When
Where
Speaker: Jake Callahan, Applied Mathematics
Title: Wrong by Design: How to Choose Experiments with an Imperfect Model
Abstract: In science, “more data” sounds like it should always mean more information. Unfortunately, that’s not always true. This talk is about what happens when we design scientific experiments using a model that is wrong—or, at best, an imperfect representation of the process we are studying. In Bayesian Optimal Experimental Design (BOED), the usual strategy is to pick the experiment that is expected to provide the most information, but under model misspecification, the one that looks most informative on paper can actually mislead you and make you more confidently wrong. I’ll present a new framework for robust BOED that accounts for this risk, and I’ll introduce Expected Generalized Information Gain (EGIG), an objective function that balances learning against model error. I’ll also show examples of how it works in practice on experiments involving nonlinear dynamical systems, variational inference, and surrogate models.