2025 workshop program
Deep Learning Photometric Redshifts for Roman
The work
Abstract
Photometric redshifts (photo-z’s) will be essential for studying cosmology, galaxy evolution, and transient science with future space-based photometric surveys, particularly with the Roman Space Telescope. Deep learning methods leverage pixel-level information from images to achieve the best photo-z’s for low-redshift galaxies, but their performance at higher redshifts that are relevant for Roman science remains untested due to limited training data. In this talk, I will discuss our efforts to deploy deep learning photo-z algorithms on high-redshift galaxies using HST CANDELS imaging and redshift labels from spectroscopic, grism, and COSMOS2020 photometric catalogs. Our results show that a semi-supervised deep learning approach which makes use of unlabeled images outperforms fully supervised methods and traditional photometry-based estimates. Furthermore, we demonstrate that our approach consistently improves photo-z estimates across varying amounts of labeled data, with no signs of plateauing – this is crucial as we scale to the vastly larger datasets from future surveys. Semi-supervised deep learning should allow us to take advantage of the information available from the full set of hundreds of millions of galaxies imaged from space, enabling the most accurate photo-z estimates for both faint and bright sources.