Machine learning has been widely applied to clearly defined tasks in astronomy and astrophysics. Contrastively, in a sequence of our recent works we have gone beyond tasks and have focused on letting deep architectures ``listen'' to the real, raw, astrophysical data, letting it speak for itself.
During the talk, I will showcase two implementations of this idea on stellar spectra: The first work, called Astro-machines, demonstrates how a machine can start to make sense of raw numerical data and begin learning known astrophysical parameters from them, without being asked to do so! The second one, called Stellar Karaoke, shows how machines can provide us with novel insights into a long-standing problem, namely the removal of adversarial atmospheric effects, just by examining a large number of raw numerical vectors.