Fairness-aware Incremental Representation Learning (FaIRL)

We present the implementation of the AAAI 2023 paper :

Sustaining Fairness via Incremental Learning,
Somnath Basu Roy Chowdhury and Snigdha Chaturvedi,
UNC Chapel Hill

Dataset

We prepare the Biased MNIST dataset using src/mnist_data_create.py.

Data for the Biographies dataset is obtained from https://github.com/microsoft/biosbias.

To access the exact data in our experiments: Biased MNIST (is available here) and Biographies dataset (available here).

Setting up the environment

• Python version: python 3.8.5

• Dependencies: To install the dependencies using conda, please follow the steps below:

  conda create -n fairl python=3.8.5
  source activate fairl
  pip install -r requirements.txt 

Running Experiments

Running the main experiments on Biographies dataset.

cd src/fairl/
python bios.py \
        --device cuda:0  \
        --num_target_class 28 \
        --num_protected_class 2 \
        --exemplar_selection prototype

Running the main experiments in various configurations of the Biased MNIST dataset.

cd src/fairl/
python mnist.py \
        --device cuda:0  \
        --num_target_class 10 \
        --num_protected_class 10 \
        --dataset .8 \
        --exemplar_selection prototype

Reference

@article{fairl,
  title = {Sustaining Fairness via Incremental Learning},
  author = {Basu Roy Chowdhury, Somnath and Chaturvedi, Snigdha},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
  volume = {37},
  number = {1},
  year = {2023}
}