Abstract

Scientific images provide a challenging testbed for computer vision because they often differ fundamentally from natural images in both how they are observed and how they appear. This is especially true in astronomy, where images exhibit characteristic noise properties, large dynamic ranges, instrumental or reconstruction effects that are not well represented by the visual features learned from natural-image datasets. Images often are not even the primary scientific data product, being reconstructed from visibilities, accompanying spectra, or serving as a basis for photometry. Astronomical observatories produce some of the largest data volumes of any scientific instrument, and much of the output lacks meaningful labels. While conventional data analysis struggles to keep up with the scale, machine learning based approaches must address important questions about representation learning under domain shift, the adaptation of pretrained vision models, and the development of computationally efficient models for scientific applications.

My work explores several approaches to these challenges. First, I investigate self-supervised learning on wide-field radio survey images, where representations are learned directly from unlabeled observations. I then consider multimodal learning, using parameter-efficient alignment of pretrained vision and spectral encoders to exploit complementary information across astronomical data modalities. Finally, I examine ways to leverage large pretrained models while reducing their computational requirements, through parameter-efficient adaptation and distillation into smaller domain-specific networks.

Speaker Bio

Erica Lastufka is a Research Scientist at the University of Geneva, working with the departments of Computer Science and Astronomy on projects for the Swiss SKA Consortium. She received her PhD from ETH Zurich, working with the Institute for Data Science at FHNW on developing instrumentation and multi-messenger analysis techniques. Her research focuses on foundation models, self-supervised learning, and efficient model adaptation for scientific imaging.