BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:AION-1: An Omnimodal Foundation Model for Astronomical Sciences - 
 Thomas Hehir (IoA)
DTSTART:20260302T160000Z
DTEND:20260302T170000Z
UID:TALK244183@talks.cam.ac.uk
CONTACT:65128
DESCRIPTION:Modern astronomical surveys produce vast\, heterogeneous datas
 ets spanning multiband photometric images\, spectroscopy\, and rich metada
 ta. While machine learning methods have achieved impressive results within
  individual modalities\, most models remain specialised and task-specific\
 , making it difficult to reuse them across surveys and scientific problems
 . In this talk\, I will present AION-1\, the first family of large-scale (
 up to 3.1B parameters) multimodal foundation models designed for the astro
 nomical sciences.\n\nAION-1 employs modality-specific tokenizers to conver
 t diverse inputs—images\, spectra\, and scalar properties—into a commo
 n representation space\, where a transformer learns joint structure across
  modalities via masked-token modeling. We pretrain on over 200 million sta
 rs\, galaxies\, and quasars from five major surveys: Legacy Survey\, HSC\,
  SDSS\, DESI\, and Gaia\, homogenising the treatment of observations with 
 unique instrument signatures.\n\nAION-1 achieves state-of-the-art performa
 nce across a broad suite of downstream tasks\, including physical property
  estimation\, morphology classification\, similarity search\, image segmen
 tation\, and spectral super-resolution\, with minimal task-specific finetu
 ning. I will discuss its effectiveness in low-data regimes and its capacit
 y to learn survey-agnostic universal representations. As a fully open-sour
 ce framework\, AION-1 provides a scalable blueprint for building multimoda
 l foundation models capable of integrating heterogeneous observations acro
 ss the physical sciences. All code\, data\, and weights are publicly avail
 able.
LOCATION:Martin Ryle Seminar Room\, KICC
END:VEVENT
END:VCALENDAR