Sparse Representations Improve Adversarial Robustness of Neural Network Classifiers