Since (Not If) AIs Will Rule Us, Let's Make Them Nice To Us

It is inevitable that AIs will soon be making many life-altering decisions on us. Because of this, we need to ensure their behavior is guided by human values??"a challenge widely known as the AI Alignment problem. I'll discuss our work on fairness and explanation alignment, presenting both theoretical and experimental results developed with collaborators, particularly our medical center. Our research has focused on classic tasks like outlier detection, classification and clustering for images, as well as more recent work involving large language models. Underlying this research is a common fundamental set of questions of how to encode constraints and extract constraints into/from deep learning. I'll conclude by discussing the need to explore alignment beyond fairness.

Speaker: Ian Davidson, UC Davis