A replication of the paper "Inducing Causal Structure for Interpretable Neural Networks" by Geiger et al.
You can use the Dockerfile to set it up. Alternatively, you can use poetry install from the root folder.
There is only one task you can run this on for now.
Running python train.py trains a model using Intercahnge Interventions + Multi Task (1)
Running python eval.py generates plots for how accurate the circuit induced is.