A crucial pillar of research at CRCS is to advance societal fairness and quity. AI systems can perpetuate unfair outcomes and increase inequity. At CRCS, we design foundational AI systems to prevent the disparate impact of algorithms. Moreover, well-designed AI systems can help further societal equity and fairness. Designing fair and equitable AI systems complements CRCS's goals of enabling AI for enhanced public and clinical health and conservation interventions. To prevent disparate impact from unfair AI predictions, we augment AI-based predictions with safeguards that prevent disparate impact. For example, these can come in the form of an algorithmic recourse that can empower human users subjected to unfair decisions to act in ways that improve their outcomes. Importantly, we design fair and equitable AI systems using targeted interventions that advance conservation, public health, and clinical health goals. For example, we develop game-theoretic approaches to improve maternal health in low-resourced communities in South Asia. Our AI systems in clinical health can seamlessly incorporate human expertise to prevent unfair interventions. We design systems that carefully capture domain-specific (un)fairness to enable equitable AI-assisted decision-making. Finally, we are committed to creating new frameworks for equity throughout the AI development pipeline with a heavy emphasis on long-term societal good.
Algorithmic Recourse: Humans may be subject to algorithmic decision-making in many high-stakes decisions, like obtaining housing loans or qualifying for insurance subsidies. When facing unfavorable outcomes in these scenarios, it is helpful or necessary to provide them with avenues to improve their outcome, e.g., qualify for a loan after six months by repaying their existing loans. Algorithmic recourse is a means to provide human users with such a set of changes to help them improve their outcome, e.g., what minimal changes could help a bank client obtain a mortgage?). In practice, these algorithms can change over time, thus invalidating the recourse provided to the user. We provide a way to generate robust recourses so that individuals are guaranteed an improved outcome even under potential changes to the underlying algorithm used for decision-making.
Algorithms for Clinical Decision-Making Under Uncertainty: Algorithms could soon be used for treatment recommendations in clinical settings. Algorithmic recommendations may not always be perfect due to the underlying physiological uncertainty of a patient. In this case, we would like to design AI systems that can safeguard against such uncertainty and have human experts, such as clinicians, take over decision-making when relying on the AI system cannot improve patient outcomes. In this work, we design such safeguards using reinforcement learning for AI-augmented decision-making in clinical settings.
@HCRCS