REVIEWS CONTROL OF GOAL-DIRECTED AND STIMULUS-DRIVEN ATTENTION IN THE BRAIN Maurizio Corbetta and Gordon L.Shulman ngnono ocon co sm works of brain areas that cary out diferent ctions.One system. udes the ter etal cortex and inferior frontal co the right hemisphere.isnot invlved in top-down selection.Instead,this system is specialized for the detection of behaviourally relevant stimull,particularly when they are salent or unexpected. This ventral frontoparletal network works as a'circult breaker for the dorsal system,directing attention to saient events.Both attentional systems interact during normal vision,and both are disrupted in unilateral spatial neglect. pantigatogetheranlhmystestertediotg by the e wards the sou fa ors tha 吃ge black-a -white dice andse that we migh The dynami ls wher we tion can be de on the been drawn to the n Th Dt175 NATURE REVTEWST SCIENC OLUME 3TMARCH 26022
REVIEWS NATURE REVIEWS | NEUROSCIENCE VOLUME 3 | MARCH 2002 | 201 Picture yourself at the Museum El Prado in Madrid while a guide explains the painting The Garden of Earthly Delights by the fifteenth-century Flemish painter Hieronymous Bosch (FIG. 1). Bosch depicts a fantastic, surreal and satirical world, which is in stark contrast to anything else represented until that time. The guide’s words cue us to attend to different aspects of the painting, such as its colour, spatial configuration or meaning. For example, if he notes “a small animal playing a musical instrument”, we can use this information to spot the rabbit playing the horn near a black-and-white dice. Knowledge and expectations allow us to focus on elements, parts or details of a visual scene that we might otherwise have missed. Cognition aids vision by enabling the brain to create, maintain and change a representation of what is important while we scan a visual scene. At the other extreme, visual perception can be dominated by external events. Initially, our eyes might have been drawn to the more salient objects in the painting, such as the large wooden musical instrument (a lute in construction) at the centre of the scene, rather than to more subtle aspects of the painting that are discussed by the guide. An event might even distract us from the painting altogether. If an alarm system started to ring and flash in a nearby room, everyone’s attention would instantly be drawn towards the source of the alarm. Unexpected, novel, salient and potentially dangerous events take high priority in the brain, and are processed at the expense of ongoing behaviour and neural activity. In everyday life, visual attention is controlled by both cognitive (TOP-DOWN) factors, such as knowledge, expectation and current goals, and BOTTOM-UP factors that reflect sensory stimulation. Other factors that affect attention, such as novelty and unexpectedness, reflect an interaction between cognitive and sensory influences. The dynamic interaction of these factors controls where, how and to what we pay attention in the visual environment. In this review, we propose that visual attention is controlled by two partially segregated neural systems. One system, which is centred on the dorsal posterior parietal and frontal cortex, is involved in the cognitive selection of sensory information and responses. The second system, which is largely lateralized to the right hemisphere and is centred on the temporoparietal and ventral frontal cortex, is recruited during the detection of behaviourally relevant sensory events, particularly when they CONTROL OF GOAL-DIRECTED AND STIMULUS-DRIVEN ATTENTION IN THE BRAIN Maurizio Corbetta and Gordon L. Shulman We review evidence for partially segregated networks of brain areas that carry out different attentional functions. One system, which includes parts of the intraparietal cortex and superior frontal cortex, is involved in preparing and applying goal-directed (top-down) selection for stimuli and responses. This system is also modulated by the detection of stimuli. The other system, which includes the temporoparietal cortex and inferior frontal cortex, and is largely lateralized to the right hemisphere, is not involved in top-down selection. Instead, this system is specialized for the detection of behaviourally relevant stimuli, particularly when they are salient or unexpected. This ventral frontoparietal network works as a ‘circuit breaker’ for the dorsal system, directing attention to salient events. Both attentional systems interact during normal vision, and both are disrupted in unilateral spatial neglect. TOP-DOWN PROCESSING The flow of information from ‘higher’ to ‘lower’ centres, conveying knowledge derived from previous experience rather than sensory stimulation. BOTTOM-UP PROCESSING Information processing that proceeds in a single direction from sensory input, through perceptual analysis, towards motor output, without involving feedback information flowing backwards from ‘higher’ centres to ‘lower’ centres. Departments of Neurology, Radiology, and Anatomy and Neurobiology, Washington University School of Medicine, St Louis, Missouri 63110, USA. Correspondence to M.C. e-mail: mau@npg.wustl.edu DOI: 10.1038/nrn755
REVIEWS no th Top-downcotenion cene when they kno This facil on dep on our a the ubiects know in advar eme ton of the (uc orthe m s but th nd motor sets on the yof the cue on of 'at d in the selectio ived with s ary or e b the guish the ight,and see 1a hom ologous nke hat haw ted these preparatory control signal logue of the FEFAccordingly,neurons in macaque 202 MARCH2002 VOLUM正 nature com/reviews/neuto
FRONTAL EYE FIELD An area in the frontal lobe that receives visual inputs and produces movements of the eye. 202 | MARCH 2002 | VOLUME 3 www.nature.com/reviews/neuro REVIEWS Recent reviews have discussed how attentional sets modulate the neural response to a target8,9. Top-down signals for spatial attention.In studies of attention, advance information is typically provided in the form of a cue (for example, a small arrow) that instructs observers about some relevant aspect of the forthcoming visual scene (such as the location or direction of motion of a target stimulus). FIGURE 2a shows the results of a functional magnetic resonance imaging (fMRI) experiment in which subjects view, for two seconds, an arrow cue that tells them to direct covertly (without moving their eyes or head) their attention to a location in the periphery of the visual field10. The pattern of brain activation is displayed on the surface of the brain. Areas in the occipital lobe (fusiform and MT+) respond transiently to the cue, whereas areas in the dorsal posterior parietal cortex along the intraparietal sulcus (IPs) and in the frontal cortex (at or near the putative human homologue of the FRONTAL EYE FIELD, FEF) show a more sustained response. The transient response in occipital areas might reflect the sensory analysis of the cue. We know that the sustained part of the response (grey arrow) is endogenous, because it is not related to either visual stimuli or motor responses, and it is time-locked to the period in which subjects pay attention to the peripheral location. The activation is predominantly bilateral for attention to either visual field10–12, but in a subset of parietal (ventral IPs) and frontal (FEF) areas, the response is spatially selective (stronger when attention is directed towards the contralateral visual field12,13). This pattern of brain activation indicates that parietal and frontal regions are involved in controlling the location of attention. FIGURE 2c (left) shows results from other recent brainimaging studies that have separated preparatory signals for attending to visual objects from signals that are related to the visual analysis of, detection of or response to those objects10–12,14. The areas most consistently activated by attention to stimulus attributes include the dorsal parietal cortex along the IPs, extending dorsomedially into the superior parietal lobule (SPL) and anteriorly towards the postcentral sulcus, and the dorsal frontal cortex at the intersection of the precentral and superior frontal sulci (the putative human FEF). During the cue period, these frontoparietal activations are sometimes accompanied by sustained activity in early occipital regions11, but these lower-level activations are not always observed10,14 (FIG. 2a), and might depend on the complexity of the cued information or on other factors. These findings complement an earlier body of imaging work using positron emission tomography (PET) and fMRI in which attention signals were not recorded in isolation but were mixed with signals that reflected the target stimulus and motor response. In various detection and discrimination paradigms, voluntary or endogenous visual selection most consistently activated the dorsal parietal and frontal cortex15–21 (FIG. 2c, right, and see REF. 22). The human parietal areas that are active during spatial attention might be homologous to areas in the monkey IPs23, whereas the frontal area might be the human homologue of the FEF24. Accordingly, neurons in macaque are salient and unattended. Here, we review the psychological, neuropsychological and physiological evidence for these two systems. Top-down control of attention Human observers are better at detecting an object in a visual scene when they know in advance something about its features, such as its location, motion or colour1–5. This facilitation depends on our ability to represent this advance information (a ‘perceptual set’), and to use it to bias the processing of incoming visual information. Similarly, responses to a stimulus are quicker when subjects know in advance what type of movement they have to make (such as which arm to move or the direction of the movement — a ‘motor set’)6,7.As most studies involve the selection of both stimuli and responses, it is sensible to define perceptual and motor sets as processes that link relevant sensory information to relevant motor information.We use the more general term of ‘attentional set’to define the representations involved in the selection of task-relevant stimuli and responses. Conceptually, it is important to distinguish control signals for the generation and maintenance of an attentional set from the top-down effects of that set on the neural activity evoked by the target stimulus that the subject must detect or identify. One way to distinguish these control signals is to separate in time the neural responses to the advance information from the responses to the target stimulus. Therefore, this review focuses on studies that have isolated these preparatory control signals. Figure 1 | The Garden of Earthly Delights by Hieronymous Bosch. Courtesy of the Corbis Picture Library
REVIEWS A 人人人 A in Ma 三怨et的 : h OLUME 3 MARCH 2002 203
NATURE REVIEWS | NEUROSCIENCE VOLUME 3 | MARCH 2002 | 203 REVIEWS FEF Attention only pIPs SPL PoCes Corbetta, 2000 Hopfinger, 2000 Kastner, 1999 Shulman, 1999 PrCes SFs Attention + stimulus a b c d aIPs pIPs vIPs vIPs Stimulus onset Fus Attend left Min Max Min Max Attend right Attend direction View passively MT+ Cue FEF aIPs pIPs MT+ Cue Vandenberghe, 1997 Nobre, 1997 Corbetta, 1993 Corbetta, 1995 Gitelman, 1996 Woldorff, 1997 Vandenberghe, 1996 Attention to motion direction Attention to location Stimulus onset Attention to location Fixation Figure 2 | Dorsal frontoparietal network for top-down control of visual attention. a | Human brain activity produced by attending to a location. Subjects see an arrow that cues one of two locations and covertly attend to the location indicated in preparation for a target at that location. The functional magnetic resonance imaging (fMRI) response is averaged over 13 subjects10. The graphs show the time course of the fMRI signal after the cue. Signals are transient in occipital regions, but sustained (grey arrow) in the parietal cortex and frontal eye field (FEF). aIPs, anterior intraparietal sulcus; Fus, fusiform cortex; MT+, middle temporal complex; pIPs, posterior IPs; vIPs, ventral IPs (junction of the vIPs and transverse occipital sulcus). b | Human brain activity produced by attending to a direction of motion. Subjects see an arrow that cues a direction of motion and prepare for a subsequent target moving in that direction. The response is averaged over 14 subjects14. The graphs contrast the response seen when the arrow cues a direction of motion with that observed when a cue instructs the subject to view the display passively. Directional cues produce sustained signals in frontal (FEF) and parietal (aIPs, pIPs) areas, but transient signals in occipital areas (MT+). c | Left: meta-analysis of studies of visual attention. Subjects expected a simple visual attribute, such as location10,12 or direction of motion14, or a more complex array11. Foci of activation from the expectation period are smoothed and projected onto the Visible Human Brain117. The area of maximal overlap between studies is in the pIPs. PoCes, postcentral sulcus; PrCes, precentral sulcus; SFs, superior frontal sulcus; SPL, superior parietal lobule. Right: meta-analysis of imaging studies of visual attention and detection (adapted with permission from REF. 22 © 1998 National Academy of Sciences, USA). The figure shows regions that are activated by attending to and detecting visual stimuli (preparatory activity has been averaged with visual- and motor-detection activity). d | Anticipatory activity in a macaque V3A neuron during a memory-guided task (adapted with permission from REF. 27 © 2000 The American Physiological Society). The graph shows the single-unit activity as the monkey performs two tasks while fixating the centre of the screen. In the fixation task, a small stimulus is presented peripherally and the monkey is rewarded for maintaining fixation. In the memoryguided task, the monkey has to remember the location of the stimulus and after a variable delay make a saccade to that location. Activity is increased in the memoryguided task before stimulus onset, perhaps reflecting attention to the stimulus location
REVIEWS M Time (s) me is Attend loft Attend right the location of the V3 in b his p which res human nantiei pation o a sti patial attention r the biects at th Th down signals for feature or obied attentio n-lab nt visu 01 or pari face in The that the the reappe f an object that quent display that contains a few cohe ently moving relevant location for an eye movement 204 MARCH2002 VOLUM正3
204 | MARCH 2002 | VOLUME 3 www.nature.com/reviews/neuro REVIEWS dots (the target) among many randomly displaced dots (the distracters). Subjects cannot use location as a cue to find the target. Similar parietal, frontal and occipital areas are recruited by the cue as were activated by the spatial cue, and this activity is sustained only in frontal and parietal regions14. Interestingly, a region in the posterior IPs that is well activated by cues for motion is poorly activated by cues for colour, indicating that there is some specialization within the parietal cortex for the type of information attended29. Posterior parietal regions are also active when subjects switch their attention between two objects at the same location30. The location of these parietal regions might be different from those that are active in attending to location or direction of motion, but this will need to be confirmed by within-experiment or within-laboratory comparisons. Finally, several studies that did not specifically isolate attention signals have reported modulations in posterior parietal cortex during tasks that require the identification of features of foveal stimuli19,31. There is also evidence from single-unit studies for a possible role of these regions in coding an attentional set for features such as motion or colour. Neurons in area LIP show directional motion-selective activity when a monkey expects the reappearance of an object that is moving behind another one, consistent with an expectation signal for moving objects32. Neurons in LIP flexibly code colour information, when colour indicates a taskrelevant location for an eye movement33. frontal and parietal cortex (FEF, lateral intraparietal area (LIP), and areas 7a and V3A) show increases in baseline firing rate when the monkey anticipates the onset of a stimulus25–27 (FIG. 2d). This prestimulus anticipatory activity could correspond to the fMRI signal that is recorded in human subjects during anticipatory attention. These regions in dorsal parietal and frontal cortex, which respond when both human and monkey observers covertly pay attention to a peripheral location in anticipation of a stimulus, might form a network (dorsal frontoparietal network) for the control of visuospatial attention. However, they also carry neuronal signals that are related to the preparation of eye and arm movements, and to stimulus processing (see below). Top-down signals for feature or object attention. Attending to location is just one way in which we can select relevant visual information. We can also attend to different features of an object, such as its shape, colour or direction of motion, or to objects, such as a familiar face in a crowd, which can be defined by many different features28. There is growing evidence that the frontoparietal cortical network that is recruited for spatial attention is also involved in other types of visual selection. FIGURE 2b shows a map of the brain activation that occurs when subjects expect to see moving stimuli. The arrow cue (as in FIG. 2a) provides advance information about the direction of motion (left or right) of a subsequent display that contains a few coherently moving Sample a b Distracter Match 85 80 75 70 65 Per cent correct Without colour responses With colour responses 0.35 0.25 0.15 0.05 –0.05 0 5 10 FEF IPs IPs FEF 15 20 0.35 0.25 0.15 0.05 –0.05 0 5 10 MT+ Time (s) Time (s) Time (s) fMRI signal fMRI signal fMRI signal 15 20 0.35 0.25 0.15 0.05 –0.05 0 5 10 15 20 Attend left Delay activity Attend right Cue Delay MT+ Figure 3 | Overlap between working memory and attention. a | In a match-to-sample task, subjects remember the location of the sample stimulus and decide after a variable delay whether it corresponds to the location of the match stimulus. Accuracy is significantly impaired when subjects have to shift attention to the distracter and report its colour (“with colour responses”). Control experiments show that this decrement is not due to increases in task difficulty. These results indicate that spatial rehearsal (maintaining a spatial location in memory) depends on spatially attending to the location (adapted with permission from REF. 36 © 2001 Elsevier Science). b | The dorsal frontoparietal network is active during spatial rehearsal13. The task is the same as that in FIG. 2a, except that subjects have to maintain attention to a peripheral location for ~7 s (2.4-s cue period + 4.7-s delay period). The functional magnetic resonance imaging (fMRI) signal remains sustained during the delay (delay activity shaded in blue) in regions of the dorsal frontoparietal system (FEF, IPs), but returns to baseline in MT+ after the sensory analysis of the arrow cue. FEF, frontal eye field; IPs, intraparietal sulcus; MT+, middle temporal complex
REVIEWS and neurons in a more anterior area of the IP with atendingtolocatk in the for arn oto aining attentional sets in memory What psy h I and neural m begun to is which defined as the ability g human areas e of in sory or mot ns for directing a ntoa it is pos ed by the premo nt in the sion of the attention ithovetoculonmotor& ng atte a lo king he bilit Too-down nulu 2 imuli at w th un tote is re ed to the tal n sin LIP that are r H d in bo be th dine for th source of a n al fire pp evs ure po d.red reg ese regions mple ring t he IPs (FG 4ergosd (specif of the p byhomologous areas LIP and oe for the d in task that a person performsn one c of n aratory activity that i he ng the digits.I area LIP code for hat th )impending ven)or letter RE REVIEWS UME 3|MARCH 2002 20
NATURE REVIEWS | NEUROSCIENCE VOLUME 3 | MARCH 2002 | 205 REVIEWS (FIG. 4a), and neurons in a more anterior area of the IPs code for impending grasping movements and threedimensional objects44. Preparatory activity has been observed for eye movements in the FEF45, and for arm movements in the premotor cortex46. Brain-imaging studies have reported both eye- and arm-related activity in the frontal cortex and IPs47–49, although these studies did not separate preparatory signals from those involved in executing the response. More recent event-related studies have begun to isolate preparatory signals that are related to response selection in what are presumed to be corresponding human areas50. Because eye movements are important in stimulus selection, mechanisms for directing attention to a location might be similar to mechanisms for preparing an eye movement, as proposed by the premotor theory of attention51. Imaging studies that have compared covert spatial attention with overt oculomotor shifts have found strong overlap in activations of both the FEF and IPs18,52,53 (FIG. 4b). Top-down signals for task sets. Although stimulus and response selection can be separated in the laboratory, stimuli and responses are inextricably linked in real life. While looking at a canvas, the brain not only selects the stimuli at which to look, but also programs the eye movements with which to look at them. This close functional linkage is related to the convergence of stimulus- and response-selection signals in areas of the frontoparietal network. Neurons in LIP and FEF show signals that are related to attention, memory and eye movements26,38,45. However, in many circumstances, the association between stimulus and response must be learned. For instance, although the natural response to the museum alarm is flight, a firefighter will respond by actively looking for the source of a potential fire. So, a crucial aspect of attentional selection is not just the isolated selection of a stimulus or a response, but the assembly and coordination of stimulus–response associations or mappings. A recent study54 found that online changes in the appropriate stimulus–response mapping for hand responses activated two clusters in posterior parietal cortex: one more posterior and medial, extending from the IPs into the SPL, and one more lateral, along the IPs (FIG. 4d, red regions). These regions were distinct from those involved in stimulus selection, more anteriorly in the IPs (FIG. 4d, green regions, and FIG. 2a,b). The spatial distribution of activity for stimulus selection and stimulus–(hand)-response mapping (specifically, the medial cluster) might correspond to the spatial distribution of the possibly homologous areas LIP and PRR in the macaque54. A change in the task that a person performs on a stimulus often changes the appropriate response (the stimulus–response mapping). Subtracting one digit from another results in a different response from that seen when adding the two digits. Performance analyses indicate that switching between task sets, particularly when they involve the same stimuli, is effortful and expensive in terms of resources and time55,56. For example, switching between categorizing numbers (as odd or even) or letters The dorsal frontoparietal network that is recruited when subjects expect to see object features other than location clearly overlaps with regions that are recruited by attending to location (FIG. 2a,b), but the exact overlap between regions recruited by different kinds of advance information is unclear, and many visual features have yet to be tested. Maintaining attentional sets in memory. What psychological and neural mechanisms are responsible for the maintenance of attentional sets? One likely candidate is working memory, which is defined as the ability to maintain and manipulate information online in the absence of incoming sensory or motor stimulation. Attentional set and working memory mechanisms overlap functionally. For example, it is possible to maintain a memory of a visual attribute for many seconds without any apparent decrement in the precision of the visual information34,35. A more direct link is provided by reports that directing attention away from a location during a delay disrupts working memory for that location36 (FIG. 3a). Therefore, spatial rehearsal, or the ability to maintain spatial information online in memory, depends crucially on spatial attention. The parietal and frontal regions that are recruited by an attentional set (FIG. 2a–c) show sustained activation during a memory delay in which the set is maintained online for up to ten seconds13,37 (FIG. 3b). The localization of working memory signals in dorsal frontoparietal areas is consistent with the presence of memory activity in neurons in macaque FEF and IPs38,39. In macaques, strong memory-related activity has also been found more anteriorly in the lateral prefrontal cortex39, which has been considered in both species to be the main source of top-down control signals to the visual cortex40,41. However, in humans, the current neuroimaging evidence does not support the involvement of dorsolateral prefrontal cortex during the encoding and maintenance of a simple visual cue in a well practised task. This discrepancy might represent a species-specific difference in the neural systems that are involved in attentional control. More anterior prefrontal areas might be recruited in monkeys because of their low memory span42. These regions are probably recruited in humans when the task is initially learned or as the selected items become more complex or increase in number. For example, human prefrontal regions are active during the delay period of match-to-sample tasks in which a sample face is maintained over a delay period and then matched to a test face37. Top-down signals for attending to effectors. Our discussion so far has emphasized a role for the dorsal frontoparietal network in preparatory aspects of stimulus selection, but other results indicate that these areas are also important for response or action selection. Different regions of macaque IPs show preparatory activity that is selective for different effectors. For example, neurons in area LIP code for impending SACCADIC EYE MOVEMENTS, whereas neurons in a more medial area (parietal reach region, PRR) code for impending reaching movements43 SACCADIC EYE MOVEMENT A rapid eye movement that brings the point of maximal visual acuity — the fovea — to the image of interest