BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:From Squiggles to Signals: Learning Useful Representations for Dis
 covery in Time-Domain Astronomy - Daniel Muthukrishna (MIT)
DTSTART:20250610T150000Z
DTEND:20250610T160000Z
UID:TALK232987@talks.cam.ac.uk
CONTACT:65128
DESCRIPTION:New large-scale astronomical surveys are observing orders of m
 agnitude more sources than previous surveys\, making standard approaches o
 f visually identifying new and interesting phenomena unfeasible. Upcoming 
 surveys such as the Vera Rubin Observatory's Legacy Survey of Space and Ti
 me (LSST) and ongoing surveys such as the Transiting Exoplanet Survey Sate
 llite (TESS) have the potential to revolutionize time-domain astronomy\, p
 roviding opportunities to discover entirely new classes of events while al
 so enabling a deeper understanding of known phenomena. The opportunity for
  serendipitous discovery in this domain is a new challenge that can be mad
 e systematic with data-driven methods\, which are particularly suitable fo
 r identifying rare and unusual events in large datasets. In this talk\, I
 ’ll explore the potential for anomaly detection and representation learn
 ing in big datasets\, and describe the challenge of applying these methods
  to real-time surveys. I’ll present novel machine learning methods for a
 utomatically detecting anomalous transient events such as kilonovae and pe
 culiar supernovae\, and characterising variable stars. I’ll explore the 
 challenge of developing representative latent spaces useful for downstream
  machine learning tasks and present a novel causally-motivated foundation 
 model. I’ll apply the approach to transients from the Zwicky Transient F
 acility (ZTF) and simulations of variable stars while discussing applicati
 ons to upcoming surveys.\n
LOCATION:Martin Ryle Seminar Room\, KICC
END:VEVENT
END:VCALENDAR