The Decoupling Network Assumptions for Optimal Learning in Recurrent Neural Networks