Grant: Alliance Grants

The increasing integration of AI in our society emphasizes the importance of models that come with guarantees of fairness. However, there are challenges to achieving this goal, as state-of-the-art AI models are based on deep learning and it remains largely unclear how these models operate and how imbalances in the training data affect their fairness properties. Furthermore, AI models are increasingly being integrated into settings such as college admissions and credit lending, where decisions made now can affect the future, and it is unclear what the appropriate fairness metric should be, let alone how to achieve it. This project aims to address these challenges by contributing to the fundamentals of fair data-driven decision-making. First, we propose a new framework to understand how data imbalances affect generalization. This framework seeks to characterize the underlying geometric structures of the models learned by deep networks and how they change due to imbalances in the training data. Second, we develop a long-term fairness framework that seeks to determine the metrics to optimize in order to achieve long-term fairness as the data distribution evolves in dynamic learning settings. Third, we integrate this knowledge to develop principled deep learning algorithms that guarantee fair decisions in both static and dynamic environments. The overarching goal of this project is to promote confidence in AI and encourage ethical AI practices that prioritize equity for all, enabling it to tackle increasingly complex problems in a safe and responsible manner. From a technical standpoint, this program aims to advance our knowledge of modern machine learning, an area in which Canada aims to be a leader. Finally, the program will provide interdisciplinary problem-solving skills to train three PhD students (two at UBC and one at the collaborating US institution) for top research positions in academia and the growing-science industry.