Calibration in the Age of AI: From Prediction to Decision Making to AI Assisted Research (in person only)

I'll explain why (multi)calibration and its variants are interesting: it serves as an interface between prediction and decision making, and it can be used to mediate human/AI collaboration. I'll go into recent results we have characterizing the fundamental cost of calibration in both batch and online learning settings, establishing the optimal sample complexity and regret rates achievable. These last two papers were themselves written as.human/AI collaborations, and while describing the results I'll also talk about the process.