In order to create appealing animation, animators define the key poses of a character by manipulating its underlying skeletons' joints. To look plausible, a human pose must respect many ill-defined constraints and the resulting realism greatly depends on the author's eye for details. Computer animation software propose tools to help in this matter, relying on various algorithms to automatically enforce some of these constraints. The increasing availability of motion capture data has raised interest in data-driven approaches to pose design, with the potential of shifting more of the task of assessing realism from the artist to the computer. In this paper, we propose such a method, relying on neural networks to learn the constraints from the data and to create an alternative representation of the pose space. We then demonstrate one application of this space by performing pose edition through optimization of a pose's latent representation.