Network Dissection: Quantifying Interpretability of Deep Visual Representations