Revisiting Bi-Linear State Transitions in Recurrent Neural Networks