Rethinking the Relationship between Recurrent and Non-Recurrent Neural Networks: A Study in Sparsity