On Learning the Statistical Representation of a Task and Generalizing it to Various Contexts