“The PAU Survey: Photometric redshifts using transfer learning from simulations” by M. Eriksen et al. has been uploaded to arXiv.

It is the first paper demonstrating we can constrain PAUS redshifts with deep learning techniques. Previously we had (Eriksen 2019) shown it worked with template fitting method. Using a deep neural network we managed to improvethe photo-z scatter with 50% for the faintest galaxies. This was possible througha set of different techniques introduced in the paper. Among the most important was combining simulated and observed data when training the neural network.We also included techniques like auto-encoders to extract information about the galaxy SEDs.

Link to arXiv