Learning rule influences recurrent network representations but not attractor structure in decision-making tasks

Neural Information Processing Systems 

RNN computation is valuable in two ways.