Training Neural Networks for Modularity aids Interpretability