Neural mechanisms for the attention-mediated propagation of conceptual information in the human brain

The visual environment contains more information than we can evaluate or act upon. To behave adaptively, we prioritize some information sources over others using a set of mechanisms collectively referred to as selective attention [1,2]. Early work characterized attention in now-famous metaphors, suggesting a spotlight or filter that acts in retinotopic space to enhance sensation and perception [3,4]. Guided by these ideas, neuroscientific investigation has focused closely on identification of the systems and algorithms that impact sensory cortex [1,5,6] and on characterization of the brain networks that control these mechanisms [7–9]. The emergent model suggests that strategic selection begins with the derivation of schemas and templates from conceptual and semantic knowledge. These are stored in working memory at frontal brain sites and translated through the frontal eye fields and posterior parietal cortex into perceptual hypotheses and biases of sensory activity [1,10].

Research on the neurophysiology of attention has thus focused closely on how concepts and categories are used to guide sensation and perception. In contrast, our understanding of the reverse inference—of how attention supports the activation and construction of concepts—is relatively poor. That is, we know that attention plays a key role in information gathering and the generation of high-level knowledge [11,12]. And, while there is debate about the degree to which concepts are grounded in sensorimotor systems [13], we know that the semantic representations that emerge in frontal and temporal cortex abstract from perception [14–17]. Theoretical treatments of attention propose that a core purpose of selection is to activate and construct these conceptual representations, in this way supporting cognition and decision-making [11,12], but we know little of how this unfolds in the brain.

To address this, in the current paper we identify and characterize neural representations of abstracted information that vary as a function of the quality of attentional selection. This approach has been adopted in earlier work. For example, results show that the semantic category of a visual stimulus is encoded in patterns of fMRI activity in the ventral and dorsal visual cortex [18,19] and that these representations vary as a function of their behavioral relevance, consistent with an effect of attention [20–22]. However, this pattern has largely been identified in brain areas that are likely to encode visual features, making it unclear that it reflects an effect on category information rather than on perception of the visual features that happen to be shared across category examples [23–25]. To circumvent this issue, a separate research line has employed multi-modal stimuli that activate the same concept—for example, by presenting an image of an apple and the word ‘apple’. This has identified category information in areas that do not carry perceptual signals, such as anterior temporal and prefrontal cortex [15–17,26]. However, the role of attention in the activation or construction of these representations remains unclear.

As a result, there are holes in our understanding of the neural mechanisms that support information gathering and abstraction. Theory suggests that that information selected through the deployment of attention activates or constructs neural representations of concepts and categories [11,12], but this has not been verified, and we do not know where in the brain this takes place. Are all such high-level representations sensitive to attention, or is the effect of attention limited to discrete brain networks? And what does this mean for our understanding of the functionality of these areas? Here, we use novel methodology and analyses to identify the effect of attention on information about the semantic category of visual stimuli. Our approach relies on identification of covariance between a well-characterized EEG index of spatial attention—the N2pc [27]—and an independent measure of semantic category information derived using machine learning of neuroimaging data. Experiment 1 employs high-density EEG in the identification of this information to allow for precise temporal resolution. Experiment 2 employs concurrently recorded EEG and MRI to provide precise spatial localization.

Participants in both experiments (n =  44, n =  31) were asked to complete the same task without moving their eyes. This involved reporting, via button press, the category of a target object (Fig 1A) that appeared at a cued location (Fig 1B), while ignoring a set of non-targets that included visually degraded objects and a single recognizable distractor. Objects were selected from a set of images that were visually heterogeneous within each category. Participants completed the task quickly (E1: 1,076 ms, 185 SD; E2: 1,088 ms, 240 SD) and accurately (E1: 87.6%, 10.6 SD; E2: 89.9%, 7.0 SD).

Fig 1 . Experimental task. (

a) Complete, recognizable examples of object categories and morphed, unrecognizable versions of the same images. In Experiment 1, the stimuli set was composed of 40 complete images in each of the four categories. In Experiment 2, a subset of 30 complete images per category was employed. All images presented in this figure are included in the BOSS image database [74] and written permission to employ these images in this figure has been kindly granted by the owner. (b) Experimental procedure. The target was cued by a central letter, either “R” or “G” to denote red or green. Participants had to report which category appeared in the cued circle via right-hand button response. Two categories were associated with index finger response and two with middle finger response, with response mapping counterbalanced across participants.

https://doi.org/10.1371/journal.pbio.3003018.g001

In both experiments, we measured the quality of attentional processing using the N2pc component of the visual event-related potential (ERP) [28]. The N2pc emerges over the posterior cortex contralateral to the location of an attended object [28,29]. It increases in amplitude as circumstances require a greater investment of attentional resources [30,31] and is therefore used as an index of the depth of attentional processing [28,32]. In Experiment 2, we conduct additional exploratory analyses of how lateral ERP effect preceding or following N2pc also impacted the emergence of category information.

In each experimental trial, the target object could appear to the left or right of a central fixation mark or directly above or below it; when the target was presented laterally, the single recognizable distractor was presented on the vertical, and vice versa (Fig 1B). This design was adopted as stimuli on the vertical meridian of the visual field initially create activity of equal magnitude in both cortical hemispheres, and therefore attentional selection of these stimuli does not generate the lateralized EEG voltage underlying N2pc [33]. By presenting stimuli on the vertical we were able to isolate the N2pc elicited exclusively by the target or distractor [29,34].

The N2pc emerged prominently in both experiments (Fig 2A and 2D) and was larger in amplitude for targets than recognizable distractors (Fig 2B and 2E). It was preceded by an earlier, positive-polarity lateral voltage that reflects an imbalance in sensory activity between the visual hemispheres. In our stimulus displays, there is a difference in the complexity and spatial frequency of the complete object, presented on one side of the screen, and the unrecognizable degraded object, presented on the other (Fig 1B). This kind of imbalance creates short-latency laterality in the visual ERP (i.e., PPN) [35]. The insensitivity of this effect to the behavioral relevance of the evoking stimulus distinguishes it from the subsequent N2pc.

Fig 2. Results from EEG. (

a) ERP results from Experiment 1. The N2pc is evident as a negativity in the posterior waveform recorded over the cortex contralateral to an attended stimulus. The blue box identifies a 125 ms interval centered on the cross-conditional peak of the N2pc. The topographical plot reflects the voltage difference between conditions when the target was in the left vs. right visual field; the N2pc expresses as a negativity in the left hemisphere but a positivity in the right hemisphere due to the subtraction. Contralateral and ipsilateral ERPs reflect mean signal across the 2 sets of 3 lateral and posterior electrodes identified in the topography by a larger marker. (b) Contralateral-minus-ipsilateral difference waves from Experiment 1. The N2pc is evident as negative deflection beginning around 175 ms post-stimulus. Shading here and in other panels indicates bootstrapped SEM. (c) Time-resolved classification of the target and distractor object category in Experiment 1. As there were four classes, the chance performance was 25%. (d) ERP results from Experiment 2. These reflect signal at the lateral posterior electrodes identified in white in the topography presented in panel A. (e) Contralateral-minus-ipsilateral difference waves from Experiment 2. (f) Time-resolved classification of target and distractor object category in Experiment 2.

https://doi.org/10.1371/journal.pbio.3003018.g002

Target and distractor category information was identified in the EEG using classification analysis. For each sample in the epoched data, an LDA model was trained on results observed in a ~85 ms window centered on that observation. A cross-fold validation scheme—described further below—was employed to test the model in each of these windowed samples, generating a time-course of classification accuracy. As illustrated in Fig 2C, when the target appeared laterally, the target category could be reliably classified from approximately 215 ms post-stimulus in Experiment 1 (Fig 2C) and from 240 ms in Experiment 2 (Fig 2F). Distractor classification also emerged in Experiment 1, though later and with reduced magnitude and extent.

To test this, we iteratively resampled the data in a bootstrap routine, building classification models and measuring the N2pc in each instance (Fig 3A). For each participant, target-lateral and distractor-lateral trials were separated into 20 chunks, each containing an equal number of trials where the classified object was a food, vehicle, animal, or tool. The chunks were allocated repeatedly into two subsets. The larger set—constituting 17 chunks or 85% of trials—was used to train a classification model that identified the target category. The smaller set—constituting the remaining three chunks or 15% of trials—was used both to test the model and to measure the target- or distractor-elicited N2pc.

Fig 3. (a) Analysis of EEG data in Experiment 1.

Trials were repeatedly partitioned into training and test sets. Classification accuracy and mean N2pc amplitude were measured for each test set, resulting in two dependent measures (identified in bold). (b) Analysis of MRI and EEG data in Experiment 2. The imaging modalities were linked during trial partitioning, such that the same trials appeared in both MRI and EEG training and test sets. This resulted in three dependent measures (identified in bold). (c) To link EEG and MRI results in Experiment 2, we correlated MRI classification accuracy with N2pc amplitude observed across the set of 1,140 sampling iterations. This occurred for each searchlight sphere.

https://doi.org/10.1371/journal.pbio.3003018.g003

For each of the target-lateral and distractor-lateral conditions, we iterated this process of sampling and model building 1,140 times, reflecting all possible combinations of 3 chunks taken from 20. The classification accuracy illustrated in Fig 2C and 2F reflects the mean of these values across iterations and subsequently participants. In each iteration, we measured the N2pc as mean voltage in a 125 ms interval centered at the peak (as identified from data collapsed across iterations, conditions, and participants; 173–298 ms). We measured mean classification accuracy in this same interval, resulting in paired sets of 1,140 N2pc and classification accuracy values for each analysis and participant. As illustrated in Fig 4A, within-participant results commonly showed a relationship between target-elicited N2pc and target classification accuracy, with negative-voltage N2pc amplitude predicting an improvement in target classification. The mean relationship is plotted in Fig 4B alongside results from each of the 44 participants.

Fig 4. Linking N2pc and EEG category information in Experiment 1.

(a) Within-participant results from three example participants. Across the 1,140 sampling iterations, the target-elicited N2pc observed in each sample predicts target classification accuracy in that sample. (b) Least-square linear fit of target-elicited N2pc and target classification accuracy for each participant, with mean relationship identified in red. (c) Linear fit of distractor-elicited N2pc and target classification accuracy for each participant, with mean relationship identified in red.

https://doi.org/10.1371/journal.pbio.3003018.g004

One interpretation of this pattern is that robust attentional selection of the target—reflected in larger target-elicited N2pc—caused the EEG to carry more information about the target category. However, an alternative is that both N2pc and target information covaried as a function of participant arousal or task engagement without any more direct relationship. We conducted a second analysis to probe this alternative. Here, we related target information to distractor-elicited rather than target-elicited N2pc. If the N2pc and category information broadly covary, this analysis should identify the same relationship as described above. However, as illustrated in Fig 4C, the magnitude of distractor-elicited N2pc predicted a decrease in target classification accuracy.

We used hierarchical linear modeling of EEG and model assessment to statistically evaluate these patterns [36]. An initial model of classification accuracy included a single random effect for the per-participant intercept. This was improved by adding a fixed effect—condition—that identified if the target or distractor was being classified (alongside a corresponding random effect for the per-participant slope; Akaike information criterion (AIC) −8746) [37]. The model was further improved by adding a fixed effect for N2pc amplitude (and corresponding random effect; AIC −702) and the interaction of these factors (and corresponding random effect: AIC −896).

(1)

ANOVA analysis of this model identified an interaction of condition and N2pc (F(1, 42.6) =  13.72, p Fig 4B and 4C. Follow-up modeling examined the relationship between N2pc and category information separately for each of the conditions.

(2)

In Experiment 2, participants completed the same task while EEG and fMRI were concurrently recorded (Fig 3B). We began analysis with the identification of category information in the unimodal fMRI signal. This analysis does not consider the laterality of the target or distractor image. We estimated single-trial fMRI responses using the GLMsingle toolbox [78] and identified category information in brain space using spatial searchlights and LDA classification. In this analytic approach, a sphere is defined around each voxel, and classification is conducted based on the pattern of activity observed in the subset of voxels within this volume [38]. Searchlight spheres had a diameter of 7 voxels (14 mm) and classifiers were trained and tested using 4-fold cross-validation. This identified an overlapping set of brain areas that contained information about both the target and the distractor, alongside a more limited set of areas that exclusively contained information about one of these two objects (Fig 5A). Brain areas expressing category information for both target and distractor included ventral visual cortex; inferior and superior parietal lobules (IPL/SPL); and pre-motor, motor, and somatosensory cortex. Target information also emerged in the left inferior frontal gyrus (IFG), bilateral orbitofrontal cortex (OFC), and precuneus/cingulate cortex; distractor information extended from the IPL into the left temporal-parietal junction (TPJ). Statistical analysis identified that the bilateral LOC, IPL/SPL, left IFG, left OFC, posterior cingulate (PC), and left basal ganglia (BG) all reliably contained more information about the target than the distractor (Fig 5B). No area reliably contained more information about the distractor than the target.

Fig 5 . Results from MRI classification.

(a) Searchlight analysis identifying the location of target and distractor information. Target and distractor category information emerge with particular strength in the aspects of occipital and posterior temporal lobes associated with the ventral visual stream. (b) Voxels identified here reflect clusters expressing more information about the target than about the distractor. In both panels, the largest clusters emerge in the ventral visual cortex; other cluster locations are identified. OFC, orbitofrontal cortex; IPL, inferior parietal lobe; SPL, superior parietal lobe; IFG, inferior frontal gyrus; TPJ, temporal-parietal junction. The left hemisphere is represented on the left here and in subsequent figures.

https://doi.org/10.1371/journal.pbio.3003018.g005

In Experiment 1, we used a resampling approach to identify within-participant covariance in independent measures derived from EEG. In Experiment 2, we used a similar approach to link variance in measures derived from the two imaging modalities (Fig 2C). For each participant, trials were again separated into 20 chunks of equal size, with 17 chunks assigned to the training set, and 3 chunks to the test set. This process of sampling and model building was iterated 1,140 times for each of the ~2.1 × 10⁵ searchlight spheres. In each of these iterations, the N2pc was measured in the test set as the mean voltage across a 125 ms interval centered at the peak (as identified from data collapsed across iterations, conditions, and participants; 189–315 ms). This generated a paired set of 1,140 N2pc and classification accuracy values for each searchlight sphere. We calculated the Fisher-transformed Pearson correlation for each of these sets, inserting the resulting values into a new brain volume at the location of each searchlight center. As illustrated in Fig 6, this identified three clusters where target-elicited N2pc amplitude predicted MRI-derived target category information. This included a cluster in the left IPL/angular gyrus extending into the SPL (92 voxels; Fig 6A), a cluster that spanned the putamen, globus pallidus, insula, and anterior temporal lobe in the right hemisphere (84 voxels, Fig 6B), and a cluster that spanned the bilateral ventromedial prefrontal cortex and OFC (VMPFC/OFC; 163 voxels, Fig 6C).

Fig 6. Linking N2pc and MRI category information in Experiment 2.

Results illustrate brain regions where target category information is predicted by target-elicited N2pc amplitude. Time-course insets show the time-resolved correlation for each MRI cluster. Significant time clusters are indicated by black lines below the x-axis and shading reflects bootstrapped SEM. Location and time-course for (a) the Inferior Parietal Lobule cluster, (b) the Insula/ Putamen/Globus Pallidus cluster, and (c) the Ventromedial Prefrontal Cortex/ Orbitofrontal Cortex cluster.

https://doi.org/10.1371/journal.pbio.3003018.g006

While these correlational results cannot demonstrate causation, one interpretation here is that attentional selection of the target—reflected in target-elicited N2pc over the occipital cortex—caused the emergence of conceptual information in the identified brain areas. However, as in Experiment 1, the alternative is that N2pc and category information broadly covaried as a function of other variables associated with participant state, like arousal. We again conducted a control analysis to test this, relating target category information to distractor-elicited rather than target-elicited N2pc. We created three regions of interest (ROIs) that described the clusters illustrated in Fig 6, calculated classification accuracy using the entire set of voxels within each of these ROIs, and used our 1,140-fold technique to relate this information to the N2pc. This showed that the relationship between target category information and target-elicited N2pc was significantly greater than the relationship between target category information and distractor-elicited N2pc in all three ROIs (IPL/SPL: p =  0.005; insula: p =  0.027; VMPFC/OFC: p p =  0.019; other areas ps >  0.230).

We conducted whole-brain analyses to relate distractor-elicited N2pc to variance in target or distractor information but failed to identify any set of voxels where these relationships survived statistical correction. Analysis of distractor-elicited N2pc and distractor information based on ROIs defined from results illustrated in Fig 6 also failed to identify significant results. The absence of reliable effects involving distractor information may in part reflect the limited magnitude and scope of this information in the MRI data (Fig 5).

To this point, we have focused on the N2pc measured at a fixed latency. However, laterality in the EEG emerges both before and after the N2pc; as discussed above, early laterality emerges as a function of sensory activity, but also as a product of stimulus novelty [39] and distractor suppression [39–41]. Later, laterality has been closely linked to the encoding and maintenance of stimuli in visual working memory [42]. We expanded our analysis to explore the role of pre- and post-N2pc EEG laterality in the control of conceptual encoding in the brain. For each of the ROIs generated from results illustrated in Fig 6, we related variance in each lateral EEG sample to variance in MRI-derived target category information. This generated a time-course of the relationship between lateral occipital EEG- and MRI-derived target category information for each ROI. These are illustrated in Fig 6. In the IPL, the relationship between EEG laterality and target category information emerged very quickly and sustained (Fig 6A). The pre-N2pc relationship may reflect distractor suppression that guides selective processing of the complete object rather than the degraded object in the contralateral visual hemifield [41,42]. The post-N2pc relationship, in contrast, is likely to index the maintenance of visual working memory representations of the target in the posterior parietal cortex [43]. We expect that these mechanisms would operate in concert with the N2pc to guide selective processing and information gathering. In contrast, in the VMPFC and insula clusters, the relationship between EEG laterality and MRI-derived target information is relatively discrete to the latency interval of the N2pc (Fig 6B and 6C). These results suggest that the specific selective processes indexed in N2pc are closely linked to the propagation of target category information to these areas.

As illustrated in Fig 7A, this identified three voxel clusters where variance in target-elicited N2pc amplitude predicted variance in voxel activation: the middle frontal gyrus (MFG; 377 voxels), the IPL extending into the TPJ (486 voxels), and the dorsomedial superior frontal gyrus (dmSFG; 296 voxels). These areas are largely independent of those identified in the corresponding analysis of category information. Significant results emerged in the IPL in both analyses, but the clusters did not overlap, with N2pc predicting activation in voxels anterior and lateral to those identified in the analysis of category information.

Fig 7. Linking N2pc and MRI voxel activation in Experiment 2.

Results illustrate brain regions where target-elicited N2pc amplitude is predicted by voxel activity. (a) Location of the three identified clusters: Left Inferior Parietal Lobule/Angular Gyrus/Temporoparietal Junction, Left Middle Frontal Gyrus/Frontal Eye Field, and Bilateral Dorsomedial Superior Frontal Gyrus. (b) Time-resolved correlation for each MRI cluster. Significant time clusters are indicated by black lines below the x-axis and shading reflects bootstrapped SEM.

https://doi.org/10.1371/journal.pbio.3003018.g007

As in the analysis of category information, we examined the relationship between distractor-elicited N2pc and voxel activity as a control. Whole-brain analysis did not identify any brain area where the distractor-elicited N2pc predicted voxel activity after statistical correction. When we defined ROIs based on the results from whole-brain analysis illustrated in Fig 7, there was no significant relationship between distractor-elicited N2pc and mean voxel activity in these areas, and the relationship between target-elicited N2pc and voxel activity was significantly greater than the relationship between distractor-elicited N2pc and voxel activity (IPL: p =  0.007; dmSFG: p p =  0.007).

As in the analysis of category information, we expanded our analysis to test the power of pre- and post-N2pc EEG laterality to predict univariate voxel activity. As illustrated in Fig 7B, a relationship between lateral EEG and voxel activity emerges sooner in IPL than in MFG, though both effects are in the 175–350 ms interval where the N2pc can emerge. Both left IPL and left MFG have been closely linked to the extraction of semantic content during reading [45,46], and left and right IPL and MFG have roles in the orientation and reorientation of spatial attention [8,47]. One interpretation of this finding is that the emergence of IPL and MFG reflect control in the deployment of attention for the extraction of semantic information.

The relationship between lateral EEG and activity in the dmSFG, in contrast, emerges from very early, including latencies that precede stimulus onset. Among other functions, this portion of medial frontal lobe contains eye fields and has been associated with fixation, including during smooth pursuit eye movements [48]. In the current results, trials with eye movements substantively preceding stimulus onset were not excluded from analysis; the association of activity in the dmSFG with EEG laterality underlying the N2pc may reflect oculomotor control supporting orientation of the eyes to the fixation mark or in maintenance of the eyes at that location. Similarly, later association of dmSFG with EEG laterality in the N2pc and post-N2pc interval may reflect the involvement of this brain area in the inhibition of saccadic movements to covertly attended locations.

We leveraged natural variability in the quality of attentional selection to identify the temporal and anatomic characteristics of the attention-mediated propagation of conceptual information in the human brain. In Experiment 1, we find that the amplitude of target-elicited N2pc, an index of spatial attention, predicts that the evoked EEG signal will carry more information about the semantic category of an attended object (Fig 4B). This relationship emerges quickly, with attention-mediated category information emerging in the same 173–298 ms window used in measurement of the N2pc. The misallocation of attention to task-irrelevant stimuli, reflected in distractor-elicited N2pc, predicts a reduction in the quality of target category information in the same interval (Fig 4C). Though these results are correlational, making it impossible to draw firm conclusions regarding causation, the pattern is in line with the idea that attentional selection of the target results in the emergence of category information in brain activity.

One goal of Experiment 2 was therefore to clarify the contribution of visual perception in the relationship between attention and category information. Initial analysis of unimodal fMRI identified a network of brain areas where category information was greater for targets than for distractors (Fig 4B). This prominently included bilateral ventral visual cortex, alongside smaller clusters in bilateral IPL/SPL, left IFG (IFG), left OFC, left BG, and bilateral PC gyrus. These results therefore include aspects of the occipital and posterior parietal cortex that are strongly associated with the representation of visual features and translation of visual information to motor control [49,50], alongside BG structures that may have a role in visual perception [51]. However, also evident is a set of left-lateralized areas including IFG and VMPFC/OFC, alongside bilateral PC, that are not putatively visual in nature, but have been implicated in earlier work as containing abstract semantic representations [26].

Further analysis identified three brain areas where trial-wise variance in target-elicited N2pc amplitude predicted the quality of target category information: the left IPL/SPL; an insula-centered cluster in the right hemisphere including portions of the putamen and globus pallidus; and a bilateral VMPFC-centered cluster extending into medial OFC (Fig 5). These are all nodes in the “semantic network” identified from a large-scale meta-analysis of fMRI studies of semantic representation [26]. The portion of IPL/SPL identified in this analysis was physically close to the posterior parietal cluster identified in the analysis of unimodal fMRI, as detailed above, but was slightly posterior with no overlap, emerging primarily in the angular gyrus. Category information in this area expressed a relationship with ERP laterality over much of 500 ms following stimulus onset, including time periods before and after N2pc. Activity in this portion of the parietal lobe is implicated in the control of visual attention [52,53] but also in the retrieval and representation of episodic [54–56] and semantic information [47]. In line with other recent theory [47], our tentative account is that this aspect of IPL/SPL is a hub in the comparison of visual characteristics of the attended stimulus to category templates retrieved from memory, in this way acting as an important waystation in the attention-mediated classification and abstraction of visual input.

Interpretation of the cluster centered on the right insula is fundamentally challenging because of the breadth of functional roles associated with this area [57,58]. A compelling possibility; however, is that this reflects the role of insula and surrounding brain tissue in the “salience network”, a set of brain areas implicated in the identification of homeostatic relevance and behavioral importance [59]. The salience network is thought to guide the flow of information through other brain networks, including prefrontal cortex, in the support of attentional processing and other cognitive operations [60]. Unlike IPL/SPL, category information in insula correlated discretely with ERP laterality in the latency of the N2pc, implicating the N2pc specifically in the propagation of information to this area. The emergence of attention-mediated category information in this cluster suggests a role for the insula and surrounding tissue in the processing of abstracted object information to guide the evaluation of behavioral relevance and the preparation of a cognitive response, possibly in the context of current participant state and task motivation.

The cluster centered on VMPFC is the largest of the three, has the strongest relationship with N2pc, and uniquely shows a link to both target-elicited and distractor-elicited N2pc, with target category information reducing in this area as a function of the mis-deployment of attention to the distractor. As with the insula, this cortical area has been associated with extensive functionality [61]. Prominent is the idea that it carries information about the abstracted economic utility of goods and actions—the “common currency” hypothesis [62,63]. This kind of signal is necessary for economically normative models of learning and decision-making to allow the utility of disparate goods and outcomes to be directly compared. However, our results show the emergence of attention-mediated category information in this area when categories do not explicitly differ in value. This is in line with the known involvement of this area in semantics [26], but is difficult to resolve with the notion that the core responsibility of this brain area is to represent value abstracted from context and stimuli. Building from other divergent findings in the literature, recent models of OFC/VMPFC function suggest that this brain region may rather be involved in the derivation of task state from disparate attributes of environmental objects and context [64,65] and in the establishment of cognitive maps of task space [66–68]. The current results are easily described in this theoretical context: strong attentional selection of a behaviorally-relevant visual object leads to the activation of a cognitive map in OFC/VMPFC that includes information about the target object category and its associated response. When attention is mis-deployed to the distractor, target information in this cognitive map is degraded.

Importantly, this does not imply that the attention-mediated representations that we identify in the posterior parietal cortex, insula, and prefrontal cortex are necessarily amodal or instantiated independent of sensorimotor systems [69]. The relationship between N2pc and category information we identify is not a simple reflection of attentional effects on visual features but is likely to contain semantic features that are grounded in sensorimotor systems, including information about the specific motor response associated with each category [70]. The pattern we identify—that attentional effects on perceptual representations in the lateral occipital cortex predict the quality of perception-abstracted information elsewhere in the brain—is conceptually in close line with modern accounts of grounded semantic representation [71,72].

In addition to our identification of attention-mediated propagation of semantic information, results from Experiment 2 provided the opportunity to identify brain structures where N2pc amplitude predicted unimodal fMRI activation. This identifies a set of brain areas including left MFG (including the left frontal eye fields) [73], left IPL extending into TPJ, and bilateral dmSFG (Fig 7). All of these areas have documented roles in attentional control. As noted above, the FEF and IPL are core nodes in the frontoparietal network implementing strategic attentional control, and the broader MFG and lateral posterior parietal cortex have been implicated in the use of attention for semantic extraction. Similarly, dmSFG contains eye fields and plays a role in saccadic control. This suggests that activity in these areas may be involved in the establishment and instantiation of attentional control, such that their activation causes an increase in N2pc amplitude. This account is reinforced by the temporal pattern of the relationship between dmSFG and N2pc, and between IPL and N2pc, where laterality in the ERP preceding the N2pc window predicts activity in these areas (Fig 7B).

Our thanks to Vinura Munasinghe, Holly Ahmed, and Tia Cainer for support with EEG data collection; to Steve Mayhew for technical expertise on the recording of concurrent EEG/MRI; to Alicia Rybicki, Alan Kingstone, Marius Peelen, and Wieske van Zoest for discussion; and to the University of Birmingham’s BlueBEAR high-performance computing service (www.birmingham.ac.uk/bear) for computational resources.