Foundation models for astrophysics

Abstract

Pre-trained representations from large volumes of unlabelled astronomical survey data are rapidly emerging as a key AI technology in astrophysics. I will review some of the recent applications of these models, including within the software pipeline for the ESA Euclid mission, highlight their advantages – and some potential issues. I will describe our recent work looking at recovering calibrated uncertainties from such models and conclude with a future outlook.

Location

Organiser