Deep neural networks provide impressive state-of-the-art performance while remaining hardly interpretable. Analysis of Linearized Layer Transformations (ALLT) aims at enhancing the network's interpretability by analysing the weights of the network. ALLT combines techniques from linearization, dimensionality reduction and feature visualisation to make network computations tractable. A detailed description of the method and results can be found here.
The schematic method
To install the package locally, navigate to the LinarizedLayerTransformations directory and run
pip install -e .
Initialise the remainder of the project skeleton:
mkdir data docs results tests
The intended contents of each of these directories is explained in the Good Research Code Handbook.
Now you should be able to run the scripts, for example:
python scripts/train_logical_network.py