Capacity and Trainability in Recurrent Neural Networks

   Abstract