Can Active Learning Preemptively Mitigate Fairness Issues?

Code for the paper "Can Active Learning Preemptively Mitigate Fairness Issues?" presented at RAI 2021.

Arxiv link: https://arxiv.org/pdf/2104.06879.pdf

This repo aims at helping researchers reproduce our paper. We welcome questions and suggestions, please submit an issue! If you are working in active learning, our library BaaL might help you!

Glossary

• query_size : Number of data labelled per AL step
• AL Step: process of training, selecting samples, and labelling.
• learning_epoch: Number of epoch to train the model before uncertainty estimation.

Datasets

We experimented with many datasets and we have one format.

# For synbols datasets
with open(dataset_path, 'rb') as f:
    train_ds = pickle.load(f)
    test_ds = pickle.load(f)
    val_ds = pickle.load(f)


x,y = train_ds
print(x[0])
# A numpy array with shape [64,64,3]

print(y[0])
# A dictionnary with all the attributes and keys.
"""
{'char':'a',
 'color': 'r',
 ...
}
"""

Models

We use a VGG-16 and we do AL with MC-Dropout.

from active_fairness import utils
from baal.bayesian.dropout import patch_module

hyperparams = {} # Dictionnary for the hparams

model = utils.vgg16(pretrained=hyperparams['pretrained'],
                          num_classes=hyperparams['n_cls'])

# change dropout layer to be able to use MC-Dropout
model = patch_module(model)

How to launch

Here is an example to run our experiment on a particular dataset.

To generate the datasets, please look in misc/biased_dataset.ipynb.

poetry run python experiments/grad_experiment.py datasets/minority_dataset_50000.pkl color char