Joshua Speagle

University of Toronto

While many machine learning (ML) applications are often rooted in statistics or particular domain applications (e.g., image processing), applications in astronomy often tend to leave behind a lot of the types of thinking involved with designing traditional astronomical or statistical models. In this somewhat informal talk, I will try to highlight a number of examples where "astrostatistical thinking" from myself and others has proven fruitful in tackling problems in several areas. Topics may include (based on audience interest): incorporating various uncertainties into ML models, dealing with outliers in training and prediction, designing robust ML workflows, vetting and interpreting ML-driven results, "calibrating" posterior uncertainties, and trying to select between various high-dimensional non-parametric ML models (based on audience interest). Applications for these methods range from photometric redshift estimation, 3-D dust mapping, chemical tagging, stellar age estimation, and more.

Date: Jeudi, le 12 octobre 2023
Heure: 11:30
Lieu: Université de Montréal
  Pavillon MIL A-3561