Layered Explanations Framework

Layered Explanations Framework aims to exploit the intrinsic architecture of neural networks, namely hidden units, to generate an interpretable explanation of the algorithmic decision. It consists of 3 main steps:

1. Identify the greatest-explanatory layer and
2. Influential units using numerical influence measures, then we
3. Reconstruct relevant input regions responsible for activating these influential units.

Documents

Interesting readers can find more details in following documents:

• Workshop abstract paper
• Presentation

Source code:

Results presented in Presentation can be found in:

• LayeredExplanationsFramework.ipynb
• DogFish_VGG16.ipynb

Preliminary experiments of MIM on textual input can be found in:

• 20newsgroup_MIM.ipynb
• IMDB_MIM.ipynb

If the file is too big for GitHub to render, you can try to view with nbviewer.