Dissociating performance from learning: An empirical evaluation of a computational model