Network Dissection

Tries to answer the question whether a Neural Network actually understood the concepts that can be visualized with Activation Maximization. So it quantifies the Interpretability of a Unit.

Disentangled Features

1. Get images with human-labeled visual concepts
2. Measure the CNN channel activations for them
3. Quantify the alignment of activations and labeled concepts

Alignment

Intersection over Union $Io{U}_{k,c}=\frac{\sum |M_k(x)\bigcap L_c(x)|}{\sum |M_k(x)\bigcup L_c(x)|}$