Euclid in a month; adaptable foundation models for science at scale
Mike Walmsley
Dunlap Fellow at the University of Toronto

Citizen science projects like Galaxy Zoo typically take many years to annotate a major survey. Deep learning has long promised to shorten this delay between collecting images and doing science - yet most work has been retrospective, fitting models to long-running surveys with plentiful labels. Label-efficient "foundation" models, pretrained on other surveys, offer the potential for accurate automated measurements in weeks rather than years. I'll discuss our experiments validating this approach on several major surveys and ultimately deploying it this past August to measure the morphology of every galaxy in Euclid. I'll reflect on how astronomers might best use such models and on the new science they might enable.

Date:	Jeudi, 3 octobre 2024
Time:	11:30
Where:	All
	En ligne seulement / Online only