Learning to create fairer AI systems for Computer Vision

The contemporary surge in artificial intelligence (AI) adoption brings automation to various daily tasks. However, biases in AI models, particularly concerning legally or morally protected attributes like ethnicity, age, or gender, pose substantial challenges. The intricacies of computer vision, dealing with high-dimensional inputs, further complicate the issue. Existing solutions have trade-offs, such as computational overhead or limited generalizability to fairness measures. Addressing attributes along a continuous spectrum, like ethnicity, challenges prevalent assumptions of a finite set of distinct classes. This thesis investigates modern fairness algorithms for such attributes, proposing a novel differential approach to enhance fairness, introduces open-set fairness methods, and explores real-world biases in protein crystallization data.