Foundation Models and Discovery in Astronomy | AI for Science

Foundation Models and Discovery in Astronomy | AI for Science At the AI for Science event, Prof. Anna Scaife (University of Manchester and an Alan Turing Institute AI Fellow) outlines how foundation models can be trained once and reused across many downstream astronomy tasks, across modalities such as images, spectra, polarization and time-domain data. She presents ZooBot, built from Galaxy Zoo citizen-science labels to predict answers to ~200 morphology questions, showing how its latent space groups morphologies and unexpectedly clusters “weird” galaxies; fine-tuning with as few as 10 examples enabled discovery of ~40,000 ring galaxies and rapid science. ZooBot has been operationalized in ESA’s Euclid pipeline to produce morphology catalogues immediately. She discusses limits of supervised scaling in astronomy, the need for self-supervised learning in radio astronomy, and methods for detecting unusual objects via energy-based out-of-distribution scoring, while noting concerns that representations may lose information for the strangest sources and emphasizing uncertainty calibration and bias quantification. Anna Scaife is Professor of Radio Astronomy at the University of Manchester. Her research focuses on the use of artificial intelligence for discovery in data-intensive astrophysics and in 2019 she was appointed as one of the five inaugural AI Fellows of the UK’s Alan Turing Institute. She is Director of Research for the UKRI AI Centre for Doctoral Training in Decision Making for Complex Systems, UK representative to the International Union of Radio Sciences (Commission J: Radio Astronomy), Deputy Editor-in-Chief of the Royal Astronomical Society’s Techniques & Instruments (RASTI) journal, and Subject Editor (Astrophysics) for the Royal Society Open Science journal. 00:00 Speaker Introduction 00:57 What Are Foundation Models 02:45 Foundation Models In Astronomy 04:51 ZooBot And Galaxy Zoo 07:58 Latent Space And Weird Galaxies 12:29 Finding Ring Galaxies Fast 15:45 From Papers To Euclid Pipeline 18:32 Why In Domain Pretraining Matters 19:49 Scaling Limits And Self Supervision 22:13 Radio Galaxies Explained 25:04 Known Unknowns Vs Unknown Unknowns 27:01 Energy Scores For Weirdness 31:04 Caveat Intrinsic Dimension Loss 32:41 Wrap Up Pros Cons And Outlook