Thanks to large public spectroscopic surveys we now have a wealth of high quality chemical abundance data of Milky Way (MW) stars. This data has been instrumental to better understand the past evolution of our Galaxy and its mass accretion history. However, due to our current technological limitations, a selection of sources has to be previously imposed, making spectroscopic surveys biased from design. With a much more modest spectral resolution, new multifilter surveys could fill the gaps left by spectroscopic surveys, providing flux-limited data for all the stars within their footprint. While and advantage, the unbiased nature of these surveys makes necessary a posterior identification and classification of all the observed sources, such as stars, galaxies, or quasi-stellar objects (QSOs).

In this talk, I will present the results of object classification using deep Bayesian neural networks in conjunction with the third data release of J-PLUS, a photometric survey to study the nearby Universe using 12 bands, seven of them narrow. Thanks to BANNJOS, we have been able to accurately classify more than 47 million sources into ~26 million stars, ~20 million galaxies and ~1 million QSOs. We are currently working on extending the methodology to derive photometric redshifts for galaxies and QSOs, and stellar parameters for MW stars. These techniques will be later implemented in J-PAS, the big sibling of J-PLUS, which will cover 8500 deg2 of the northern hemisphere with a staggering 59 narrow-band filters and up to a magnitude r~23 mag, providing low-res spectra-like data for hundreds of millions of sources in the MW and the distant Universe.