[back]

Attention As A Result of Distributed Competition
Fred H. Hamker

Recordings in V4, IT, MT, MST, PFC and FEF reveal influences of attention on the average rate activity of neurons. However, it is still missing a global picture of the process of attention, i.e. the origin of spatial attention and the interactions between feature-based and spatial attention. We investigate the possibility of a spatial stimulus reentry from the frontal eye field into extrastriate visual areas by means of a quantitative comparison between simulations and experimental data. Let us therefore assume the following cue-guided search task (Chelazzi et al., 1998). A subject has to perform an eye movement to a target, but not to non-targets. An item is defined as the target by presenting it right before the visual search task. First, this item has to be held in memory. Once an array containing two or more items is presented, the subject has to identify the target. We ask the question: Is identification only possible after the deployment of attention as suggested by early limited capacity theories (Treisman & Gelade, 1980; Wolfe et al., 1989)? If an identification could be done in parallel without spatial attention, how do the systems for action planning detect the location of the identified item and where could there be limitations on the interaction between the system for recognition and action selection? The frontal eye field (FEF) in known for its involvement in the planning of volountary saccades. Recordings in the FEF during natural scanning revealed that neurons with visual activity increase their rate when the target of the next saccade lies within their receptive field. Similar effects have been observed during a conjunction search. Visually responsive cells discriminated whether the target or a distractor appeared in their receptive field, and the activity was stronger for distractors that shared a feature with the target (Schall, 1995; Bichot and Schall, 1999).

We address this issue by means of a computational approach. We assume that the brain processes stimuli in a parallel and fast bottom-up manner and integrates the processing in different brain areas by a reentry from higher brain areas into extrastriate visual areas. Thus, our model consists of reciprocally connected functional blocks (V4, IT, PFC, FEF). Its population dynamics is described by a functional mapping between groups of neurons and simulated on the basis of differential equations.

Our population code approach based on continuous nonlinear dynamics allows the description of elementary brain functions like recognition, attention and decision. The simulation of such dynamics allows us to overcome the limitations of several models, which define attention by the competition of only one spatially organized map. Our simulations show that the attentional effects in this experiment can be described without a map that explicitly represents attention Ð instead attention is an emergent result of the inherent competition in processing stages. According to our model, memory guided visual search consists of two phases: parallel feature-based identification in PFC+IT and spatial selection in FEF (Fig. 1). A match of the pattern in working memory with the one that enters IT during the array presentation leads to an increase of activity in IT. The receptive field of cells in IT comprises large areas of the total visual field. They can hardly indicate the target location. But the advantage of IT cells is transferred downwards and again, those bottom-up patterns matching the top-down pattern can increase their activity. Increased local activity in these feature streams enhances the visually responsive neurons in the frontal eye field. These cells reflect the task-relevance of a location, since under normal conditions the FEF is not sensitive to specific features.

According to our simulations, FEF cells with a strong phasic component, like visual and visuomovement cells, are not the source of spatial attention. We predict that spatial attention is tightly connected to premotor movement neurons in the FEF. Their increase depends on the activity of the visually responsive neurons, but their activation is suppressed until the match of the pattern in working memory with one of the patterns that enter IT is successful. The increasing activity of a movement cell is sent into extrastriate visual areas and facilitates the processing within its movement field. Thus, neurons in V4 at the location of the intended eye movement gain a further advantage. These processes clean up the population activity in higher stages like IT from all unimportant stimuli so that a full recognition can take place.

Figure 1. Activity during a visual search experiment. The presentation of the cue elicits a response in the M-IT cells, which is stored by cells in a prefrontal memory through recurrent excitation. During the delay, the active memory cells project into M-IT and increases the baseline activity of the cell sensitive for the good stimulus. Such feedback is called feature based attention. As soon as the display containing the good and the poor stimulus is presented their patterns are processed bottom-up without any specific bottleneck. However, when the stimuli enter M-IT, the good stimulus receives an advantage by the active feedback from memory cells. This advantage is sent to M-V4 cells, which have smaller RFs. Since M-FEF visual cells receive their main input from M-V4 the advantage of a stimulus feature is transferred into an advantage of a location. Initially, the FEF movement cells with the good and the poor stimulus in its movement field are able to gain activity, but differently to the visual cells the good stimulus quickly outperforms the poor stimulus. The activity of the movement cells enters extrastriate visual cortex and modulates the activity in M-V4 and M-IT, which results in space based attention. This reentry is essential for a sustained activity of the good stimulus at the selected location and allows the further suppression of the behaviorally unimportant stimulus.


References
Effects of similarity and history on neural mechanisms of visual selection. Bichot, N.P.; Schall, J.D. Nature Neuroscience, 2: 549-554, 1999.

Responses of neurons in inferior temporal cortex during memory-guided visual search. Chelazzi, L.; Duncan, J.; Miller, E.K.; Desimone, R., J. Neurophysiol., 80: 2918-2940, 1998. Schall JD. Neural basis of saccade target selection. Rev Neurosci., 6:63-85, 1995.

A feature integration theory of attention. Treisman, A.; Gelade, G. Cognitive Psychology, 12:97-136, 1980.

Guided Search: An alternative to the feature integration model for visual search. Wolfe, J.; Cave, K.; Franzel, S.: Journal of Experimental Psychology, 15:419-433, 1989.