The PAU Survey and <i>Euclid:</i> Improving broadband photometric redshifts with multi-task learning

Current and future imaging surveys require photometric redshifts (photo-zs) to be estimated for millions of galaxies. Improving the photo-z quality is a major challenge but needed advance our understanding cosmology. In this paper we explore how synergies between narrow-band data large can exploited improve broadband redshifts. We used multi-task learning (MTL) network estimates by simultaneously predicting photometry from photometry. The only required in training field, which also enables better predictions galaxies without wide field. This technique was tested with Physics Accelerating Universe Survey (PAUS) COSMOS find that method predicts photo-zs are 13% more precise down magnitude i_{AB} < 23; outlier rate 40% lower when compared baseline network. Furthermore, MTL reduces bias high-redshift galaxies, improving redshift distributions tomographic bins z>1. Applying deeper samples crucial such as \Euclid or LSST. For simulated data, on sample <23, scatter 16% all i_{AB}<25. studied effects extending using PAUS high-precision photo-zs, 20%