performance on benign, "naturally occurring"