Modular Training of Neural Networks aids Interpretability