AstroAI Lunch Talks - February 24, 2025 - Dimitri Zagorulia
10 Feb 2025 - Joshua Wing
The video can be found here: https://www.youtube.com/watch?v=jb1XsqaXnCs
Speaker: Eddie Berman (Northeastern)
Title: On the Uncertainty Calibration of Equivariant Functions
Abstract: Uncertainty estimation shows pronounced importance in the data-sparse settings where equivariant models tend to excel, including pick-and-place robotics tasks, galaxy morphology classification, and chemical physics. This talk explores the relationship between equivariance, model calibration, and model confidence. In particular, we discuss bounds on calibration error for functions that are arbitrarily expressive except that they are constrained to be equivariant with respect to a group G. Following this, we discuss experiments related to Astrophysics and beyond that explore how these bounds manifest in practice.