Attention control in the primate brain

Abstract

Attention is fundamental to all cognition. In the primate brain, it is implemented by a large-scale network that consists of areas spanning across all major lobes, also including subcortical regions. Classical attention accounts assume that control over the selection process in this network is exerted by ‘top-down’ mechanisms in the fronto-parietal cortex that influence sensory representations via feedback signals. More recent studies have expanded this view of attentional control. In this review, we will start from a traditional top-down account of attention control, and then discuss more recent findings on feature-based attention, thalamic influences, temporal network dynamics, and behavioral dynamics that collectively lead to substantial modifications. We outline how the different emerging accounts can be reconciled and integrated into a unified theory.

Introduction

The term attention refers to a set of mechanisms aimed at selectively prioritizing the neural representations that are most relevant to one’s current behavioral goals. Attention mechanisms are highly flexible and provide the foundation on which most other cognition can operate (e.g. memory, decision making, action selection). In the visual domain, attentional selection can be based on spatial location, feature, or object information (e.g. looking for a person with an orange hat standing at a traffic light). Even though a typical selection takes only a few hundred milliseconds, the network in the primate brain that is engaged during this process encompasses a multitude of areas and is distributed across all major lobes as well as subcortical structures including the thalamus, basal ganglia, and the superior colliculus [1–5]. It has been a longstanding goal of the field of cognitive neuroscience to identify mechanisms deployed across this distributed network that account for the efficiency of the selection process. We will refer to these broadly as ‘attentional control mechanisms’. Of note, these are different from control mechanisms related to why an item is selected, which involve motivational, emotional, rule-based, and other cognitive processes that are beyond the scope of this review [6–8].

The first evidence that the attention network contained regions exerting critical control functions came from neuropsychological studies in human patients showing that unilateral lesions, especially of higher-order cortex, may cause impairments in directing attention to contralateral visual space. This syndrome is known as visuospatial hemineglect and, in severe cases, may lead to complete disregard of the affected space, not only with respect to the external space, but also in reference to the body or internal representations of memory systems [9,10]. For example, patients may eat from only half of their plate, dress half of their body, or recall landmarks from ipsi-, but not contra-lesional cognitive maps. Visuospatial neglect may also follow unilateral lesions at different sites including the basal ganglia and thalamus [11,12], thereby underlining the distributed nature of possible control functions across the network. However, emphasis on the most frequently observed sites, that is, the inferior parietal lobule and superior temporal cortex [13,14], have contributed to attention control models that center on fronto-parietal ‘top-down’ control mechanisms that operate by providing feedback to sensory cortices (extensively reviewed in [15]).

In this review, we will discuss the changing views of attention control mechanisms that have emerged in recent years. We will primarily focus on studies using intracranial electrophysiology in non-human primates, but will also make connections to the human literature. We will take the traditional model of attentional top-down control of fronto-parietal over visual cortex during space-based selection as a starting point and introduce three more recent control accounts along with evidence in their support: advances in feature-based control, thalamic control, and control of temporal dynamics in network interactions. We propose that the control mechanisms associated with each account are not mutually exclusive but may rather synergistically interact and complement one another. Together, the view that emerges from these different lines of work is that attention control is distributed across multiple network nodes exhibiting specific control functions over space, objects, and time.

Fronto-parietal control of the selection process

Neuroimaging studies have provided a detailed account of the functional neuroanatomy of the attention network. When human subjects are attending to a location in visual space in anticipation of a target stimulus, activity in frontal and parietal cortex is sustained relative to activity in visual cortex, reflecting the cognitive operations of the task, but not sensory processing (Figure 1a bottom; [16]). Such results are in line with a top-down model of spatial attention control, in which the fronto-parietal network generates feedback signals that influence visual processing in downstream areas. The fronto-parietal activation during directed attention encompasses a multitude of topographic areas that have a preferential representation of the contralateral visual hemifield (Figure 1a top, left; [17]). Accordingly, spatial attention to the contralateral hemifield evokes a stronger signal than to the ipsilateral hemifield, thus generating a contralateral spatial biasing signal in each topographic area (Figure 1a top, right; [18]). The sum of biasing signals across all fronto-parietal areas is similar in magnitude across the two hemispheres suggesting a balanced spatial attention control system in which visual space within a hemifield is largely controlled by the contralateral hemisphere (interhemispheric competition account; [18–20]. The large-scale attention network that has been observed in humans is generally conserved in non-human primates, including key regions in frontal cortex, particularly the frontal eye fields (FEF) and adjacent areas in dorso- and ventrolateral prefrontal cortex (PFC; [21]), as well as in posterior parietal cortex, particularly the lateral intraparietal area (LIP) and adjacent parietal regions such as V6/V6A, and the ventral intraparietal area (VIP) [22–24].

Figure 1. Fronto-parietal top-down control of attention.

a) Top: A spatial attention task activates fronto-parietal regions (color indicates p < 0.001 of contrast: ‘attend to periphery’ vs. ‘attend to fixation’) in humans including FEF and IPS of the contralateral hemisphere (right; adapted from [18]) that largely overlaps with topographic maps of the fronto-parietal network (color wheel maps visual field angle to color overlay; left; [17]). Bottom: Directing attention to a peripheral target (horizontal black lines) location activates FEF and V4 during an expectation period (grey bars) preceding target presentation (red bars). V4’s activity is further increased by the visual stimuli, whereas FEF’s activity reflects the cognitive operation involved with the task (i.e. allocation of attention in space) (adapted from [16]). b) Electrical microstimulation using currents below saccade-inducing thresholds in FEF induces attention-like effects (red) in V4 neurons with receptive fields overlapping the saccade field. Black indicates control data when FEF is not stimulated (adapted from [27]). c) A schematic of the macaque brain, the flow of sensory bottom-up (black arrows) signals, and the trajectory of control signals (red arrows) originating from higher-order cortical areas in the fronto-parietal network that backpropagate to sensory, lower-order cortex.

Important evidence that frontal cortex is a source of top-down feedback signals comes from studies demonstrating that electrical microstimulation of FEF can induce attention-like effects. First, stimulation of FEF can increase the behavioral performance to discriminate a stimulus, similar to the effects of attentional allocation on discriminability [25,26]. And second, FEF stimulation leads to attention-like effects in downstream area V4 such as increasing neuronal firing rates (Figure 1b; [27]). Causal manipulations in humans using transcranial magnetic stimulation have corroborated these findings by showing qualitatively similar effects [28]. Further evidence comes from lesion studies of PFC demonstrating that modulatory effects of attention are attenuated in visual cortex [29]. Together, these and other lines of research have provided compelling evidence for a hierarchical top-down model (c.f. [30]), in which the allocation of attention to locations, features, and objects are controlled by a higher-order fronto-parietal network that influences processing in visual cortex via feedback signals (Figure 1c).

New perspectives on feature-based control

There is extensive evidence of non-spatial effects of attention that enhance neural responses to behaviorally relevant stimulus features (e.g. color or motion; see [31–33]). Feature-based attention operates in parallel across the visual field, such that an attended feature receives a boost independent of its spatial location, supporting a notion that feature- and space-based attention mechanisms may be controlled by partially non-overlapping networks [34–38]. In contrast, based on human neuroimaging findings, the fronto-parietal network is engaged during non-spatial selections of stimulus features [33] or objects [39,40], characterizing it as a ‘domain-general’ controller with little functional specialization. However, the limitations of imaging methods make it unclear whether the underlying control functions associated with space-, feature- and object-based attention are linked to specialized subpopulations of neurons within the network or recruit common circuits. Recent studies in non-human primates have shed light on this issue.

Several studies have probed the role of FEF and surrounding prefrontal areas in the control of feature-based attention. FEF neurons have been shown to exhibit activity that reflects the behavioral relevance of relevant features in a visual search task [41]. However, FEF’s neurons do not exhibit intrinsic feature selectivity. Therefore, FEF may receive feature-relevant information from a source area that represents not only a spatial map, but also incorporates feature information. Neurons in the ventral prearcuate gyrus (VPA), located lateral to the principal sulcus and anterior to the arcuate sulcus (Figure 2a, left, inset), were found to be feature selective and were modulated by spatial attention [42]. VPA exhibits sustained feature-selective responses that are obtained at earlier response latencies than in FEF (Figure 2a, left). Inactivation studies of VPA showed that feature- (but not space-) based attention effects were abolished in FEF (Figure 2a, right), and behavioral performance in a feature-based attention task was impaired [42]. These findings suggest that VPA is a source for top-down feedback signals related to feature-based attention control that may propagate to other higher-order areas such as FEF [42] as well as visual cortex [43] to boost the processing of relevant stimuli across the visual field.

Figure 2. Advances in feature-based attention control.

a) Left: During a visual search task in macaques, VPA (location schematically presented in inset) exhibits feature-based attention effects earlier than those observed in FEF. Right: Feature-based effects (red) observed in FEF (top) are abolished following inactivation of VPA (bottom). Spatial attention effects (green) are not affected following inactivation (adapted from [42]). b) PITd is strongly activated during a motion discrimination task (top; adapted from [46]) and its neurons exhibit considerable attentional modulation (bottom; adapted from [47]). c) Top: Inactivation of SC induces a decrease in activation in an area at the floor of the STS during a motion detection task. Bottom: fSTS exhibits attention modulation in its neuronal responses (left). This modulation is reduced, but not abolished following inactivation of SC (right; adapted from [48,49]). d) New nodes discovered may factor into the control of feature-based attention. VPA may generate a behaviorally-relevant template for object and feature information to FEF (solid red arrow) and to V4 (dotted arrow). Relevant anatomical connectivity is denoted in red, highlighting projections between the PITd/fSTS area (abbreviated PITd*) with LIP, FEF, and VPA (red lines). The SC causally influences temporal cortex in the PITd/fSTS area (dashed red line) most likely by way of pulvinar (black lines).

Feature-based attention control is not confined to PFC. A temporal area in dorsal posterior inferotemporal cortex (PITd), located on the inferior bank and lip of the superior temporal sulcus (STS), anterior to MT (Figure 2b, top) has recently been implicated in attention control function. PITd is strongly connected with prefrontal areas VPA, FEF, as well as LIP (Figure 2d, [44,45]). This area has been found to be activated in monkeys performing a motion detection attention task in a neuroimaging study [46], and was targeted for recording studies that revealed strong attention modulation (Figure 2b, bottom). Further, microstimulation of PITd causally shifted the location at which attention was allocated [47]. While PITd appears to be responsive to visual stimuli of different feature categories, neurons do not exhibit specific feature selectivity [47].

PITd is not only connected with fronto-parietal cortex, but may also receive modulatory input from the superior colliculus. In neuroimaging studies, inactivation of the superior colliculus (SC) during a motion discrimination attention task [48] has been shown to reduce activation in the vicinity of PITd, a region termed fSTS due to its location at the floor of STS (Figure 2c, top; Bogadhi et al., 2019). Recording studies showed a reduction (but not abolishment) of attentional modulation effects as a consequence of SC inactivation (Figure 2c, bottom; Bogadhi et al., 2021). While it is not clear whether fSTS and PITd are overlapping or separate but adjacent areas, inactivation of this part of cortex underlines its role in attention control by inducing behavioral deficits akin to visuo-spatial hemineglect [48]; however, exact pathways affected by perturbing this area remain unexplored and may be best probed with systematic causal manipulation studies of simultaneous multi-area recordings. Neurons in this area were also responsive to different feature categories as well as objects. Although the exact relationship between PITd and fSTS remains to be determined, these studies show that this part of cortex exhibits functional characteristics of an attention control area that may be specialized in feature- and object-based attention [46,48], and highlights afferent signals from SC that may be routed to this temporal area through pulvinar (Figure 2d). The anatomical projections between this temporal region and VPA (Figure 2d, [45]) may also enable the flow of feature-based control signals between PFC and the feature representative areas close to PITd/fSTS.

Coordination of large-scale networks: Thalamic control

How can a large-scale network that consists of a multitude of nodes process information from the sensory periphery, integrate it with internal representations, and drive motor output to generate behavior within just a few hundred milliseconds? There is evidence that a central thalamic structure interconnected heavily with the cortex, the pulvinar, may play an important role in organizing the cortical attention network in time and aiding its processing efficiency [3,50]. Pulvinar is the largest nucleus in the primate thalamus and its expansion during evolution scales with that of the primate neocortex [51–53]. It is almost exclusively interconnected with the cortical visual system through two well-established cortico-pulvinar pathways that parallel the canonical input-output relationships constituting the visual processing hierarchy [54,55]: (i) a transthalamic feedforward pathway that originates from layer 5 of a cortical area, loops through a dedicated pulvinar projection zone and terminates in layer 4 of an upstream area, and (ii) a feedback pathway that originates from layer 6 of a cortical area and projects to its pulvinar projection zone. An important characteristic of the transthalamic feedforward pathway is that it indirectly connects two cortical areas that share a direct cortico-cortical projection [56,57]. Based on this anatomical connectivity, pulvinar is in an ideal position to coordinate functional interactions across a cortical large-scale network [58]. This idea was probed in studies that recorded simultaneously from two directly interconnected cortical areas and their projection zone in the pulvinar, while monkeys performed a spatial attention task [58,59]. Focusing on the cue-target interval, which represents a pure cognitive state (i.e. the allocation of attention at a cued location in expectation of target onset), during which the attention network is prepared for target selection, neuronal spiking activity was not only enhanced, but also showed increased correlations in a common frequency band (alpha, low beta, <20 Hz) of the local field potential. The pulvinar appeared to have the strongest influence on this coordinated activity in the cortical areas, whereas the cortical areas had little influence on these interactions (Figure 3a; [58]). These functional thalamo-cortical interactions have been shown for extrastriate visual cortex [58], fronto-parietal cortex [59], and interactions between higher-order and extrastriate cortex [60], and may thus constitute a general principle of functional thalamo-cortical organization during attention. These results are consistent with a role for pulvinar as ‘time-keeper’, coordinating and optimizing functional interactions across cortical networks to enhance the efficiency of signal processing between nodes.

Figure 3. Thalamic control and temporal dynamics of attention.

a) Granger causality analysis shows pulvinar influences on cortical extrastriate areas V4 and TEO (top) during an attention task when attention is directed at the RF (red) vs. away from the RF (blue). However, Granger causality analysis shows no direct influences between the two extrastriate areas (bottom; adapted from [58]), suggesting that pulvinar initially organizes the cortical attention network in time. b) Left: Detection behavior (as represented by hit-rate [HR] in a modified Egly-Driver task) is predicted by the phase of neural theta rhythms in three major nodes involved in attention control: LIP, FEF, and pulvinar. Using a sine fit (black), HR is higher in one phase of theta (termed ‘good theta’) than in the opposing phase (termed ‘poor theta’; adapted from [59,73]). Right: Functional connectivity between nodes in the attention network is modulated by neural theta phase. Spike-LFP coupling between pulvinar and LIP in monkeys has two broad states during the cue-target delay interval of an attention task. In the ‘good theta’ phase, pulvinar spikes are coupled to LIP LFP in alpha frequencies (top), while LIP spikes are coupled to pulvinar LFP during the ‘poor theta’ (bottom; adapted from [59]). c) Schematic of connectivity between pulvinar and cortical regions involved in attention function. For cortico-cortical connections (dotted black lines) there exists an indirect route through pulvinar (red lines). The directionality of the connectivity evolves dynamically over time, organized by a theta rhythm; however, interactions beyond LIP, mediodorsal pulvinar, and FEF (panel B) during this rhythmic reweighing remain unexplored. Further, whether similar temporal dynamics apply to areas engaged in feature-based attention is an open question (Box 1).

Pulvinar inactivation studies have corroborated these ideas by showing that, as a consequence, functional interactions between cortical areas are weakened [61,62], demonstrating that cortico-cortical information transmission is greatly compromised without pulvinar’s influences. Further, during pulvinar inactivation, neuronal activity in visual cortex is greatly attenuated [62,63], indicating a loss of responsiveness to visual input. Thus, it is possible that pulvinar controls the excitability of neurons to respond to visual information from downstream areas, thereby playing a role of an ‘enabler’ of cortical function [64].

Attention and the dimension of time: Control of temporal dynamics

According to classical attention theories [15,65–67] and our subjective experience, neural processing appears to be continuous during attentional deployment. However, recent studies have shown that attention function unfolds over time and is characterized by alternating periods of relatively enhanced or diminished visual processing at the attended location, which are associated with respective better or worse behavioral performance in detecting visual targets. Behavioral studies have shown that these alternations occur at a rhythm in the theta range (~4–6 Hz; [68–71]). Importantly, these behavioral rhythms have been found both in humans and monkeys, thus constituting an evolutionary conserved and possibly fundamental property of attention [59,68,72].

The behavioral rhythms have been linked to neuronal theta rhythms in fronto-parietal cortex and the thalamus. The phase of the neuronal theta rhythm in relation to target appearance in a visual detection task is predictive of behavioral outcome on a trial-by-trial basis in both humans and monkeys, thereby linking behavior and network-level neuronal signals (Figure 3b; [72,73]). The theta rhythm across the fronto-parietal-thalamic network appears to present a mechanism that temporally coordinates potentially conflicting sensory and motor functions within the network. Specifically, the network contains both neuronal populations for visual processing that gets boosted during attentional deployment and for shifting attention or executing eye movements to a new target of interest [74–78]. Sampling and shifting cannot occur at the same time and require temporal coordination of the associated neuronal populations and the wider circuits they are embedded in. Theta-dependent attentional rhythms appear to set up two alternating states – for ‘sampling’ and ‘shifting’ – that are characterized by complex temporal dynamics across fronto-parietal cortex and the thalamus.

The ‘sampling’ state emphasizes visual processing (associated with better behavioral performance) and is characterized by greater synchronization of gamma (>35 Hz) and beta (~15–30 Hz) activity within and between frontal and parietal cortex [73]. Specifically, spiking activity of visual neurons is exclusively correlated with gamma activity, which has often been observed during enhanced visual processing [79], while spiking activity of oculomotor neurons is exclusively correlated with beta activity, which has been linked to the suppression of motor actions [80,81]. The pulvinar coordinates these cortico-cortical dynamics across the attention network only during the sampling state in similar ways as described in the previous section, thereby underlining its proposed function in sensory gating (Figure 3b; [59]). The ‘shifting’ state (associated with relatively worse performance) is characterized by a release of suppression in the motor system observed as decreased motor-suppressive beta oscillations in LIP and FEF [73], which may also affect SC. This is accompanied by greater synchronization of alpha activity between parietal cortex and pulvinar (Figure 3b; [59]), which is hypothesized to temporarily interrupt the transthalamic gating function, thereby leading to a relative suppression of visual processing. The latter state provides windows of opportunities to disengage from the presently attended location and to direct attention to a new location or event in the environment. The rhythmic alternations of sampling and shifting states during attention deployment provide critical flexibility, not only to prioritize visual processing at an attended location, but also to provide opportunities to re-allocate attentional resources, if behaviorally desirable, without locking them into a particular state for extended periods of time [82]. Thus, even when attention is sustained at a spatial location, the underlying network interactions are highly dynamic, providing a neural model for the flexibility of a cognitive function (see [83] for a circuit model).

Conclusions: Attentional control 2.0

Taken together, findings from the recent literature introduce new functional nodes and mechanisms that are implicated in attention control beyond a fronto-parietal, top-down framework. Substantial evidence points to pulvinar as a critical node that is involved in the active propagation of attention control signals. Putative temporal control areas discussed in this review [84,85], as well as SC [86,87] may communicate with fronto-parietal nodes using pathways involving pulvinar. This places pulvinar as a potential hub for control signals to route through efficiently, either as a means to coordinate processing between control areas or to propagate from control nodes to sensory areas that exhibit attentional modulation.

Increasing evidence indicates a dynamic unfolding of attention function and its exerted control over time. The large-scale network with thalamus at its center may dynamically alter functional connectivities across nodes at timescales similar to those obtained in behavioral studies, setting up alternating attentional states that are characterized by complex temporal network dynamics. We propose that these shifting functional connectivities enable behavioral strategies for efficient exploration of the environment as well as provide a means to alternate between enhanced sensory processing and motor-engaged shifts of attention. Specifically, we propose that these dynamic interactions, which are coordinated by theta rhythms, not only influence the fronto-parietal attention network, but may also modulate the putative temporal control areas, as well as putative orienting signals originating in SC (see Box 1).

Box 1. Open Questions.

• Do thalamocortical interactions support and control other cognitive functions? Higher-order thalamus, and particularly pulvinar, has been shown to contribute to cognitive control. It is an open question whether thalamocortical circuits support other cognitive functions in the primate brain, as suggested by recent studies in rodents [88]. For example, working memory recruits overlapping mechanisms in PFC [89]. It will be an exciting venue for future research to investigate whether similar thalamocortical influences, possibly exerted by other nuclei such as the mediodorsal nucleus, would subserve the control of working memory in a manner similar to what has been found in attention function.

• What is the role of temporal putative control areas in the large-scale primate attention network? The causal effects of PITd and fSTS perturbation on behavior and the puzzling functional similarity to fronto-parietal control areas make temporal cortex a fruitful target for future investigation using tasks that dissociate feature- from space-based attention effects, from a network perspective to probe its interactions with VPA, FEF, and LIP, and within the framework of rhythmic attention.

• Do the temporal dynamics of attention factor into the control of feature-based or object-based attention? While recent work suggests that feature-based attention has rhythmic properties similar to spatial attention [90], it remains to be examined how that is reflected in underlying neural activity specific to feature-based effects. That is particularly of interest in PITd/fSTS and VPA as control areas that exhibit effects of both feature- and space-based attention.

• What is the role of the signals that pass from SC to temporal cortex in attention function? The causal influence of SC on the PITd/fSTS region suggests the existence of SC afferents, likely by way of pulvinar, that modulate neuronal signals in temporal cortex. However, the functional role of these signals is not clear, as SC inactivation only reduces attention effects in fSTS, but does not abolish them. The ability to detect novel events is an orienting function with localized cortical regions in humans [91–93] but not in monkeys [94]. One hypothesis is that detection of novel events may be more closely linked to SC in monkeys. SC afferents in fSTS may carry coarse signals that inform about the behavioral relevance of objects across the visual field that then propagate from temporal cortex to other control areas either directly or through pulvino-cortical projections.

Although future studies need to elucidate the role of the reported temporal control areas and how they relate to prefrontal feature-based attention signals originating in VPA (Box 1), current evidence indicates that neurons in PITd/fSTS exhibit feature- and space-based attention effects as well as object selectivity [47,49]. The finding of spatial and non-spatial effects in that region and its location in the visual processing hierarchy along the object processing ventral visual stream suggest that PITd/fSTS may act as a control node that integrates sensory information relating to the basic form, color, and motion and may be involved in the control of object-based attention.

In summary, we propose that the non-canonical nodes of attention control discussed in this review are integral parts of one conjoined large-scale network, or constitute subdivisions of smaller subnetworks that function in tandem in a dynamic fashion, as attention function unfolds over time.