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:
- Identify the greatest-explanatory layer and
- Influential units using numerical influence measures, then we
- Reconstruct relevant input regions responsible for activating these influential units.
Documents
Interesting readers can find more details in following documents:
Source code:
Results presented in Presentation can be found in:
Preliminary experiments of MIM on textual input can be found in:
If the file is too big for GitHub to render, you can try to view with nbviewer.