/ACE

Code for the paper, Neural Network Attributions: A Causal Perspective (ICML 2019).

Primary LanguageJupyter Notebook

ACE

Neural Network Attributions: A Causal Perspective
Aditya Chattopadhyay, Piyushi Manupriya, Anirban Sarkar, Vineeth N Balasubramanian
Presented at ICML 2019

Dependencies:
scikit-learn (0.19.1)
scipy (0.17.0)
torch (0.4.0)
joblib (0.11)
matplotlib (1.5.1)
numpy (1.14.5)

Usage-
MNIST:
  sh run_mnist_mod.sh
  MNIST.ipynb
  
Iris:
  python decision_tree.py
  train.ipynb
  ACE.ipynb
  
Synthetic Dataset:
  toy_dataset.ipynb
  python evaluate_lstm.py
  
Aircraft:
  python lstm.py
  python find_tau.py
  python aircraft_causal_interventions.py foldername
  eg. python aircraft_causal_interventions.py "40"
  python learn_causal_regressors.py learn effect_num_header effect
  eg. python learn_causal_regressors.py learn 5 LATG
  python causal_analysis_final.py predict effect foldername start_time
  eg. python causal_analysis_final.py predict GS "40" 100

NASA dataset used in Aircraft code is uploaded at https://drive.google.com/open?id=1rEZ3veRpcKH5OZKAoXuVTyC9oMnn78ra
Class-conditional Beta VAE code used in MNIST experiments is adapted from Beta VAE code from https://github.com/1Konny/Beta-VAE

If you use this code, please cite our paper:
Aditya Chattopadhyay, Piyushi Manupriya, Anirban Sarkar, Vineeth N Balasubramanian. "Neural Network Attributions: A Causal Perspective", in International Conference on Machine Learning (ICML), 2019.
Bibtex
and consider giving a star to our repository.

References:
https://github.com/1Konny/Beta-VAE
https://c3.nasa.gov/dashlink/projects/85/resources/?type=ds