Common and Distinct Neural Mechanisms of Attention

Abstract

Despite a constant deluge of sensory stimulation, only a fraction of it is used to guide behavior. This selective processing is generally referred to as attention, and much research has focused on the neural mechanisms controlling it. Recently, research has broadened to include more ways by which different species selectively process sensory information, whether due to the sensory input itself or to different behavioral and brain states. This work has produced a complex and disjointed body of evidence across different species and forms of attention. However, it has also provided opportunities to better understand the breadth of attentional mechanisms. We summarize the evidence which suggests that different forms of selective processing are supported by mechanisms both common and distinct.

Different forms of selective processing

The senses are constantly bombarded with stimulation. Although early stages of sensory processing extract a wealth of information from the environment, only a small fraction of that information is used to guide behavior at any given moment. Instead, the sensory input that is perceived, remembered, decided about, and acted upon is largely, and conveniently, confined to contextually relevant stimuli. This selective processing is generally referred to as attention, a basic cognitive function that interfaces with all other high-level functions of the nervous system, from memory to motor control.

In testing hypotheses about the neural circuitry of selective attention, as with hypotheses about the circuit basis of any class of behaviors, it is crucial to rigorously consider the precise form or forms of attention employed. Sensory stimuli can be preferentially detected and/or discriminated in a variety of behavioral contexts in human subjects [1,2], and this is likely so in most model organisms. In this review, we highlight two widely studied types of attention, goal-driven and stimulus-driven, but there are other important types, e.g., cross-modal attention (selection of stimuli across sensory modalities) [3,4], and other dichotomies, e.g., spatial versus feature and object-based attention [2]. Each of these types seems likely to have different and dissociable neural circuit components, if only because the behavior differs qualitatively between each. Moreover, different species likely vary considerably in their engagement in these different types of attention.

To date, the types of tasks employed to probe attention mechanisms have varied significantly across different laboratories (e.g. behavioral response, type of cues), even when the same model organism and type of attention is studied [5–8], and indeed even within human studies [9,10]. These differences seem likely to yield discrepancies in the conclusions drawn about underlying circuit mechanisms (Box 1). In this review, we briefly survey a broad swath of recent studies aimed at identifying mechanisms of attention across model species, including humans, and across different types of attention, with an eye toward identifying neural mechanisms both common and distinct. We highlight evidence from nonhuman primates as it provides the best mechanistic bridge between evidence from human and other animal models. We begin by discussing new evidence on the mechanisms of goal-driven and stimulus-driven attention. Next, we summarize recent insights into the relationship between mechanisms of attention and arousal. Lastly, we consider evidence of the neural oscillations that reflect attentional control.

Box 1. Choosing valid behavioral paradigms.

There are myriad technical challenges faced in identifying the neural mechanisms of particular behaviors. Unsurprisingly, among them is the challenge of designing valid behavioral paradigms, where valid refers either to congruence with natural behavioral phenotypes of a given organism or congruence with a particular human behavior. The choice of behavioral paradigms naturally plays a pivotal role in the conclusions that can be drawn from any given experiment. Consider, as an example, the early conclusions that the SC and the frontal eye field (FEF) were solely involved in oculomotor control in primates, and not involved in covert spatial attention [135,136]. That conclusion followed reasonably from the observation that whereas visually driven activity in both areas is enhanced in advance of eye movements directed to neuronal receptive fields (RFs), no enhancement is observed when monkeys covertly attended to RF stimuli. Crucially however, the covert task employed in those early studies required monkeys to divide attention between a peripheral RF stimulus and the central fixation point, as either could change luminance with equal probability. In stark contrast, during the same (divided attention) task, neurons in parietal cortex exhibited clear enhancement [136]. By now, it is well established that the activity of neurons in all three structures is strongly modulated when monkeys perform tasks in which they covertly attend to a single one of multiple visual stimuli [5,88,137]. Thus, it seems that the discrepant results between parietal cortex and the SC and the FEF might simply reflect differences in the spatial precision and competitive interactions within the visuotopic maps of the latter and former. Thus, the choice of the behavioral paradigm is clearly consequential in interpreting the role of putative underlying mechanisms.

Goal-driven vs. Stimulus-driven attention

The tendency of organisms to selectively process only a subset of sensory input appears to come in a variety of forms. Among these, goal-driven and stimulus-driven attention represent two forms of attention that have been characterized extensively in human psychophysical studies [1,11]. Goal-driven attention describes selective sensory processing due to an endogenously generated signal (e.g., representation of a rule, strategy, or motivational state). By contrast, stimulus-driven attention describes selective processing based solely on the physical properties of stimuli. Stimulus-driven and goal-driven attention have been widely assumed to be dichotomous, and indeed seem to involve distinct underlying neural mechanisms [12]. In humans, two separate frontoparietal networks are thought to contribute differently to goal-driven and stimulus-driven attention, specifically a dorsal network to the former and a ventral network to the latter [13,14]. Nevertheless, these two types of attention may not be sufficient to account for all instances of stimulus selection. For example, some human psychophysical studies suggest that selection history contributes independently of task goals and stimulus salience [15,16], adding a third potential factor in attentional selection. Others have suggested that these types can be broken down even further [17]. This added complexity places even greater importance on the careful design of behavioral paradigms used in future studies aimed at identifying underlying neural mechanisms of attentional selection, particularly in animal models. Nonetheless, these two forms of attention provide a good starting point in the effort to identify common and distinct neural mechanisms of attentional selection.

Mechanisms of goal-driven attention

Insights into the mechanisms of goal-driven attention have been noted for more than a century, or at least since early neurological descriptions of attentional disorders resulting from brain damage in humans [18] and in animals [19]. Furthermore, sensorimotor and premotor networks have been implicated in the control of goal-driven, spatial attention for nearly as long, gaze control networks being a prime example ([20] for review). In humans, the preparation of gaze movements necessarily results in perceptual benefits at targeted locations [21,22], and those perceptual benefits signify the defining effect of attentional deployment. Additional human psychophysical evidence indicates that similar perceptual benefits accompany planned hand movements [23,24], suggesting that multiple effector systems may independently influence sensory processing as a means of adaptively selecting relevant stimuli during sensory-guided behavior. By now, it is clear that the preparation of eye movements is sufficient to alter activity in the visual systems of human [25] and nonhuman primates [26–28]. Moreover, the influence of eye movement preparation on visual activity is robust even when dissociated from covert attention [29], consistent with human psychophysical data [30]. Correspondingly, there is now considerable evidence of a causal role of gaze control networks in the control of goal-directed, spatial attention, having first been shown in nonhuman primates [31,32], and subsequently in humans [33,34], birds [35] and mice [36]. Collectively, this evidence demonstrates that circuits involved in the preparation of eye and head movements also exert a direct influence on the visual system and visual performance, indicating that these circuits represent both intended and attended goals.

In recent years, more has been learned about the specific mechanisms by which gaze control circuits contribute to goal-driven attention. In the primate brain, gaze control is achieved principally via the SC, the lateral intraparietal area (LIP), and the frontal eye field (FEF), three sensorimotor structures that serve as interfaces between the visual and saccadic systems [37–39]. In the human brain, the FEF and parietal cortex comprise much of what is referred to as the dorsal frontoparietal, attention network [40]. In the macaque, both the SC and the FEF have been implicated in the control of visual spatial attention [32,41]. However, the attention-related modulation of visual cortical activity observed during goal-directed attention does not depend on the SC [6], in contrast to a clear role of the FEF [7,28,42,43]. The influence of FEF neurons on visual cortical activity appears to be achieved, at least in part, by direct retinotopic projections of FEF neurons to extrastriate visual cortex [44,45]. Similar to neurons within dorsolateral prefrontal (dlPFC) and parietal cortex, many FEF neurons signal the location of remembered visual targets via spatially tuned persistent spiking activity, which is proposed to emerge from local and long-range recurrent connectivity [46,47]. To date, the function of FEF persistent activity either in motor preparation or spatial working memory, or both, remains uncertain (although see [48]). One study identified visual cortex-projecting FEF neurons by antidromic stimulation and found that whereas only half exhibited visual properties and virtually none exhibited motor properties, all of them exhibited persistent activity [49]. These neurons have also been shown to preferentially express dopamine D1 receptors compared to D2 receptors, and compared to locally projecting neurons [50], which is consistent both with the unique role of D1 receptors in memory-related persistent activity [51] and their involvement in the gating of visual cortical activity by the FEF [43]. The fact that persistently active FEF neurons preferentially project to visual cortex suggests that they play a distinct role in the selection of visual signals. Indeed, within the FEF, memory-related persistent activity is accompanied by an increased efficacy of visual cortical inputs to FEF neurons within the same local network [52]. Combined, this evidence points to a possible circuit within primate prefrontal cortex by which gaze control networks select visual information during goal-directed, spatial attention (Fig. 1A).

Figure 1. Neural circuitry of goal-driven attention.

A. Connectivity between visual and prefrontal cortex in the macaque. Diagram summarizes several lines of evidence. Connectivity between dorsolateral prefrontal cortex and posterior visual areas exists largely between the FEF and retinotopic areas in extrastriate cortex (e.g., V3, V4, MT). Projections to extrastriate cortex from the FEF originate largely from layer 2/3 [133]; projection neurons (peach) disproportionately exhibit delay-period activity during memory-guided saccade tasks [49]. Inputs to visual cortex from the FEF terminate predominantly, and disproportionately, onto the distal dendritic spines of pyramidal neurons across layers [44]. Visual-cortex projecting FEF neurons, along with other long-range projecting pyramidal neurons, express dopamine D1 receptors (red dots) at a much higher rate than inhibitory interneurons [50]. Input to the FEF from visual cortex targets neurons exhibiting both visual and motor properties, and those inputs are facilitated when monkeys remember locations matching the location of visual input [52]. B. Summary of prefrontal-visual cortical circuitry in the mouse (adapted from [8]). Prefrontal projections principally target VIP+ inhibitory interneurons, which in turn inhibitory somatostatin+ inhibitory interneurons thereby disinhibiting excitatory neurons. PV, parvalbum+ inhibitory interneurons.

In addition to the evidence implicating gaze control networks in goal-directed spatial attention, newer evidence suggests an additional source of attentional control in the primate brain. Using fMRI, several different investigators have independently identified a region in temporal cortex within the floor of the superior temporal sulcus (STS) of the macaque brain that is remarkably sensitive to attentional deployment [53–55]. This region of the STS sits near a confluence of motion-selective (MT, MST) and shape-selective (TEO) visual areas, yet itself appears to exhibit only modest stimulus selectivity [54,55]. However, the activity of neurons within this area is modulated approximately fourfold on average when attention is directed into the neuronal receptive field compared to outside [55]. Furthermore, electrical microstimulation of sites within this area significantly influences attentional deployment in spatial specific manner [55]. Both visual and attention-related activity in this area are dependent upon input from the SC [56] and the FEF [57], and this area is directly connected with the areas implicated in attentional control within parietal and prefrontal cortex [58]. This region of temporal cortex is also strongly activated during different types of visually guided eye movements, specifically vergence, smooth pursuit, and saccadic eye movements [59]. Nonetheless, its unique role in attention and visually guided behavior remains an open question. Notably, the location of this area within temporal cortex raises the possibility of homology with areas of the human ventral attention network [57,60,61].

Given the relative limitations in the tools available for dissecting neural circuits of attentional control in primates, attempts to do so in more genetically tractable animal models, particularly mice, have increased dramatically in recent years (see [8] for review). As would be expected given the obvious behavioral and cognitive limitations of the mouse model, the development of suitable tasks in mice to model selective attention in humans has posed a significant challenge for investigators [36,62]. Indeed, one might argue that focusing on mechanisms of more species-specific forms of goal-directed attention in the mouse, e.g., during locomotion, may be more fruitful (Box 2). Nonetheless, more recent attempts have employed behavioral tasks that seem remarkably analogous to goal-directed attention used with human and nonhuman primate subjects [63–65]. During such tasks, as in primates, improved psychophysical performance at attended locations is accompanied by spatially selective enhancements in visual cortical activity [64]. In addition, as in primates, prefrontal cortex directly influences the activity and sensitivity of neurons in visual cortex [36,66].

Box 2. Locomotion and Visual Cortical Enhancement in Rodents.

In primate species, it may not be terribly surprising that spatial attention arises in part from direct modulation of visual activity by gaze control neurons given the importance of gaze control for visually guided behavior, largely due to the need to continuously reposition the fovea for high acuity vision. However, this instance of a clear influence of motor control signals on sensory processing, one that extends even to non-foveate, non-visually dominate species [35,36], suggests that other instances of motor-driven sensory modulation might be viewed in the context of visual attention. A prime example of this is the robust influence that locomotion exerts on visual cortical activity in mice. First described by Neill and Stryker [138], locomotor activity (running) in mice is associated with a dramatic increase in the visual responsiveness of area V1 neurons, without concurrent changes in spontaneous firing rate. In many ways, the effects of locomotion on neuronal activity resembles the effects of attention described previously in the primate visual system [139], including increased stimulus-driven spike rates [138,140], reduced spike count correlations [141], and spiking variability [141,142], and changes in gamma local field potential power [138]. Notably, visual detection is enhanced during locomotion, compared to when mice are stationary [142]. Moreover, both the behavioral and visual cortical effects depend on the type of visual task mice are engaged in during locomotion [143]. Thus it may be reasonable to consider the neural and behavioral effects of locomotion as a form of visual attention. As another similarity with attention, the cortical circuitry involved in the locomotor effects on visual cortical activity appears to involve a similar disinhibitory circuit that has also been identified in the mouse in studies employing standard attention tasks [64,144]. Lastly, although it remains unclear to what extent the influence of locomotion on V1 activity includes changes in global arousal and the associated effects of arousal [97,145], the observed effects nonetheless involve input from neurons specifically involved in locomotor control. Optogenetic stimulation of the mesencephalic locomotor (MLR) region in awake mice induces both locomotion and increases in the gain of V1 responses [146]. Notably, MLR stimulation below the threshold needed to evoke overt movements exerts similar changes in V1 activity. Thus, similar to the effects of FEF stimulation in primates [28], the effects of activating the MLR in mice are dissociable from motor output.

But more importantly, studies in the mouse have revealed considerably more cellular and synaptic detail to circuits implicated in attention. For example, projections from prefrontal to visual cortex appear to influence visual activity via local disinhibitory circuits in which vasoactive intestinal peptide (VIP) local inhibitory interneurons are directly targeted by prefrontal cortical projections [64] (Fig. 1B) VIP neurons in turn inhibit somatostatin (SST) inhibitory interneurons, resulting in disinhibition of excitatory neuronal activity. These two classes of interneurons may roughly correspond, respectively, to calretinin+ and calbindin+ interneurons [67–69], though there is yet evidence of a similar disinhibitory circuit motif in prefrontal-visual connections in the primate brain. Instead, thus far pyramidal neurons seem to be disproportionately targeted by prefrontal projections [44]. In addition, in the mouse, a subset of frontal cortical neurons exerts modulatory effects on visual cortex indirectly via projections to the superior colliculus and the pulvinar [70], thus raising the possibility of a similar circuit in primates.

Combined, the above evidence not only suggests the existence of similar mechanisms of attentional control across phylogenetically distant species but holds the promise that such mechanisms can be studied in more genetically tractable animal models. Nevertheless, regarding both points, the homology between neural circuit mechanisms of attentional control in rodents and primates remains very much in question. For example, the frontal cortical area in mice implicated in attentional control, specifically anterior cingulate cortex [36,66], is likely not anatomically homologous with dorsolateral prefrontal cortex in primates, and instead may be homologous with primate medial prefrontal areas [71]. Notably, however, medial prefrontal cortex in primates contains a supplementary eye field (SEF) that is heavily interconnected with the SC, the FEF and parietal cortex [72]. Primate SEF is also connected with areas within posterior visual cortex [73] and is activated during attention tasks [74]. Yet the extent to which neurons in the SEF directly contribute to goal-directed attention in primates remains to be examined.

Mechanisms of stimulus-driven attention

Unlike the extensive evidence demonstrating the selective modulation of sensory representations throughout the brain during goal-driven attention [40,54,70,74], details about the effects of stimulus-driven attention are notably less clear [20]. By definition, a salient stimulus draws attention, and the salience of a stimulus can be conferred in several ways, including simply by virtue of greater physical energy, e.g., brighter, larger, or moving objects [1,75] or by ethological relevance, e.g., face images, or looming objects [76]. Presumably, this form of salience can be detected already at very early stages of sensory processing in a winner-take-all fashion, such as in the midbrain [77,78]. Physical salience can also be conferred by the uniqueness of a stimulus relative to all others. In fact, models of stimulus-driven attention have been developed largely from human psychophysical studies investigating the influence of unique visual features on the allocation of attention [79–82]. Consistent with observations in human psychophysics[11,20,83], these models suggest that the neural mechanisms of stimulus-driven attention may be separable at some level from those controlling goal-driven attention. A key aspect of these models is the proposition that the visual system analyzes different feature dimensions (e.g., color) more or less in parallel, to compute local feature contrasts. The resulting feature contrasts can then be combined to form salience maps in which unique objects can be localized in space, regardless of the features that define them [80,84]. Neurophysiological studies in nonhuman primates have provided at least some evidence that this feature-independent global salience is computed, or at least amplified, within the posterior parietal cortex (PPC) [12,84–87], prefrontal cortex (PFC) [12,32,87,88], and the SC [78,89,90]. However, the role of each candidate area has not been causally examined, and the relative contributions of these cortical and subcortical regions to salience computations remain largely unknown.

A recent study [91] provides the first causal evidence of a role of PPC in computing visual salience. In the study, the authors reversibly inactivated PPC in behaving monkeys and measured the effects of inactivation on salience representations downstream of PPC, in prefrontal cortex, specifically the FEF (Fig. 2). PPC inactivation selectively reduced the coding of visual salience by neurons in the FEF. Like neurons in visual cortex, FEF neurons exhibit relatively discrete classical receptive fields and thus can signal the location of isolated visual stimuli (Fig. 2A, top). In addition, despite a general lack of stimulus selectivity [92], FEF neurons signal the location of stimuli with unique features (Fig. 2B, top). Inactivation of the PPC however, drastically reduces this effect. Specifically, whereas FEF neuronal responses to single isolated visual stimuli were largely unchanged during PPC inactivation (Fig. 2A, bottom), responses to unique stimuli, those with distinctive visual features, presented among competing stimuli, were diminished (Fig. 2B, bottom).

Figure 2. Contribution of PPC to visual salience representations and salience-driven behavior.

A. Classical receptive fields (RFs), mapped with a single-colored stimulus presented in isolation, of three example FEF neurons during control (top) and PPC inactivation (bottom). Following PPC inactivation, the classical RFs remained largely unchanged. B. Salience-derived receptive fields, mapped with a unique stimulus within an array of other stimuli that were differently colored, of the same example neurons during control (top) and PPC inactivation (bottom). Following PPC inactivation, in contrast to A, the salience-derived RFs were dramatically diminished. C. Example image from a free-viewing task (left) and the corresponding salience map within that image (right) [80]. D. Correspondence between the computed salience and fixations made in the example image before (left) and after (right) PPC inactivation. Top-row fixations are labeled in eye-centered coordinates as contralaterally (triangles) or ipsilaterally (circles) directed saccadic eye movements. The bottom row shows the same fixations labeled in head-centered coordinates as landing in the contralateral or ipsilateral half of the image. Figures adapted from [91].

In addition, to test for corresponding changes in salience-driven behavior, the authors measured the relationship between the salience contained within naturalistic images (Fig. 2C) and the pattern of fixations of those images. Following PPC inactivation, monkeys exhibited a reduced influence of visual salience on the pattern of fixations within the contralateral visual space, when measured in head-centered or in eye-centered coordinates (Fig. 2D). These results indicate that in the monkey neural activity within PPC contributes both to the emergence of salience signals in prefrontal cortex and the influence of salience on behavior. However, it remains unclear how the results in the monkey map onto human dorsal and ventral attention networks, where the latter, not the former is proposed to underlie salience-driven attention. Nevertheless, some studies in humans have reported evidence that the dorsal network areas (e.g. intraparietal cortex) are also engaged during stimulus-driven attention [93–95]. An important goal of future studies in nonhuman primates and other animal models might be to examine the relative roles of other structures, including temporal cortical areas and the SC, in the control of stimulus-driven salience and salience-driven behavior.

Arousal, cortical states, and attention

Neural circuits controlling selective attention have generally been considered independent from mechanisms regulating global arousal. Yet, attention and arousal have similar effects on neural population activity in sensory cortical areas [96]. Both processes involve similar modulations of tuning curves, spike-count correlations, and low-frequency oscillatory activity in local field potentials (LFPs) [97,98], and they appear mediated by common neurotransmitter systems [99,100]. Moreover, optogenetic activation of basal forebrain—the primary source of arousal-promoting cholinergic input to the cortex—produces perceptual benefits and neural activity modulations typical of selective attention [101]. These similarities suggest that global arousal and selective attention may tap into similar underlying circuitries, and recent work has provided more support for this idea.

The link between attention and arousal transpires within the synchronized neural activity fluctuations characteristic of cortical state. Slow synchronous transitions between periods of high activity and quiescence define cortical state at low arousal, such as during slow wave sleep or anesthesia [96,102]. In the awake brain, slow fluctuations are less conspicuous and cortical state appears desynchronized in macroscopic measurements (e.g. electroencephalogram). However, even during active task engagement, local neural populations spontaneously transition between episodes of high and low activity, akin to synchronous transitions during sleep but on a much smaller spatial scale [103,104]. These local synchronized fluctuations are evident as spontaneous waves propagating laterally across nearby cortical columns, as observed within marmoset visual cortex (Fig. 3A) [105,106]. The local waves have irregular timing and varying degree of spatial organization, which together predict the magnitude of stimulus-evoked responses and the animal’s ability to detect a stimulus [105]. In single columns, synchronized fluctuations appear as transitions between phases of vigorous (On) and faint (Off) spiking across all cortical layers (Fig. 3A) [98,107–109]. The On-Off phase of population activity also predicts the animal’s ability to detect stimulus changes in the attended location. The On-Off dynamics in single columns and lateral waves are likely related, with columns switching to the On phase when hit by a passing wave. However, a direct relationship between On-Off dynamics and lateral LFP waves has not been demonstrated yet.

Figure 3. Coordination of local cortical states during selective attention.

A. Cortical activity fluctuations are organized along anatomical dimensions. Lateral recordings from nearby columns in marmoset visual cortex (e.g., using rectangular Utah arrays) reveal spontaneous waves in local field potentials (LFPs) that travel horizontally across the cortex (top, adapted from [105]). Laminar recordings from single columns (e.g., using linear multielectrode arrays) reveal spontaneous transitions between phases of vigorous (On) and faint (Off) spiking that occur synchronously across cortical layers (lower right, adapted from [107]). The population spiking activity is segmented into On (yellow) and Off (blue) phases with a Hidden Markov Model (HMM). B. Left: Cross-correlation between sequences of On and Off phases of activity in columns of macaque areas V1 and V4 consisting of neurons with overlapping receptive fields (RFs) shows that On and Off phases in V1 and V4 tend to align (dashed gray line - shuffled data). The cross-correlation asymmetry indicates that On-Off transitions in V4 tend to precede V1 transitions during attention. Right: The correlation strengths (area under the curve) decrease with the receptive field separation. Thus, the alignment of On-Off dynamics between visual areas is coupled to their retinotopic alignment. C. Left: Simultaneous spiking activity in V1 (bottom) and V4 (top) segmented with HMM (example trial of a spatial attention task) illustrates four possible configurations of On-Off phases in two areas (circles with arrows). Right: Alignment of local On-Off phases between V1 and V4 predicts reaction times (RT) on an attention task. RT are fastest when both areas align in the On-phase and slowest when both areas are in the Off phase at the time of stimulus change. No effect is observed when attention is directed to a location outside the receptive fields. B-C are adapted from [134].

Changes in the local On-Off dynamics highlight the entanglement between attention and arousal. The mean On-episode duration increases with arousal (indexed by a pupil dilation) [98,107]. Longer On-durations result in fewer transitions, marking a shift toward desynchronized state as expected at higher arousal levels. During spatial attention, the On-episode durations also increase, but only in columns with receptive fields at the attended location [107,108]. Thus, during attention, cortical state shifts toward desynchronization locally within a retinotopic map. This parallel suggests that attention and arousal may recruit the same circuit mechanisms but at different spatial scales. This notion is further supported by network models of On-Off dynamics in interacting cortical columns [110]. On-Off transitions in single columns can be captured as metastable dynamics, similar to models of synchronized activity during sleep or anesthesia [111,112]. In the network, On-Off activity propagates across columns as irregular waves. Attentional inputs modulate the spatial extent of On-Off dynamics giving rise to distance-dependent changes in spike-count correlations as seen in the data [110]. Thus, local metastable dynamics may define the operating regime of cortical networks, where attention and arousal control the spatial extent of synchronized activity on local and global scales.

On-Off dynamics are also coordinated with retinotopic precision across macaque visual cortical areas (Fig. 3B) [108]. The On and Off phases tend to align between V1 and V4 columns with overlapping receptive fields. The On-Off phase alignment gradually decreases with the receptive field separation, suggesting a mirroring of spontaneous waves in V1 and V4. During attention, the inter-areal coordination of cortical state is enhanced and follows the reversed hierarchy, i.e. transitions in V4 tend to precede V1 transitions, and transitions in the FEF precede those in V4. Moreover, the alignment of On-Off phases across visual areas predicts faster reaction times in an attention task (Fig. 3C) [108]. Thus, attentional modulation of cortical state appears to be local within each visual area and coordinated globally across areas. These links between attention and arousal open unique opportunities for studying how fundamental properties of neocortical circuitry that are likely common in mammalian species interact with mechanisms of sensory processing.

Neural oscillations reflecting attentional control

In addition to attentional modulation of neuronal activity and firing-rate states, a large body of research focusing on local field potentials (LFP), electroencephalography (EEG) and magnetoencephalography (MEG), has established a strong correlation between neural oscillations and attentional deployment. Importantly, this insight has emerged from work that includes data obtained in human subjects, thus facilitating comparisons with similar measurements in animals.

One widely observed effect is the attenuation of alpha activity, specifically a power reduction of the intrinsic alpha-band oscillation (8–15 Hz), among regions in the dorsal attention network contralateral to the attended hemifield, as well as in visual areas that are tuned to the attended information [113] (Fig. 4A). In addition, some studies also show that alpha activity can be evoked or enhanced where irrelevant stimuli are processed [114]. These effects suggest that an alpha rhythm may be involved in selective disinhibition of attended representations through decreased alpha power, and/or selective inhibition of task-irrelevant sensory processing through increased alpha power [115]. Such a role of alpha activity was recently tested using MEG neurofeedback [116]. By training subjects to manipulate their relative alpha power between two hemispheres of parietal cortex, this study provided evidence of a causal relationship between alpha asymmetry and behavioral measures of spatial attention. Interestingly, despite subjects showing similar behavioral benefits, the trained alpha asymmetry varied across subjects, some with only an ipsilateral increase and others with only a contralateral decrease, and some with both, consistent with the variability in the observed inhibition/disinhibition effects [115].

Figure 4. Two frameworks for neural oscillations underlying attention.

A. Sustained attentional modulation of behavioral performance and neural oscillations. When endogenous attention is directed in a task (e.g. toward the right hemifield), the power of alpha oscillations in contralateral (left) frontal cortex decreases compared to ipsilateral cortex, and this is associated with a mean difference in behavioral performance between attended (right) and unattended (left) conditions. In visual cortex, a similar alpha asymmetry can be observed, accompanied by an opposite modulation on the stimulus-induced gamma activity, i.e. an increased gamma synchrony in the attended condition compared with the unattended condition. B. Rhythmic attentional modulation of behavioral performance and neural oscillations. In a cued detection task, detection performance fluctuates as a function of the stimulus onset asymmetry (SOA), or the phase of theta-band neural oscillation found in frontal-parietal areas and visual cortex. The slow theta oscillation provides a phasic modulation on the amplitude of other neural oscillations (e.g. gamma). When attention is directed (e.g., toward the right hemifield), the theta-phase dependent fluctuations in behavioral performance at the attended target is attenuated compared to the unattended condition. This is accompanied by reduced theta power and phase-amplitude coupling in visual areas contralateral to the attended location.

In contrast to lower frequency bands, stimulus-driven gamma-band activity (30–70 Hz) shows increased power and synchrony within representations of attended locations or features [117,118], similar to neuronal firing rate modulation. In macaque visual cortex, gamma oscillations have been shown to propagate from granular to supra- and infragranular layers, whereas alpha oscillations propagate in the reverse direction [119]. Thus, they are thought to reflect feedforward and feedback processing, respectively. Indeed, attention modulates these rhythms in a similar laminar-specific manner [120], suggesting the recruitment of both feedforward and feedback pathways during attentional modulation. Moreover, MEG recordings in humans reveal that the trial-by-trial fluctuations in attentional alpha modulation during the pre-stimulus period predict the power of subsequent stimulus-driven gamma activity [121].

Together, this evidence suggests that alpha-band oscillations may participate in attentional modulation through the gating of feedforward gamma activity. In addition to the potential gating modulation by alpha activity, studies in nonhuman primates also find attention-related gamma synchrony within higher-level areas (e.g. prefrontal areas and anterior cingulate cortex), and between higher areas (e.g. FEF) and the visual cortex (e.g. V4), implying that gamma activity might play a more direct role in the top-down modulation of attention [122,123].

Although attentional deployment is often thought of as a “sustained” process, whereby behavioral and neural activity remain stably modulated for a period of time, recent studies have also revealed a “rhythmic” nature of attention. During goal-driven covert attention, in addition to the measured behavioral benefit at the attended target (e.g. increased sensitivity), performance nonetheless fluctuates in a periodic pattern as if subjects are shifting their attention between the target and distractors [124] (Fig. 4B). For example, these fluctuations have been observed for two stimuli at different locations [125] or two overlapping stimuli with different features [126]. In humans, these performance fluctuations are found to depend on the phase of theta-band activity (3–8 Hz) in the frontal-parietal network as well as in visual cortex [127]. In addition to behavioral effects, the phase of theta oscillations also modulates the amplitude of oscillations in other frequencies both locally and remotely, a phenomenon called phase-amplitude coupling (PAC). Specifically, the “good” theta phase (i.e. periods with enhanced behavioral performance) is associated with increased stimulus-driven gamma power, whereas the “poor” phase features stronger alpha oscillation [128]. Such fluctuations of higher-frequency neural oscillations and behavioral performance in theta cycles suggest a mechanism of rhythmic attentional scanning when multiple items are represented simultaneously [129].

The theta rhythm has been shown to propagate across brain areas in various directions. For example, in the monkey, studies show a greater theta-band influence from anterior cingulate cortex (ACC) to the FEF during sensorimotor mapping [130], and from the FEF to LIP during an endogenous attention task [128], highlighting the role of the frontal cortex in leading the theta-rhythmic control of attention. In contrast, Granger Causality (GC) analysis reveals a feedforward direction of theta-band influence along the macaque ventral visual pathway, which is actually suppressed during attention. This observation suggests that a different type of theta activity is driven by simultaneously presented stimuli and progresses from early to late visual areas [131]. These differences in the direction of theta modulation (feedforward and feedback) could correspond to two different forms of attention, e.g. stimulus-driven and goal-directed attention, respectively.

Future research is needed to clarify the cellular- and circuit-level bases of the above effects during attention. What are the sources and pathways of different neural oscillations and their attentional modulation? How does the oscillatory field activity reflect and/or interact with collective neuronal activity, e.g. how might the theta-rhythm oscillations relate to the modulation of On/Off states? Studies using different animal models, could shed light on these questions by combining behavioral metrics, genetic tools (e.g. gamma-band optogenetic stimulation within frontal-cingulate network and its attentional effect in mice [132]), and different levels of recordings (e.g. combined recordings of lower-frequency LFP and intracellular membrane potentials during attention-like improvement of mice behaviors [64]).

Concluding Remarks

In recent years, a wealth of research has focused on the neural mechanisms and circuits involved in different forms of attentional control, in both humans and nonhuman model species, and at different levels of analysis. Nonetheless, much more work will be needed to isolate the mechanisms by which sensory information is selectively processed, and to identify the mechanisms that are both common and distinct across species and different types of attention (see Outstanding Questions). However, in some ways, the expanded breadth of attention studies has already yielded progress on both fronts. For example, it has revealed a common role of long-range cortico-cortical and subcortical input to sensory cortex from motor-related structures across species and types of attention. Furthermore, recent evidence has highlighted the importance of examining the interface between sensory coding and the dynamics of internal brain states, regardless of what those dynamics reflect, e.g., global arousal or behavioral goals. We suspect that future work will continue to benefit from a focus on that interface.

Outstanding Questions.

1. How similar or different across species are the circuit motifs (e.g. long-range disinhibitory motifs within neocortex) involved in the modulation of sensory responses during goal-driven attention? Likewise, are the circuit motifs similar or different across different forms of goal-driven attention, e.g. spatial vs. feature or object-based?

2. To what extent is the salience of a stimulus a function of the proportion of neural resources involved in processing it? For example, we might assume that olfactory cues will be more salient than visual cues in an olfactory-dominant brain (e.g. mouse) than a visually dominant one (e.g. marmoset). If true, does that principle generalize to within sensory modalities of the same species, e.g. face vs. non-face objects in the primate brain where faces are disproportionately represented? Could this be a mechanism underlying ethological salience?

3. How do neuromodulators (e.g. acetylcholine, norepinephrine) influence the dynamics of cortical traveling waves and states? What are the unique contributions of different neuromodulators to arousal and attention?

4. Fluctuations in cortical states and traveling waves (e.g. within retinotopic areas) have been shown to influence sensory processing across space [105]. Given the modularity of sensory cortices (e.g. tonotopy in auditory cortex, feature and object-selective modules in visual cortex), might the same fluctuations also influence sensory processing across features and objects?

5. What is the functional role of oscillatory LFP activity during attentional control? What is its relationship with changes in spiking activity and information coding within neural circuits?

Highlights.

1. The quest to identify neural mechanisms of selective attention has broadened in recent years to include multiple forms of attention, multiple species and multiple levels of analysis.

2. New data reveals distinct, and causal roles of the superior colliculus and parietal cortex in stimulus-driven visual attention

3. Convergent evidence across species has identified a role of motor-related signals in the control of goal-directed attention.

4. Fluctuations in cortical activity associated with global arousal interact with goal-driven attention and both predict perceptual performance.

5. Neural oscillations reflecting attentional control in human and non-human primates suggest an underlying rhythmic mechanism synchronized to theta.