Abstract
The landmark experiments by Posner in the late 1970s have shown that reaction time (RT) is faster when the stimulus appears in an expected location, as indicated by a cue; since then, the so-called Posner task has been considered a "gold standard" test of spatial attention. It is thus fundamental to understand the neural mechanisms involved in performing it. To this end, we have developed a Bayesian detection system and small integrate-and-fire neural networks, which modeled sensory and motor circuits, respectively, and optimized them to perform the Posner task under different cue type proportions and noise levels. In doing so, main findings of experimental research on RT were replicated: the relative frequency effect, suboptimal RTs and significant error rates due to noise and invalid cues, slower RT for choice RT tasks than for simple RT tasks, fastest RTs for valid cues and slowest RTs for invalid cues. Analysis of the optimized systems revealed that the employed mechanisms were consistent with related findings in neurophysiology. Our models predict that (1) the results of a Posner task may be affected by the relative frequency of valid and neutral trials, (2) in simple RT tasks, input from multiple locations are added together to compose a stronger signal, and (3) the cue affects motor circuits more strongly in choice RT tasks than in simple RT tasks. In discussing the computational demands of the Posner task, attention has often been described as a filter that protects the nervous system, whose capacity is limited, from information overload. Our models, however, reveal that the main problems that must be overcome to perform the Posner task effectively are distinguishing signal from external noise and selecting the appropriate response in the presence of internal noise.