Interpretability in Gated Modular Neural Networks