Semantic Bottlenecks: Quantifying and Improving Inspectability of Deep Representations - International Journal of Computer Vision