FLAC: Fairness-Aware Representation Learning by Suppressing Attribute-Class Associations.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence Ethics

Background:

• Bias in computer vision systems can perpetuate societal discrimination.
• Existing fairness methods often require protected attribute labels, limiting their practical use.
• Label-unaware fairness approaches typically yield lower performance.

Purpose of the Study:

• To introduce FLAC (Fairness through Learning Attribute Correlations), a novel methodology for training fair computer vision models.
• To develop a label-unaware approach that minimizes bias without relying on protected attribute labels.
• To improve the fairness and accuracy of computer vision models on diverse datasets.

Main Methods:

• FLAC minimizes mutual information between model features and protected attributes without using attribute labels.
• It employs a sampling strategy to emphasize underrepresented data samples.
• The method frames fairness as a probability matching problem using a bias-capturing classifier.

Main Results:

• FLAC achieves state-of-the-art performance on Biased-MNIST, CelebA, and UTKFace, surpassing existing methods by significant margins (29.1%, 18.1%, 21.9%).
• The approach demonstrates improved accuracy on ImageNet-A (2.2%) and Corrupted-Cifar10 (up to 4.2%).
• FLAC often outperforms even bias label-aware state-of-the-art methods in experimental evaluations.

Conclusions:

• FLAC effectively learns fair representations that are independent of protected attributes, even without explicit labels.
• The methodology offers a practical and high-performing solution for mitigating bias in computer vision.
• FLAC represents a significant advancement in developing equitable AI systems.