Compositional Learning of Visually-Grounded Concepts Using Reinforcement