In this paper, we push forward the idea of machine learning systems for which the operators can be modi ed and netuned for each problem. This allows us to propose a learning paradigm where users can write (or adapt) ...