Qu'est-ce que « Representation learning » ?

Representation learning fits a map $h = e_\theta(x)$ from raw inputs to a lower-dimensional code $h$ on which later tasks are easier; the map is judged by what those tasks achieve with it.