Learning Hierarchical Information Flow with Recurrent Neural Modules

   Abstract