CSE5519 Advances in Computer Vision (Topic H: 2024: Safety, Robustness, and Evaluation of CV Models)

Efficient Bias Mitigation Without Privileged Information

TAB: Targeted Augmentation for Bias mitigation

Loss history embedding construction (use Helper model to generate loss history for training dataset)
Loss aware partitioning (partition the training dataset into groups based on the loss history, reweight the loss of each group to balance the dataset)
Group-balanced dataset generation (generate a new dataset by sampling from the groups based on the reweighting)
Robust model training (train the model on the new dataset)

Tip

This paper is a good example of how to mitigate bias in a dataset without using privileged information.

However, the mitigation is heavy relied on the loss history, which might be different for each model architecture. Thus, the produced dataset may not be generalizable to other models.

How to evaluate the bias mitigation effect across different models and different datasets?