Deep k-Nearest Neighbors: Towards Confident, Interpretable and Robust Deep Learning