A modular design encourages neural models to disentangle and recombine
different facets of knowledge to generalise more systematically to new tasks.
In this work, we assume that each task is associated with a subset of latent
discrete skills from a (potentially small) inventory. In tur