TARA: Training and Representation Alteration for AI Fairness and Domain Generalization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• AI models can exhibit bias due to sensitive attributes in data.
• Existing debiasing methods have limitations in addressing complex bias sources.
• Fairness metrics often fail to capture the trade-offs in debiasing.

Purpose of the Study:

• To propose a novel dual strategy for enforcing AI fairness.
• To mitigate AI bias by addressing both data representation and training set imbalances.
• To introduce new metrics for evaluating debiasing performance and Pareto efficiency.

Main Methods:

• Training and Representation Alteration (TARA) employs adversarial independence for representation learning.
• TARA uses intelligent augmentation with generative models for training set alteration.
• Novel conjunctive debiasing metrics are proposed to assess Pareto efficiency.

Main Results:

• TARA significantly reduces or eliminates bias in baseline models.
• Experiments on Eye-PACS and CelebA datasets show improved accuracy and reduced accuracy gaps.
• TARA outperforms competing debiasing methods with equivalent information.

Conclusions:

• TARA is an effective method for mitigating AI bias in image analytics.
• The proposed conjunctive debiasing metrics provide a more robust evaluation of debiasing performance.
• The TARA method demonstrates Pareto efficiency in fairness-accuracy trade-offs.