Foundation models and representation learning for SKA precursor images

Our research develops scalable deep learning frameworks to automate the analysis of massive  astronomical datasets. We focus on bridging the gap between general-purpose computer vision and the specialized requirements of astrophysics images by evaluating and adapting vision foundation models for tasks like galaxy morphology classification and source detection. Using self-supervised learning, we have demonstrated that models can learn highly meaningful representations directly from source-rich wide-field images without the need for manual pre-processing or individual galaxy cutouts. Additionally, we find that generalized representations learned by commercial models trained on natural images or image-text pairs preserve features useful for certain astrophysical downstream tasks.
 

Involved researchers

Erica Lastufka, Omkar Bait, Mariia Drozdova, Vitaliy Kinakh, Davide Piras, Marc Audard, Miroslava Dessauges-Zavadsky, Taras Holotyak, Daniel Schaerer, Slava Voloshynovskiy