Department Seminar: Information Elicitation – Past, Present, and Future by Bo Waggoner

This talk will give a whirlwind survey of the field of information elicitation — designing reward functions for predictions — and its applications from 1950 to 2050-ish. The mathematical foundation is in proper scoring rules (a.k.a. proper loss functions), which we’ll see have deep roots in convex analysis. In a game theory context, we’ll see how proper scoring rules underly prediction market models, which can be viewed as convex optimization algorithms. In a machine learning context, we’ll see how theory of scoring rules can inform design of loss functions and valuation of data. We’ll look ahead to open directions and potential future applications as well. Based on several joint works.