Debiasing Global Workspace: A Cognitive Neural Framework for Learning Debiased and Interpretable Representations

This is the official implementation of Debiasing Global Workspace (DGW).

Requirments

conda create --name py38DGW python=3.8
conda activate py38DGW
pip install -r requirements.txt

Datasets

Please check the repo of Learning Debiased Represntations via Disentangled Feature Augmentation (LFA). You can download all dataests via the link.

Usage

• For [vanilla, lfa, dgw], check script files in the folder scripts to execute models. You can execute [vanilla, lfa, dgw].
• For [ReBias, LfF], we re-implemented them based on their repos. Please check dev branch in our repo to see our implementations of the two baselines.

Our Pretrained Models

You can download our pretrained models of DGW via the following links to check their test accuracies in our paper.

• Colored MNIST
• Corrupted CIFAR-10
• BFFHQ

Acknowledgement

Our source codes are based on:

• Concept-Centric Transformers [WACV2023]: We implemented our method referring to this repo and used this as one of our baselines.
• Learning Debiased Represntations via Disentangled Feature Augmentation (LFA) [NeurIPS2021]: We implemented our method based on this repo and compare our model with this as one of our baselines.
• Learning De-biased Representations with Biased Representations (ReBias) [ICML2020]: This is one of our baselines.
• Learning from Failure: De-biasing Classifier from Biased Classifier (LfF) [NeurIPS2020]: This is one of our baselines.