Abstract

Errors indicate the need to adjust attention for improved future performance. Detecting errors is thus a fundamental step to adjust and control attention. These functions have been associated with the dorsal anterior cingulate cortex (dACC), predicting that dACC cells should track the specific processing states giving rise to errors in order to identify which processing aspects need readjustment. Here, we tested this prediction by recording cells in the dACC and lateral prefrontal cortex (latPFC) of macaques performing an attention task that dissociated 3 processing stages. We found that, across prefrontal subareas, the dACC contained the largest cell populations encoding errors indicating (1) failures of inhibitory control of the attentional focus, (2) failures to prevent bottom-up distraction, and (3) lapses when implementing a choice. Error-locked firing in the dACC showed the earliest latencies across the PFC, emerged earlier than reward omission signals, and involved a significant proportion of putative inhibitory interneurons. Moreover, early onset error-locked response enhancement in the dACC was followed by transient prefrontal-cingulate inhibition, possibly reflecting active disengagement from task processing. These results suggest a functional specialization of the dACC to track and identify the actual processes that give rise to erroneous task outcomes, emphasizing its role to control attentional performance.

Introduction

The anterior cingulate cortex (ACC) has been suggested to be of particular relevance to optimize attentional control processes, required for a wide variety of tasks (Hayden et al. 2011; Hyman et al. 2011; Euston et al. 2012; Holroyd and Yeung 2012; Khamassi et al. 2013; Narayanan et al. 2013; Shenhav et al. 2013). Lesion and recording studies have shown that the ACC is critical (1) to optimize choice behavior from integrating reward histories of prior choices, (2) to convey stimulus- and action-specific reward expectations (Kennerley et al. 2006; Quilodran et al. 2008; Kaping et al. 2011; Luk and Wallis 2013), and (3) to provide early onset neuronal task control signals specifically after erroneous or suboptimal choices have been made (Johnston et al. 2007; Quilodran et al. 2008; Womelsdorf et al. 2010; Bryden et al. 2011).

According to recent conceptualizations, the ACC accomplishes such a high-level control function by evaluating whether behavioral outcomes indicate the need to adjust control demands for better future performances (Khamassi et al. 2013; Shenhav et al. 2013). Consistent with this view, cell populations in the ACC are sensitive to a variety of different outcomes (Alexander and Brown 2011): Subsets of 7–15% of cells in the ACC encode separable types of outcomes, including positive and negative reward prediction errors (Matsumoto et al. 2007; Sallet et al. 2007; Quilodran et al. 2008), unsigned prediction errors including surprising events (Hayden et al. 2011; Kennerley et al. 2011), and outcome signals that vary with the expected reward magnitude (Amiez et al. 2005; Seo and Lee 2007).

These diverse outcome-specific firing-rate modulations in the ACC have in common that they can be used to adjust stimulus- and response-outcome predictions in future trials (Alexander and Brown 2011; Khamassi et al. 2013; see also Holroyd and Coles 2008). However, beyond optimizing reward expectancies, subjects utilize errors to identify whether they allocated sufficient attentional control to task aspects to accomplish desired goals. This perspective emphasizes that errors act as interruptions of ongoing processes, serve as “circuit breakers,” and function as an immediate “vigilance signals” to identify what processing aspects needs refinement (Dehaene et al. 1998; Alexander and Brown 2011; Hayden et al. 2011; Shenhav et al. 2013). This control-adjusting function of errors has recently been formalized in the Expected Value of Control (EVC) framework (Shenhav et al. 2013). The EVC predicts that the ACC is functionally specialized to track information about current processing states, to rapidly detect errors (changes in processing states), and to specify which processes require enhanced control in subsequent trials. According to this framework, separate cells in the ACC should not only detect erroneous outcomes, but should also be specifically tuned to the attentional processing requirements that failed at the time of error commission.

Here, we tested this “process specificity” of attentional error signals in the ACC and across a large extent of medial and lateral prefrontal cortices of macaques. Monkeys performed an attention task that dissociated errors originating from failures of attentional focusing, attentional filtering, and stimulus-response mappings. Our results reveal an apparent functional specialization of ACC (area 24) to encode specific errors of attentional control, and suggest that ACC error detection triggers a widespread posterror inhibition that disengages ACC and lateral PFC from ongoing task processing.

Materials and Methods

Experimental Procedures and Paradigm

We collected data from 2 male macaques following the guidelines of the Canadian Council of Animal Care on the use of laboratory animals and of Western University's Council on Animal Care. The following experimental procedures and data acquisition protocols have been described in detail in Kaping et al. (2011).

Extracellular Recordings

Extracellular recordings commenced with 1–6 tungsten microelectrodes (impedance 1.2–2.2 MΩ, FHC, Bowdoinham, ME, USA) through standard recording chambers (19 mm inner diameter) implanted over the left hemisphere in both monkeys. The recording chambers allowed access to more anterior aspects of the prefrontal cortex and cingulate sulcus in both animals (see Kaping et al. 2011 and see below the section reconstruction of Recording Sites). Electrodes were lowered in guide tubes with software controlled precision microdrives (NAN Instruments Ltd, Israel) on a daily basis, through a recording grid with 1 mm interhole spacing. Before recordings began, anatomical 7-T magnetic resonance imagings (MRIs) were obtained from both monkey's to allow reconstruction of electrode trajectories and recording sites.

Data Acquisition

Data amplification, filtering, and acquisition were done with a multichannel processor (Map System, Plexon, Inc.), using unity-gain headstages. Spiking activity was obtained following a 100- to 8000-Hz bandpass filter, further amplification and digitization at a 40-kHz sampling rate. During recording, the threshold was adjusted to have always a low proportion of multiunit activity visible against which we could separate single neuron action potentials (APs) in a 0.85- to 1.1-ms time window. Sorting and isolation of single-unit activity was performed offline with a Plexon Offline Sorter (Plexon, Inc., Dallas, TX, USA), based on principal component analysis (PCA) of the spike waveforms, and strictly limiting unit isolation to periods with clear temporal stability and separation to waveforms from other isolated neurons and multiunit activity. The monkey's eye position was tracked continuously with an infra-red system (ISCAN, Woburn, USA) running on a DOS platform, with eye fixation controlled within a 1.4–2.0° radius window. Liquid reward following correct trials was delivered by a custom-made, air-compression controlled, mechanical valve system with a noise level during valve openings of 17 dB within the isolation chamber.

Visual Stimulation

Stimuli were presented on a 19′ CRT monitor placed 57 cm from the monkey's eyes, running at 1024 × 768 pixel resolution and 85-Hz refresh rate. Behavioral control and visual stimulation were accomplished with Pentium III PCs running the open-source software Monkeylogic (http://www.monkeylogic.net/), which has been benchmarked and validated in 2 previous publications (Asaad and Eskandar 2008a, 2008b). We used grating stimuli with “rounded off” edges moving within a circular aperture at 1.0° per sec, a spatial frequency of 1.4°, and radius of 1.5–2.2°. Gratings were presented at 4.2° eccentricity to the left and right of fixation. The grating on the left (right) side always moved within the aperture upwards at −45° (+45°) relative to vertical. Monkeys had to detect a transient and smooth clockwise/counterclockwise rotation of the grating movement (see below). The rotation was adjusted to ensure ≥85% of overall correct responses to the grating and ranged between ±13° and ±19°. The rotation proceeded smoothly from standard direction of motion toward maximum tilt within 60 ms, staying at maximum tilt for 235 ms, and rotating back to the standard direction within 60 ms, and continued moving at the standard ±45° thereafter.

Experimental Paradigm

Monkeys performed a selective attention task requiring a two-alternative, forced-choice discrimination on the attended stimulus (Fig. 1A). Monkeys initiated a trial by directing their gaze to a centrally presented, gray fixation point. Following a fixed 0.4-s period with 2 black/white moving grating stimuli, the moving grating stimuli were colored red/green (“color cue” onset). The location of the red and green grating color was randomized across trials. Within 0.05–0.75 s after color onset, the central fixation point changed to red or green cueing the monkeys to covertly shift attention toward the location with the color-matching stimulus. We label the period of sustained spatial attention the “Attention Epoch” (Fig. 1A, time epoch with red-colored solid line). Errors during the Attention Epoch were fixation breaks—indicating failures to control covert peripheral attention and central fixation—and triggered abortion of the trial. At random times (drawn from a uniform distribution) within 0.05–4 s after cue onset, the cued target grating transiently rotated clockwise or counterclockwise. In half of the trials, the uncued distractor grating transiently rotated before the target. The stimulus rotation of the distractor stimulus had to be ignored, or filtered, which led us to label this trial period the “Filter Epoch” that required control of bottom-up stimulus interference (Fig. 1A, time epoch with red-colored dashed line). Monkeys had to discriminate the rotation of the target stimulus by making a saccadic eye movement up- or downwards to 1 of 2 response targets within 70–550 ms following rotation onset. We label the time after target stimulus rotation that required a choice from the monkeys the “Choice Epoch” (Fig. 1A, blue-colored time epoch). The monkeys received fluid reward after a further delay of 0.4 s after correct saccadic responses. Errors during the Choice Epoch were saccadic responses to the incorrect response target (indicating a wrong discrimination of the rotation direction), or saccadic responses elsewhere (e.g., toward a peripheral stimulus). Stimulus colors were differentially associated with high and low rewards in alternating blocks of trials, which is, however, not subject of this research report (for more details, see Kaping et al. 2011).

Figure 1.

Task design and time epochs of error occurrences. (A) Progression of task epochs within a trial. The animals shifted covertly attention to a peripheral stimulus that matched the color of the centrally presented cue stimulus and maintained selective attention on that stimulus in the “Attention Epoch”. In half of the trials, the attended stimulus transiently rotated before the distractor stimulus rotated. The target stimulus rotation signifies the begin of the “Choice Epoch,” triggering the animal's overt choice to saccade up-/downwards to a response target in response to clockwise/counterclockwise rotations. In the other half of the trials, the distractor stimulus rotated before the Choice Epoch. This distractor rotation had to be ignored, constituting what we term the “Filter Epoch”. (B–D) Time (x-axis) at which errors (y-axis) occurred across all sessions and both monkeys relative to: cue onset (B, Attention Epoch), the distractor event onset (C, Filter Epoch), and the target rotation event (D, Choice Epoch). (E) Illustration of the prefrontal cortex and ACC subdivisions where cells were recorded. Details on the reconstruction of recording sites to different subfields are illustrated in Supplementary Figures 3 and 4.

Figure 1.

Task design and time epochs of error occurrences. (A) Progression of task epochs within a trial. The animals shifted covertly attention to a peripheral stimulus that matched the color of the centrally presented cue stimulus and maintained selective attention on that stimulus in the “Attention Epoch”. In half of the trials, the attended stimulus transiently rotated before the distractor stimulus rotated. The target stimulus rotation signifies the begin of the “Choice Epoch,” triggering the animal's overt choice to saccade up-/downwards to a response target in response to clockwise/counterclockwise rotations. In the other half of the trials, the distractor stimulus rotated before the Choice Epoch. This distractor rotation had to be ignored, constituting what we term the “Filter Epoch”. (B–D) Time (x-axis) at which errors (y-axis) occurred across all sessions and both monkeys relative to: cue onset (B, Attention Epoch), the distractor event onset (C, Filter Epoch), and the target rotation event (D, Choice Epoch). (E) Illustration of the prefrontal cortex and ACC subdivisions where cells were recorded. Details on the reconstruction of recording sites to different subfields are illustrated in Supplementary Figures 3 and 4.

Reconstruction of Recording Sites

The anatomical site of each recorded neuron was reconstructed and projected onto a flat map representation of a normalized, average macaque brain (see Kaping et al. 2011), which allowed assigning each recorded location to a region on a two-dimensional (2D) flat map (Fig. 1E, see also Supplementary Figs 3 and 4, and Fig. 5). We used the area subdivision scheme outlined by Barbas and Zikopoulus (2007) and refer to prefrontal areas 46, 8, and 9 as “lateral prefrontal cortex” (latPFC), area 24 as the ACC, and area 32 as the (ventro-) medial prefrontal cortex (medPFC) (see also Passingham and Wise 2012). Very similar area assignments would follow when considering 2 other major anatomical subdivision schemes of the prefrontal and cingulate cortex (Petrides and Pandya 2007; Saleem et al. 2013; see Supplementary Fig. 4). To briefly summarize the reconstruction steps, we began by projecting each electrodes trajectory onto the 2D brain slice obtained from 7-T anatomical MRIs, using the open-source OsiriX Imaging software and custom-written Matlab programs (Mathworks, Inc., Natick, MA, USA), and utilizing the iodine visualized electrode trajectory within the electrode grid placed within the recording chamber during MR scanning. We drew the coronal outline of the cortical folding of the MR gray scale image to ease the comparison of the individual monkey brain slices to standard anatomical atlases, and to ease using major landmarks the projection of the electrode tip position into the standardized F99 brain available in Caret (Van Essen et al. 2001). Note that we initially reproduced the individual monkey brains within the Caret software to validate similarity and to derive the scaling factors to match the lower resolution monkey MRs to the higher resolution standard F99 brain. We then projected manually and under visual guidance the electrode position to the matched location in the standard F99 brain in Caret. We estimate that the complete procedure from documenting precisely the recording depth, identification of the recording location in the monkeys MR slice, and up to the placement of the electrode position in the standard F99 brain introduces a potential maximal error of 3 mm. However, we felt that despite this potential distortion, which we cannot rule out despite our confidence that the typical (unsystematic) error is more in the <1-mm range, the assignment of recording locations to standard brains is highly beneficial. Note, in particular, that anatomical reconstruction was conducted entirely independent of the analysis of neuronal data and of projecting functional results onto the anatomical 2D map. After identifying all recording sites within the standard F99 brain, we used the Caret software package to render the standard brain into a 3D volume, which was then spherically inflated and cut in order to unfold the brain into 2D space, that is, the flat map (for an illustration of the procedure, see Supplementary Fig. 3).

Data Analysis

Analysis was performed with custom Matlab code (Mathworks), utilizing functionality from the open-source fieldtrip toolbox (Oostenveld et al. 2011). Analysis of spiking activity was based on convolving spiketrains of individual trials with a Gaussian (SD 30 ms) sampled every 5 ms.

Classification of Errors

Analysis of errors was time-aligned to the onset of the saccade that triggered the erroneous trial outcome. For the analysis, we considered errors that were committed during either of 3 time epochs. First, we considered errors committed during the sustained Attention Epoch within 0.3–3 s following attention cue onset (Fig. 1B). These errors occurred during a state of selective attention and may best reflect rather unspecific, motivational lapses of attentional control, which entails lost inhibitory control of eye fixation. The second error type occurred during the Filter Epoch (Fig. 1C). We restricted analysis of these errors to the typical response time window that animals were granted for the true attentional target stimulus (0.05–0.55 s following the distractor event). Errors in this Filter Epoch reflect erroneous bottom-up capture of attention by the salient distractor, or erroneous top-down attention to the wrong stimulus that is of a color different to the attention cue color. In both scenarios, the source of the error is faulty control of top-down attention. The third error type was seen in the Choice Epoch, where the animals either responded to the wrong response target (i.e., made a wrong choice), or incorrectly directed attention to the wrong stimulus. We restricted analysis of errors in the choice period to the time window that would have been allowed for the correct response to occur (Fig. 1D).

Analysis of the Behavioral Consequences of Errors

To quantify the behavioral consequence of the different types of errors, we calculated for each experimental session the accuracy in trials following the errors relative to the accuracy in the trial that preceded the error. Sparse errors (with a cutoff of <10 errors within a session) were not included in the analysis. For each error type, we calculated the proportion of correct responses across the trials that preceded the error trial (Pbefore) and the proportion of correct responses across the trials on the trial following the error (Pafter) using the “Behavioral Adjustment Index” (PafterPbefore)/(Pafter + Pbefore). To compare posterror adjustment across cells, we calculated the difference in accuracy before versus after the error using the index that takes on positive values if accuracy increased following the error, and negative values if the accuracy decreased following errors. We tested the distribution of error adjustment indices against the null hypothesis of zero mean (no adjustment) using t-test statistics.

Classifying Saccade Directions

We categorized the saccade direction of errors in each of the 3 time epochs during the trial relative to the 4 major quadrants (up: 0 ± 45°, left: 90 ± 45°, down: 180 ± 45°, and right: 270 ± 45°) (Fig. 1B–D, bottom panels). For each of the 3 errors, we statistically tested whether erroneous saccades were biased in any particular direction (χ2 test). First, we compared whether saccadic responses up versus down (and right vs. left) were statistically equally likely, and observed no bias in these dimensions. Next, we tested the null hypothesis that saccade directions were equally distributed in the horizontal versus vertical axis, since the response targets were along the vertical axis and the peripheral stimuli on the horizontal line.

Identifying Error-Specific Firing Rate Modulation

We imposed 2 main criteria to identify whether cells encoded error-specific information in their firing. First, firing in response to the error had to be significantly different from pre-error baseline firing. Baseline was defined as the average firing within 300 ms before error onset. Posterror activity was averaged in sliding windows of ±150 ms in steps of 50 ms. Secondly, the activity modulation following errors had to be significantly different from the firing in correct trials. For this comparison, we randomly realigned correct trials so that their activity at time “zero” corresponded to the time when errors were committed (i.e., drawing the alignment times for the times of error commission). For both criteria, we required a cell to show significant modulation for at least 5 consecutive sliding windows in the time windows between 0 and 0.7 s after error onset (P < 0.05, Mann–Whitney U-test). To focus on transient detection processes, we additionally required that there was no significant modulation after 0.7 s after error onset. For some cells with low and no firing in a large subset of trials, we used a nonparametric two-part model for statistical comparison that has been shown to have more power and be more accurate for data with large proportions of discrete (0s) over continuous values (Lachenbruch 2002). For all neuronal analysis, we used only errors committed ≥0.2 s following the onset of cue (Attention Epoch errors), the onset of distractor rotation (Filter Epoch errors), and the onset of target rotation (Choice Epoch errors) in order to prevent confounding influences from transient onset-related effects due to the sensory stimulus change. We further required at least 10 error trials per isolated cell to test for error selectivity and proceed with subsequent analysis. Error modulated firing was either characterized as enhanced or suppressed firing, based on the sign of the firing-rate difference relative to baseline firing.

Anatomical Distribution of Cells with Error-Selective Firing

To test for fine anatomical clustering of error-detecting cells, we calculated the proportion of error-selective cells, relative to all recorded cells, at intersections of a virtual grid within a 4-mm radius of the flat map. The intersections of the grid were 2 mm apart and constituted the center of the pixels shown in the flat map (see below). We only considered pixels in the map that included at least 10 recorded cells. This anatomical mapping of error activity allowed testing for spatial clustering of error detection using a randomization statistics that controlled for uneven sampling of cells across the grid (see Kaping et al. 2011). To test for a higher or lower proportion of error-detecting cells than expected by chance, we calculated N = 2000 random distributions of proportions of cells with a significant effect, after randomly shuffling the location label while maintaining the number of recorded cells per pixel. A cluster was significant if the observed proportion of error-selective cells exceeded the 95th percentile of the random distributions, corresponding to a one-tailed P < 0.05 threshold.

Independent of the fine anatomical mapping, we also compared the proportions of error-detecting cells between 3 major subdivision of the prefrontal-cingulate cortex using multiple comparison corrected (Bonferroni threshold: P < 0.0165) χ2 tests of independence, comparing ACC versus latPFC, latPFC versus medPFC, and medPFC versus ACC.

Latency Analysis of Error-Selective Firing

For each cell with significant error-selective firing, we calculated the latency as the time of maximal difference in post- to pre-error firing baseline. We aimed then to test for significant discrepancies in the latency associated with different response types (enhanced vs. suppressed firing modulations), different area subdivisions (latPFC, ACC, and medPFC), and different error epochs (Attention-, Filter-, and Choice Epochs). However, the small sample size of some cell subsets limited the reliability of the latency distribution (i.e., proportion of error-selective cells with a given latency). Therefore, instead of simply comparing medians from less reliably characterized distributions, we computed the cumulative latency distribution for each cell subset, fitted them with sigmoidal functions, and utilized the C50 parameter of the fits—latency at which the cumulative distribution reaches 0.5—as an improved estimate of the overall cell subset latency.

Analysis of Error-Selective Firing: Enhancement Versus Inhibition

We directly tested for ACC and latPFC whether the proportion of cells with error-locked firing enhancement exceeded, or fell behind, the proportion of cells with error-locked firing inhibition across time relative to error onset. To this end, we first estimated the distribution of cells showing error-specific enhancement/suppression, by differentiating the fitted cumulative distribution of error latencies (see above). The subtraction of both derivatives indexed the relative balance of enhanced and inhibited firing. We evaluated statistical significance using randomization statistics by shuffling the assignment of cells to the enhanced/suppressed cell subset and repeating the procedure 2000 times.

Classification of Putative Cell Types: Spike Waveform Analysis

We aligned, normalized, and averaged all APs for the set of highly isolated neurons (N = 404) of the sample. Each neuronal waveform was then fitted with cubic interpolation from an original precision of 25–2.5 μs. On the resultant waveform, we analyzed 2 measures (Fig. 8A): The peak-to-trough duration and the time for repolarization. The time for repolarization was defined as the time at which the waveform amplitude decayed 25% from its peak value. These 2 measures were highly correlated (r = 0.68, P < 0.001, Pearson correlation). We computed the PCA and used the first component (84.5% of the total variance), as it allowed for better discrimination between narrow (NS)- and broad-spiking (BS) neurons, compared with any of the 2 measures alone. We used the calibrated version of the Hartigan Dip Test (Hartigan and Hartigan 1985) that increases the sensitivity of the test for unimodality (Cheng and Hall 1998; Henderson et al. 2008). Results from the calibrated Dip Test discarded unimodality for the first PCA component (P < 0.01) and for the peak-to-trough duration (P < 0.05), but not for the duration of 25% repolarization (P > 0.05). In addition, we applied Akaike's and Bayesian information criteria for the two- versus one-Gaussian model to determine whether using extra parameters in the two-Gaussian model is justified (Fig. 8A). In both cases, the information criteria decreased (from −669.6 to −808.9 and from −661.7 to −788.9, respectively), confirming that the two-Gaussian model is better. We then used the two-Gaussian model and defined two cutoffs that divided neurons into 3 groups (Fig. 8A). The first cutoff was defined as the point at which the likelihood to be a narrow-spiking cell was 10 times larger than a broad-spiking cell. Similarly, the second cutoff was defined as the point at which the likelihood to be a broad-spiking cell was 10 times larger than a narrow-spiking cell. Thus, 95% of neurons (N = 384) were reliably classified: Neurons at the left side of the first cutoff were reliably classified as NS (20%, N = 79) and those at the right side of the second cutoff were reliably classified as BS (75%, N = 305). The remaining neurons were labeled as “fuzzy” neurons as they fell in between the two cutoffs and were not reliably classified (5%, N = 20).

Results

Behavioral Error Classification

We recorded single neuron activity across a large extent of the lateral prefrontal cortex and medPFC in 2 macaques performing a sustained selective attention task in 131 experimental sessions (63 and 68 for monkeys M and R, respectively; see Kaping et al. 2011). The task required monkeys to use an instructional cue stimulus to covertly shift attention to one of two peripheral stimuli and to maintain spatial attention on that stimulus until it transiently rotated. The rotation triggered a choice on the rotation direction (Fig. 1A). In half of the trials, the nonattended, distracting stimulus transiently rotated before the attended stimulus and thus had to be filtered (ignored) by the animals. Overall, the task was performed at 73% accuracy (monkeys M and R: 75/71%; SDs: 8/11%; proportions are calculated relative to all trials initiated by the animals), with errors falling into 3 major task epochs: first errors committed during sustained spatial attention, but prior to any stimulus rotations, were breaks of fixations without a particular directional bias (n.s., χ2 test) to the response targets or peripheral stimuli (Fig. 1B). These errors fall in the “Attention Epoch” and mostly reflect lapses in inhibitory control of eye fixation during attention, indicative of lapses in the motivational control of attention. Monkeys committed errors in the Attention Epoch in an average of 8% of trials per experimental session (monkey M and R: 11/6%; SD 5/4%). The second epoch with large numbers of errors followed the distractor rotation that had to be filtered from influencing the behavior of the monkeys. Saccadic errors in this Filter Epoch partly reflect erroneous top-down attention to the distractor as inferred from the vertical preference in the saccade directions (Fig. 1C), and bottom-up capture of attention by the peripheral stimuli (Fig. 1C). Monkeys made on average 3% of trials errors in the Filter Epoch (monkey M: 4% and R: 2%; SD: 2/1%). The third type of error was committed within the allowed response time window to the rotation of the attended target stimulus (Fig. 1D). Errors in this “Choice Epoch” were predominantly saccades to the wrong response target located vertically up and down (vertical preference, P < 0.05, χ2) from the fixation point (Fig. 1D). Choice Epoch errors are thus mostly reflecting incorrect sensory-response mappings from rotation directions to saccade directions. Monkeys made on average 6% of trials errors in the Choice Epoch (monkey M/R: 3/10%; SD: 2/4%). An average of 10% of errors (SD 6%, monkey M/R: 7/12%) were made outside of the defined epochs and not considered further.

To test whether the commission of errors in the described task epochs led to an adjustment of attentional performance, we calculated for each experimental session the proportion of correct responses in the trial following an error and compared it to the performance prior to the error. Figure 2A–C illustrates that the average performance significantly improved after errors committed in the Attention Epoch (P < 0.001, t-test; significant for early and late times after cue onset) and the Filter Epoch (P < 0.01, t-test). There was no significant change in performance following errors in the Choice Epoch (Fig. 2D). This behavioral pattern is consistent with the assumption that errors trigger an increase in attentional demand when they likely originate in a failure of attentional control as in both the Attention Epoch and the Filter Epoch (see above) (Shenhav et al. 2013). Notably, overall accuracy levels following these errors remained at a constant (higher) level over subsequent trials, suggesting that these errors are transient lapses that are compensated abruptly following error commission (Fig. 2E). The lack of an apparent adjustment (in terms of average accuracy) following errors in the Choice Epochs may partly relate to the fact that these errors are committed using “full” selective attention on the target stimulus, and thus may not trigger an increased attentional demand to prevent subsequent errors.

Figure 2.

Behavioral adjustment following errors in 3 task epochs. (A–D) Distributions of the Behavioral Adjustment Index showing the performance on the trials after an error relative to the trials prior to an error, for errors committed in the Attention Epoch (A), in the Attention Epoch at early and late times after cue onset (B), in the Filter Epoch (C), and in the Choice Epoch (D). Means (red triangles) and standard errors are given as text. The P-value shows whether the average adjustment is significant (i.e., significantly different from zero). Values larger than zero denote improved accuracy after errors. (B) The distributions of the Behavioral Adjustment Index for errors committed in the Attention Epoch at early times (within 1.5 s after cue onset) and at later times (≥1.5 s after cue onset). (E) Illustration of the Behavioral Adjustment Index for the full sequence of 10 trials following errors (the value for the first trial corresponds to the mean of the histograms shown in A–C). The plot reveals that following errors the performance in the Attention and Filter Epochs improves. Shading denotes SE.

Figure 2.

Behavioral adjustment following errors in 3 task epochs. (A–D) Distributions of the Behavioral Adjustment Index showing the performance on the trials after an error relative to the trials prior to an error, for errors committed in the Attention Epoch (A), in the Attention Epoch at early and late times after cue onset (B), in the Filter Epoch (C), and in the Choice Epoch (D). Means (red triangles) and standard errors are given as text. The P-value shows whether the average adjustment is significant (i.e., significantly different from zero). Values larger than zero denote improved accuracy after errors. (B) The distributions of the Behavioral Adjustment Index for errors committed in the Attention Epoch at early times (within 1.5 s after cue onset) and at later times (≥1.5 s after cue onset). (E) Illustration of the Behavioral Adjustment Index for the full sequence of 10 trials following errors (the value for the first trial corresponds to the mean of the histograms shown in A–C). The plot reveals that following errors the performance in the Attention and Filter Epochs improves. Shading denotes SE.

The observed, average proportion of error trials across task epochs allowed separate neurophysiological analysis of error signals in each of the 3 major task epochs. However, finer-grained analysis of errors as a function of the saccade direction of the fixation break was not feasible due to the low number of trials (see polar plots in Fig. 1BD). We observed on average across cells 6.3, 3.7, and 7.1 error trials per directional quadrant in the Attention Epoch, the Filter Epoch, and the Choice Epoch, respectively.

Selective Error-Detection Signals in Neuronal Firing

We next tested whether single cells encoded error outcomes in the 3 task epochs described above. Figure 3A–C shows that we found a similar proportion of cells that transiently modified their firing upon error commission in each of the trial epochs (for example cells, see Supplementary Fig. 1): for the Attention Epoch, of all (N = 867) recorded cells with at least 10 error trials 128 (15%) cells showed error selectivity; for the Filter Epoch, the recordings of 544 cells included at least 10 error trials. From these cells, 77 (15%) showed error selectivity; and for the Choice Epoch, the recordings of 728 cells included at least 10 error trials. From these cells, 84 (12%) showed error selectivity. The error-locked firing modulations were evident not only in increased firing, but for a substantial number of neurons (44% of error-selective cells) became evident through transient posterror response inhibition (Fig. 3AC).

Figure 3.

Error-related firing-rate modulations for errors in 3 different task epochs. (A) Normalized firing of cells that increase (upper panels) and decrease (lower panels) their firing rate transiently following error commission in the Attention Epoch (see Fig. 1A). Dashed and solid lines correspond to error and correct trials. (B and C) Same format as in A, but for the set of cells with significant activity modulation in response to error commissions in the Filter Epoch (B), and the Choice Epoch (C). Normalization of single cell firing rates used the formula [Rate – min(Rate)/max(Rate) – min(Rate)]. Gray shading shows SE.

Figure 3.

Error-related firing-rate modulations for errors in 3 different task epochs. (A) Normalized firing of cells that increase (upper panels) and decrease (lower panels) their firing rate transiently following error commission in the Attention Epoch (see Fig. 1A). Dashed and solid lines correspond to error and correct trials. (B and C) Same format as in A, but for the set of cells with significant activity modulation in response to error commissions in the Filter Epoch (B), and the Choice Epoch (C). Normalization of single cell firing rates used the formula [Rate – min(Rate)/max(Rate) – min(Rate)]. Gray shading shows SE.

Figure 4.

Proportion of cells encoding error outcomes in more than one task epoch. (A) Proportion of cells (y-axis) that encode errors in one task epoch (as indicated on the x-axis) and in at least one other epoch (“joint encoding” cells). The dark gray bars reflect join encoding for cells that increased their firing upon errors. The light gray bars denote cells showing transient firing suppression following errors. (B) The distribution of joint encoding of errors for cells that increase firing upon error commission. The length of arrows in each plot denotes the proportion of cells that jointly encode errors in the epoch where the arrows originate and where they point to. The sketched circles on the left illustrate the main result of partly overlapping error types. (C) Same as B for cells with firing-rate suppression upon error commission.

Figure 4.

Proportion of cells encoding error outcomes in more than one task epoch. (A) Proportion of cells (y-axis) that encode errors in one task epoch (as indicated on the x-axis) and in at least one other epoch (“joint encoding” cells). The dark gray bars reflect join encoding for cells that increased their firing upon errors. The light gray bars denote cells showing transient firing suppression following errors. (B) The distribution of joint encoding of errors for cells that increase firing upon error commission. The length of arrows in each plot denotes the proportion of cells that jointly encode errors in the epoch where the arrows originate and where they point to. The sketched circles on the left illustrate the main result of partly overlapping error types. (C) Same as B for cells with firing-rate suppression upon error commission.

Cells may show error-locked responses only for errors in one task epoch, or they could generalize and signal errors across epochs. We quantified this ‘tuning’ to specific errors in different epochs by calculating for each the proportion of cells that showed joint error-locked response modulation (Fig. 4). We defined “joint error selectivity” of a cell when it significantly encoded an error in one task epoch and in at least one other task epoch. Joint selectivity of error outcomes across epochs was evident in 28%, 33%, and 24% of the cells that were error-selective in the Attention-, Filter-, and Choice-Epochs, respectively (Fig. 4A). Joint error coding through response inhibition was less equally distributed compared with response enhancement. Figure 4A shows that joint error coding in 2 epochs ranged from 9% to 28% among cells with error-locked response inhibition. Thus, less than one-third of cells encoding errors in one epoch encoded errors in another task epoch. Figure 4B,C illustrates the specific combinations of task epochs to which single cells showed error-locked firing increases (Fig. 4B) and decreases (Fig. 4C). Joint error selectivity through response enhancement was similarly likely for all combinations of error types. In contrast, cells showing response inhibition following errors in the Choice Epochs were largely unaffected by errors committed during the Attention and Filter Epochs (Fig. 4C).

Functional Topography of Error Detection

We next tested whether neurons encoding errors in a particular task epoch were anatomically located in specific subareas within the prefrontal-cingulate cortex of the macaques. For this purpose, we reconstructed the recording locations of error-detecting neurons and projected them onto a 2D flat map of the prefrontal-cingulate cortex (see Materials and Methods). Across the 3 major subdivisions of the prefrontal-cingulate cortex that we recorded from, the ACC (area 24) hosted the largest proportion of error cells (Fig. 5A). For a finer-grained anatomical depiction of error-detecting cell groups, we plotted data-driven functional maps of the anatomical distribution of error-selective cells (Fig. 5B). The salient cluster in area 24 (Fig. 5B, randomization test, P < 0.05) is highly consistent with existing evidence about the functional role of the ACC in error detection and reveals its anatomical specificity to encode error outcomes (Alexander and Brown 2011; Shenhav et al. 2013). Beyond the ACC, we found small satellite spots with significantly higher proportions of error-detecting neurons than expected by chance in dorsomedial PFC (area 9) and at the border of medial frontal area 32 (Fig. 5B, randomization test, P < 0.05). Taken together, these findings show first that error detection is particularly prominent in the ACC with >25% (and locally up to 40%) of cells signaling error outcomes across task epochs in an attention task. Secondly, the results also show that error signals are not strictly confined to the ACC.

Figure 5.

Anatomical distribution of error-encoding cells in the prefrontal cortex and ACC. (A) The overall proportion of cells with significant error-related firing-rate modulation in the lateral prefrontal cortex (latPFC, areas 46, 8, and 9), the ACC (area 24), and the (ventro-) medPFC (area 32). The contour map shows these areas as patches on a 2D flat map (see also Fig. 1E). (B) Fine-grained functional topography of the proportion of error-encoding cells (color scale) in prefrontal and cingulate cortex. Significant spatial clustering, particularly within the ACC (area 24), is shown in the contour map to the right. (C–E) Same format as in B for cells that signal error outcomes in the Attention Epoch (C), in the Filter Epoch (D), and in the Choice Epoch (E). The bar histograms show the average proportions of error-encoding cells per subarea. The small contour maps show significant spatial clustering of error-encoding cells for each error type. The color bars are scaled with the distribution medians as the intermediate (white) color.

Figure 5.

Anatomical distribution of error-encoding cells in the prefrontal cortex and ACC. (A) The overall proportion of cells with significant error-related firing-rate modulation in the lateral prefrontal cortex (latPFC, areas 46, 8, and 9), the ACC (area 24), and the (ventro-) medPFC (area 32). The contour map shows these areas as patches on a 2D flat map (see also Fig. 1E). (B) Fine-grained functional topography of the proportion of error-encoding cells (color scale) in prefrontal and cingulate cortex. Significant spatial clustering, particularly within the ACC (area 24), is shown in the contour map to the right. (C–E) Same format as in B for cells that signal error outcomes in the Attention Epoch (C), in the Filter Epoch (D), and in the Choice Epoch (E). The bar histograms show the average proportions of error-encoding cells per subarea. The small contour maps show significant spatial clustering of error-encoding cells for each error type. The color bars are scaled with the distribution medians as the intermediate (white) color.

Topographic Clustering of Cells Encoding Errors in Different Task Epochs

One reason for the anatomical clustering of error signals in frontal areas beyond the ACC could be an anatomically specific tuning to errors committed in selective task epochs. As shown above (Fig. 4), more than two-thirds of error-encoding neurons were exclusively selective for one error type. We therefore analyzed the proportion of error-encoding neurons for each task epoch separately (Fig. 5C–E). For the Attention Epoch, the overall proportion of error tuned cells did not vary between the latPFC, medPFC, and ACC, but fine-grained functional mapping revealed a significant cluster of error-encoding cells in latPFC areas 46 and 8 (Fig. 5C, randomization test, P < 0.05). Error encoding during the Filter Epoch was evident in up to 31% of cells in local clusters within the ACC (area 24) (Fig. 5D, randomization test, P < 0.05). Errors committed during the Choice Epoch were similarly encoded in local clusters of cells within the ACC (Fig. 5E, randomization test, P < 0.05). The anatomical clusters of error-detecting cells in the Filter and Choice Epochs overlapped only partly, suggesting that largely independent populations of ACC cells encode errors in each epoch (see below).

Anatomical Dissociation of Error-Locked Enhancement and Inhibition

In principle, the functional topographies described above could reflect circuits of cells “genuinely” tuned to detect error outcomes by increased firing. However, there could also be anatomical clusters where cell activity reflects error commission only indirectly, for example, through transient disengagement from ongoing processes or through inhibition coming from other cells that encode error outcomes (see also Fig. 3). We tested whether these possibilities may be realized in the prefrontal-cingulate cortex by investigating the anatomical distribution of neurons that showed error-locked inhibition independent from neurons that signal errors through response enhancement. Figure 6A shows that, across ACC and all PFC subregions, the proportion of neurons that transiently increased their firing upon error commission was statistically indistinguishable across different task epochs. In contrast, error-locked inhibition varied within the different PFC subregions as a function of the task epoch (Fig. 6A). Within the ACC, error-locked inhibition was more than twice as likely to occur in the Choice Epoch than in the Attention Epoch (P < 0.016, multiple comparison corrected); the opposite pattern was found within the lateral PFC (P < 0.016, multiple comparison corrected). As illustrated in Figure 6B, in the latPFC, a large proportion of cells (up to 18% of cells in local clusters) showed transient error-locked inhibition in the Attention Epoch, whereas there were virtually no error-signaling cells with response inhibition in the Choice Epoch. Analogously, large proportions of cells in the ACC (up to 21% of cells) showed error-locked inhibition in the Choice Epoch, whereas virtually no cells signaled errors through response inhibition in the Attention Epoch. To validate this double dissociation (ACC–latPFC and Attention–Choice Epoch), we evaluated significant spatial clustering of cells with a permutation test that corrected for uneven sampling of cells across the flat map. This analysis (Fig. 6C) confirmed that a larger proportion of cells than expected by chance (P < 0.05) showed error-locked inhibition during the Choice Epoch within the ACC, and during the Attention Epoch within the latPFC (areas 8 and 46d).

Figure 6.

Anatomical distribution of cells with enhanced and suppressed firing upon errors across anterior cingulate and prefrontal cortices. (A) Proportion of cells encoding error outcomes with enhanced firing (dark gray bars) and suppressed firing (light gray bars). Significantly different proportions are indicated with a star. (B and C) Finer-grained functional topography of cells with suppressed firing following errors in the Attention Epoch (top), the Filter Epoch (left bottom), and the Choice Epoch (right bottom). The double pointing arrows connect the spatial clusters, where there were significantly larger proportions of error encoding cells than expected by chance as shown in the clustering map in C. Red and blue bounded gray patches denote cell clusters with error-locked inhibition in the Attention Epoch (red) and Choice Epoch (blue). The color bars in B are scaled with the distribution medians as the intermediate (white) color.

Figure 6.

Anatomical distribution of cells with enhanced and suppressed firing upon errors across anterior cingulate and prefrontal cortices. (A) Proportion of cells encoding error outcomes with enhanced firing (dark gray bars) and suppressed firing (light gray bars). Significantly different proportions are indicated with a star. (B and C) Finer-grained functional topography of cells with suppressed firing following errors in the Attention Epoch (top), the Filter Epoch (left bottom), and the Choice Epoch (right bottom). The double pointing arrows connect the spatial clusters, where there were significantly larger proportions of error encoding cells than expected by chance as shown in the clustering map in C. Red and blue bounded gray patches denote cell clusters with error-locked inhibition in the Attention Epoch (red) and Choice Epoch (blue). The color bars in B are scaled with the distribution medians as the intermediate (white) color.

Relative Spike Timing of Error-Locked Responses

We next tested the relative timing of error-locked enhancement and inhibition. For each cell with significant error-selective firing, we first calculated the latency of maximal error-locked response modulation, and then compared the proportion of cells with error-locked firing increases versus the proportion of cells with error-locked firing inhibition. We restricted this analysis to cells in the ACC and the latPFC, as our aim was on explaining the dissociation between both areas described above (Fig. 6B). Figure 7A–C illustrates the cumulative distribution of error-specific response enhancement and inhibition for cells. The cumulative distributions were well fitted by a sigmoidal function (or, in one case, with the combination of 2 sigmoid curves, see Materials and Methods), which allowed using the C50 parameter of the fits—latency at which the cumulative distribution reached 0.5—as a good estimate of the latency for error encoding per subarea (see Materials and Methods). Figure 7D,E summarizes the latencies of error signaling through response enhancement and response inhibition, illustrating 2 main findings: first, error-locked firing increases significantly preceded in time error-locked inhibition in all task epochs and brain areas with only one exception, namely the latPFC error signaling in the Filter Epoch showed similar timing of firing increases and inhibitions.

Figure 7.

Latency of error encoding in the ACC and lateral PFC. (A–C) Cumulative latency distributions (y-axis) in the Attention Epoch (A), the Filter Epoch (B), and the Choice Epoch (C), with respect to time (x-axis). Raw data (small point or cross-symbols) were fitted with a sigmoidal function (see text and Materials and Methods for details). Green/blue lines show results from ACC and lateral PFC, respectively. Solid and dashed lines show cumulative latency distributions for cells with enhanced (solid) and suppressed (dashed) firing to errors. (D) The latency of cells with error-related response inhibition (y-axis) against the latency of cells with error-related response enhancement (x-axis). Latency is estimated by the C50 parameter of the sigmoidal fits shown in (A–C). Vertical and horizontal error bars denote the 95% confidence intervals around the C50 estimates. (E) The table lists the latencies (in seconds) from D for cells with error-locked enhancement (first number) and inhibition (second number). (F–H) Difference in the proportion of error-selective cells that showed firing enhancement versus inhibition within ACC and latPFC in the Attention Epoch (F), the Filter Epoch (G), and the Choice Epoch (H) (see Material and Methods). The green (blue) shading denotes the time intervals for which the difference in proportion between enhancement and inhibition neurons was significantly (P < 0.05) different from zero (null hypothesis) in the ACC (latPFC).

Figure 7.

Latency of error encoding in the ACC and lateral PFC. (A–C) Cumulative latency distributions (y-axis) in the Attention Epoch (A), the Filter Epoch (B), and the Choice Epoch (C), with respect to time (x-axis). Raw data (small point or cross-symbols) were fitted with a sigmoidal function (see text and Materials and Methods for details). Green/blue lines show results from ACC and lateral PFC, respectively. Solid and dashed lines show cumulative latency distributions for cells with enhanced (solid) and suppressed (dashed) firing to errors. (D) The latency of cells with error-related response inhibition (y-axis) against the latency of cells with error-related response enhancement (x-axis). Latency is estimated by the C50 parameter of the sigmoidal fits shown in (A–C). Vertical and horizontal error bars denote the 95% confidence intervals around the C50 estimates. (E) The table lists the latencies (in seconds) from D for cells with error-locked enhancement (first number) and inhibition (second number). (F–H) Difference in the proportion of error-selective cells that showed firing enhancement versus inhibition within ACC and latPFC in the Attention Epoch (F), the Filter Epoch (G), and the Choice Epoch (H) (see Material and Methods). The green (blue) shading denotes the time intervals for which the difference in proportion between enhancement and inhibition neurons was significantly (P < 0.05) different from zero (null hypothesis) in the ACC (latPFC).

Secondly, error-locked firing increases in the ACC preceded error-locked signaling in the lateral PFC in all task epochs. This finding of the earliest latency of detection signals in the ACC corroborates the special role of the (dorsal) ACC to detect errors in the attention task (Fig. 5). In summary, error detection is fastest in the population of ACC cells that increases firing upon error commission, followed by transient, more homogeneous response inhibition in the ACC and the lateral PFC.

We further validated these conclusions with an additional analysis that, unlike the cumulative distributions, takes into account the unequal proportion of cells signaling distinct error types in different prefrontal regions through either enhancement or inhibition (Fig. 7F–H). For this analysis, we subtracted the estimated distribution for the proportions of cells showing error-locked firing increases with respect to the cells showing firing decreases, at each time point relative to the error onset (see Materials and Methods). As shown in Figure 7F–H, an error-net locked increased activity becomes evident in time, clearly emerging earliest and strongest in the ACC compared with the lateral PFC in the Attention Epoch, the Filter Epoch, and the Choice Epoch.

Putative Neuron Types Underlying Error Detection and Error-Locked Inhibition

The previous results illustrate that transient posterror inhibition follows in time the error-locked firing increase. This finding suggests that posterror enhanced firing is instrumental in triggering the posterror response inhibition. This would predict that error signaling relies to a large extent on inhibitory interneurons whose error-locked activation will impose error-specific inhibition on the connected pyramidal cell populations (Medalla and Barbas 2012).

We tested this prediction by classifying the set of N = 404 maximally isolated neurons into putative inhibitory interneurons and putative pyramidal cells according to their AP waveform parameters (Fig. 8A, see Materials and Methods). Using the peak-to-trough duration and the time of 25%—repolarization of the cells’ APs provided a bimodal distribution of waveform parameters with a clean separation of “NS” (20%, N = 79), “BS” (75%, N = 305), and 5% (N = 20) of neurons without unequivocal classification (which we labeled as fuzzy cells). Of the classified cells, we found that N = 70 cells showed significant error-locked firing modulation (split in 38 and 32 cells with error-locked enhancement and suppression, respectively), comprising N = 11 NS cells (16%), N = 55 BS cells (79%), and N = 4 fuzzy cells (Fig. 8B). The overall proportion of NS-to-BS cells did not differ significantly between lateral PFC (NS: N = 4, 14%; BS: N = 22, 79%), medial PFC (NS: N = 2, 10%; BS: N = 18, 90%), and ACC (NS: N = 5, 23%; BS: N = 15, 68%). Figure 8C documents that, on average, the population of cells with error-locked response enhancement contained 74% BS cells (N = 28) and 18% (N = 7) NS cells. Similarly, overall proportions of 84% BS cells (N = 27) and 13% NS cells (N = 4) showed error-locked response inhibition. However, considering the subareas of recorded cells revealed that a significantly larger proportion of narrow-spiking cells encoded errors in the ACC than in the lateral PFC through response enhancement (Fig. 8CE, randomization statistics, P < 0.05). This finding is consistent with the outlined scenario that error signaling in the ACC involves a significant fraction of putative inhibitory interneurons who use their error information to inhibit connected cells within ACC and possibly within the connected lateral PFC (Medalla and Barbas 2009, 2012; Morecraft et al. 2012).

Figure 8.

Putative cell types encoding errors in the ACC and lateral PFC. (A) Bimodal distribution of “narrow-spiking” and “broad-spiking” cells (and unreliably classified spiking, “fuzzy” cells in between) as indexed by a combination of AP waveform parameters (indexed as a principal component score, see Materials and Methods): The peak-to-trough duration and time to 25% repolarization. The middle and right panels show the average normalized waveforms across NS and BS cells. (B) Normalized waveforms of all recorded single cells (left) and of those single cells that encoded error outcomes by enhanced and suppressed firing (middle and right panels, respectively). Red/blue-colored waveforms reflect grouping into NS/BS cells. (C) Overview of the number of NS and BS cells encoding error in the latPFC, medPFC, and ACC. (D) Proportion of NS cells within the classes of error-encoding cells in the lateral PFC, medPFC and ACC. The difference between proportions of error-selective NS cells in latPFC and ACC was statistically significant for cells with enhanced firing following an error. (E) Shuffled random distribution of the relative difference in the proportion of NS cells increasing firing upon error commission between latPFC and ACC. The observed difference is statistically lower than expected by chance, showing that there are more putative inhibitory interneurons in the ACC that encode errors through enhanced firing (the comparison is highlighted in bold font in the table in C).

Figure 8.

Putative cell types encoding errors in the ACC and lateral PFC. (A) Bimodal distribution of “narrow-spiking” and “broad-spiking” cells (and unreliably classified spiking, “fuzzy” cells in between) as indexed by a combination of AP waveform parameters (indexed as a principal component score, see Materials and Methods): The peak-to-trough duration and time to 25% repolarization. The middle and right panels show the average normalized waveforms across NS and BS cells. (B) Normalized waveforms of all recorded single cells (left) and of those single cells that encoded error outcomes by enhanced and suppressed firing (middle and right panels, respectively). Red/blue-colored waveforms reflect grouping into NS/BS cells. (C) Overview of the number of NS and BS cells encoding error in the latPFC, medPFC, and ACC. (D) Proportion of NS cells within the classes of error-encoding cells in the lateral PFC, medPFC and ACC. The difference between proportions of error-selective NS cells in latPFC and ACC was statistically significant for cells with enhanced firing following an error. (E) Shuffled random distribution of the relative difference in the proportion of NS cells increasing firing upon error commission between latPFC and ACC. The observed difference is statistically lower than expected by chance, showing that there are more putative inhibitory interneurons in the ACC that encode errors through enhanced firing (the comparison is highlighted in bold font in the table in C).

Error-Locked Firing and Oculomotor Activity

Previous studies have shown that error-related activity in the ACC and adjacent prefrontal cortex is largely separate, and goes beyond, motor-related activation (Ito et al. 2003; Hyman et al. 2013). To relate to these findings, we tested whether the error-locked modulation that we have described is separated from saccade-related activity modulations.

For this analysis, we aligned the data to the saccade onset in correct trials and compared it to the saccade onset aligned activity modulation on error trials. Figure 9 shows across task epochs and error-locked response types the maximal difference in postsaccadic activity on error trials versus correct trials. Cells that significantly increased their firing upon error commission showed on average a stronger activity increase on error-aligned trials compared with postsaccadic modulation on correct trials (P < 0.05 for all error types, Wilcoxon rank test, see Fig. 9A–C). This finding is notable, because the postsaccadic response of 84% of these cells showed already significant modulation on correct trials (Wilcoxon rank test relative to baseline, P < 0.05). An identical response pattern was found for cells with error-locked inhibition, which showed on average significantly stronger firing inhibition on error trials than on saccade-aligned correct trials (P < 0.05, for all error types). This difference was evident despite the finding that 75% of the neurons with significantly decreased error-locked modulation also had a significant postsaccadic modulation on correct trials (Wilcoxon rank test relative to baseline, P < 0.05). In summary, error-locked firing modulation often co-occurs with postsaccadic-related modulations, but shows significantly larger modulations following errors.

Figure 9.

Error-related modulation exceeds oculomotor aligned activity modulation. (A–C) Comparison of maximal firing after errors (y-axis' alignment is the break-fixation saccade on error trials) and after saccadic responses on correct trials (x-axis). (A–C) Results for the Attention Epoch, the Filter Epoch, and the Choice Epoch are shown. Left panels show results for the population of cells with enhanced firing after errors and right panels shows results for cells with suppressed firing following errors. Statistical P-values are based on paired Wilcoxon signed-rank test.

Figure 9.

Error-related modulation exceeds oculomotor aligned activity modulation. (A–C) Comparison of maximal firing after errors (y-axis' alignment is the break-fixation saccade on error trials) and after saccadic responses on correct trials (x-axis). (A–C) Results for the Attention Epoch, the Filter Epoch, and the Choice Epoch are shown. Left panels show results for the population of cells with enhanced firing after errors and right panels shows results for cells with suppressed firing following errors. Statistical P-values are based on paired Wilcoxon signed-rank test.

Error-Locked Firing and its Relation the Omission of Expected Reward

So far, our main analysis identified errors strictly aligned to the time of the erroneous saccadic response of the animals. For errors of the Attention Epoch, this erroneous response was the fixation break with no apparent directional bias (Fig. 1B). However, for subsets of errors in the Filter and Choice Epochs, the erroneous response could have been a saccadic response toward a regular response target within the regular response time window that we allowed during correct trials (Fig. 1C,D). For these trials, there is thus no overt indication to the animal that an error occurred until the time when the reward would be expected but failed to occur. In our task, the reward was given on correct trials with a fixed delay of 0.4 s following a correct saccadic response, suggesting that the omission of the expected reward becomes apparent following 0.4 s after the erroneous response. Thus, a certain proportion of the cells that were identified as detecting errors in the Filter and Choice Epochs might have been influenced by reward omission. Additionally, a subset of cells could have contributed to error detection by selectively responding to reward omission (see Ito et al. 2003), and we hypothesized that these signals should anatomically cluster within the ACC.

In fact, we identified that about 30% of error-encoding cells described above were also modulated significantly when the analysis was aligned to the time of reward omission (0.4 s after the response). In particular, among all the cells with error selectivity in the Filter Epoch, 28.6% (N = 22 of 77) showed significant modulation following the time of reward omission. For the Choice Epoch, 37% (N = 31 of 84) of cells with error selectivity showed significant outcome selectivity when tested time-aligned to the omission of expected reward. We also found a separate population of cells that responded to the erroneous outcome only after the time of expected reward (in the Filter Epoch N = 34, and in the Choice Epoch N = 32). The average responses of these cells are illustrated in Supplementary Figure 5A,B. Remarkably, cells that responded closely time-locked to the reward omission during the Filter and Choice Epochs were most prevalent in local clusters within the ACC (Supplementary Fig. 5C,D). Nonparametric statistics that controlled for possible sampling biases showed significant (P < 0.05) anatomical clustering for both error types at locations that were anatomically close to the locations that showed the highest proportions of cells with error selectivity emerging earlier than the time of reward omission (Supplementary Fig. 5C,D).

Discussion

We have shown that a large proportion of cells in the ACC respond transiently to the commission of errors in an attention task across 3 epochs that varied in their specific attentional processing demands. The error-encoding cell populations clustered predominantly within the dACC (area 24c) and showed 5 main characteristics: first, they were partly selective for specific task epochs reflecting different processing demands. Secondly, they emerged with the shortest latency across the prefrontal-cingulate cortex. Thirdly, they were separated from oculomotor-related modulations. Fourthly, they emerged earlier than and largely independent from reward omission signals. Fifthly, they were comprised of a significant fraction of putative inhibitory interneurons.

Consistent with the prevalence of putative interneurons to encode errors in the ACC, we found that error-selective response modulation is apparent not only by firing-rate enhancement, but also by selective inhibition of firing in large populations of cells in the ACC and the lateral PFC. Across all prefrontal-cingulate cortical subareas, error-locked inhibition followed in time error-locked enhancement, suggesting that transient inhibition does not reflect genuine error detection, but rather indexes a transient disengagement process that follows error detection and could reflect the transition to later outcome evaluation processes (Rothe et al. 2011). Consistent with such an interpretation, we identified selective error-locked inhibition in a specific cluster of cells in lateral PFC that did not otherwise show particular firing increases to errors. Cells in this lateral PFC cluster (spanning the border of areas 46 and 8) were thus most likely showing a transient disengagement response following errors in the Attention Epoch of the task. The origin of such a transient inhibition may be those cells that detected the erroneous outcome in the first place, including local cells from within the lateral PFC, but likely including those cells in the ACC that showed the earliest positive error signals (via response enhancement) across all types of analyzed errors (see below). The relative delay in lateral PFC inhibition (Fig. 7D,E) is consistent with the second account of a nonlocal influence arriving from those ACC cell populations that showed a widespread increased firing upon error commission. Consistent with this scenario, a recent study has shown that outcome-specific high-frequency gamma activity in the local field potential of ACC sites correlates with the high gamma amplitude envelopes within the lateral PFC at a temporal delay (Rothe et al. 2011). Moreover, functional ACC-to-lateral PFC inhibition is anatomically supported by specialized inhibitory connectivity originating from the ACC (Medalla and Barbas 2009, 2010, 2012). A note of caution should be raised with respect to the longer latencies of error-locked firing suppression, because they could partly be overestimated because response suppression is typically of lower magnitude than response enhancement and it is thus more difficult to statistically detect onsets of suppression (in our sample, the absolute modulation depth of response suppression was overall lower than the modulation depth for error-locked response enhancement by 1.9–7 Hz firing rate across error epochs).

Taken together, these findings suggest that the neuronal circuitry in the ACC plays a pivotal role to encode errors that signal failures in attentional control processes. These error signals are the prerequisites for an organism to adjust control demands of ongoing processing, as recently formalized by the Expected Value of Control framework of ACC function (Shenhav et al. 2013). The following discusses these processing aspects and their anatomical specificity within the ACC and PFC.

Process Specificity of Error Outcome Signals in the ACC and Lateral PFC

Our findings suggest that the commission of errors in maintaining focused attention and failures to filter distracting bottom-up inputs triggers a sequence of events that starts with error-specific firing-rate increases in populations of ACC cells. Following initial error detection in the ACC, (1) local firing was transiently suppressed in the lateral PFC when the error was committed during focused attention, (2) local firing was transiently suppressed in the ACC following erroneous choices, and (3) local firing in ACC and lateral PFC is similarly suppressed when errors were committed during the filtering of distractors (Fig. 6). This pattern of results was unexpected and followed from data-driven topographic mapping of error-locked response modulations. We believe that it conveys novel insights about the functional specialization of attentional control processes within the prefrontal-cingulate cortex under the assumption that the source of errors as initially detected within the ACC is conveyed to those prefrontal-cingulate cells that are concerned with resolving and preventing a similar processing error in future trials (Womelsdorf et al. 2010; Rothe et al. 2011; Khamassi et al. 2013; Shenhav et al. 2013).

In particular, Figure 10 illustrates that the Attention Epoch represents a processing stage that places particular strong control demands on the oculomotor system (to prevent overt movements during covert attention) and on motivational perseverance (to sustain the attentional focus). When these processes fail, as evident by error commission in the Attention Epoch, the organism ideally recalibrates those circuits that realize these control processes (Khamassi et al. 2013). The inhibition in the lateral PFC following these errors may thus be linked to the process of recalibration or re-evaluation of the invested motivational efforts and the top-down task control of oculomotor centers, such as the frontal eye field (FEF) or the superior colliculus (Fig. 10) (Kaping et al. 2011; Everling and Johnston 2013).

Figure 10.

A process-specific perspective of attentional performance and adjustment following errors. The figure explicitly associates specific processing stages (task epochs, left column) with specific functional demands (middle column). As a consequence of the link between task epochs and functions, errors committed in a task epoch specific are hypothesized to give rise to adjustments of specific processing aspects as outlined in the rightmost column.

Figure 10.

A process-specific perspective of attentional performance and adjustment following errors. The figure explicitly associates specific processing stages (task epochs, left column) with specific functional demands (middle column). As a consequence of the link between task epochs and functions, errors committed in a task epoch specific are hypothesized to give rise to adjustments of specific processing aspects as outlined in the rightmost column.

In contrast to errors in the Attention Epoch, erroneous sensory-motor mapping performed on a fully attended stimulus in the Choice Epoch did not result in inhibition within the lateral PFC, but in localized, delayed inhibition within the dorsal ACC. According to the “process specificity” hypothesis of error processing, this response pattern suggests that processes to prevent these “choice error” in future trials reside predominantly within ACC circuits (Shenhav et al. 2013). Consistent with this hypothesis, the ACC is implicated to convey specific information supporting posterror adjustments of stimulus-response mapping rules and of behavioral strategies that optimize reward harvesting by choices in future trials (Kennerley et al. 2006; Procyk and Goldman-Rakic 2006; Matsumoto et al. 2007; Quilodran et al. 2008; Womelsdorf et al. 2010; Hayden et al. 2011; Rothe et al. 2011).

Similar to the Attention and Choice Epochs, the errors in the Filter Epoch were followed by response suppression in a large number of neurons. However, Figure 6B revealed that inhibition was not as strictly localized, but evident in about 10% of cells across a larger extent of the lateral PFC as well as in the ACC. This response pattern could reflect that errors in the Filter Epoch originate largely in failures of a function that is subserved by the distributed cell population across the prefrontal-cingulate cortex. We have indeed shown previously that attentional filtering, that is, the control of interference, recruits a wide band of cells spanning lateral to medial prefrontal-cingulate cortices (see Fig. 5 in Kaping et al. 2011). Beyond the failure of filtering the bottom-up change of the distractor stimulus, a fraction of errors in the Filter Epoch will originate in a faulty attentional allocation to the wrong stimulus. We could not separate these 2 major sources of erroneous responses in the current analyses (due to too small number of trials per subcondition and cell), but suggest that future studies may reveal a functional specificity of neuronal signals detecting on the one hand, failures in interference control, and on the other hand failures in correct attentional focusing.

Such a hypothesized division of labor in attentional control processes has long thought to be realized by an interplay of cells in the ACC and the lateral PFC: According to the Expected Value of Control framework, the ACC specifies the control requirements for a task and monitors performance (Botvinick 2007; Shenhav et al. 2013). In contrast, the lateral PFC implements top-down control according to the strength and task context specified by ACC circuitry (Holroyd and Yeung 2012; Shenhav et al. 2013). These complementary roles of ACC and lateral PFC for task control predict a specific temporal relation of activation: The ACC should become error selective earlier and its error detection should serve as an interrupt signal that disengages the lateral PFC circuitry from ongoing processes and possibly triggers the transitioning to evaluation and adjustment processes (Kerns et al. 2004; Shenhav et al. 2013) (Fig. 10). This theoretically predicted temporal dynamics is what we have shown in the latency analysis in each brain area separately. Future studies will be needed to resolve how and when the ACC and lateral PFC cell populations exchange information and mutually influence each other during outcome processing. Our data did not provide sufficient simultaneous interareal recordings of cells with error-locked transient response modulations to allow such direct testing.

Anatomical Specificity of Error Encoding in the ACC

Single neuron outcome signals that relate to either reward prediction errors, surprising outcomes, choice errors, or errors of commission have been reported in cells recorded in the dorsal ACC (areas 24c and 24b) (Ito et al. 2003; Amiez et al. 2005; Matsumoto et al. 2007; Quilodran et al. 2008; Sallett et al. 2007; Hayden et al. 2009, 2011; Kennerley et al. 2011). Our finding of an apparent local clustering of error-encoding cells in ACC (area 24), rather than in areas 32, 10, 8, 9, or 46, provides strong corroborative evidence of the highly specialized role of ACC circuitry to detect errors (Fig. 5) (Holroyd et al. 2009; Khamassi et al. 2013). We observed in local anatomical spots in the ACC up to 40% of error-encoding cells when all task epochs were considered together. This number goes down to an average 10–15% of error-selective firing cells when error encoding in individual task epochs is considered. These findings resonate well with previous findings reporting of 7–15% of ACC cells encoding specific outcomes such as reward prediction errors or choice errors in nonhuman primates (Matsumoto et al. 2007; Quilodran et al. 2008; Kennerley et al. 2011) and similar proportions in rodent ACC (Totah et al. 2009; Bryden et al. 2011, but see Narayanan and Laubach 2008 for higher proportions when cells are included with sustained error-specific firing modulation). Taken together, the documented error-locked firing suggests that separate groups of cells in the ACC are tuned to respond to outcomes originating from failures in different processing regimes. This finding can be readily seen in single cell firing patterns as illustrated in process-specific error-locked modulations of 3 example cells (Supplementary Fig. 2). Our findings are thus strongly supporting the notion of process specificity of error-locked firing by showing significant clustering of cells encoding errors during sustained focusing of attention, filtering, and choice processes in anatomically partly nonoverlapping regions of the ACC.

To our knowledge, such an anatomical dissociation of encoding different types of errors has only rarely been reported (Quilodran et al. 2008; Kennerley et al. 2011). In human Electroencephalogram (EEG) studies, slightly different scalp EEG topographies and source locations of error-related negativity amplitudes have been reported for errors of different origins (erroneous motor executions vs. suboptimal choices) (Grundler et al. 2009; Cavanagh et al. 2010; see also Holroyd et al. 2009). However, the human EEG literature may best serve to illustrate the generality of the error outcome encoding of the ACC. Error- and feedback-locked potentials with a generating source in the ACC are observed in a wide variety of tasks, independent of the (sensory) input modality, and independent of the (motor) output modality of the erroneous task outcome (Debener et al. 2005; Mitchell et al. 2008; Riesel et al. 2013). Importantly, error-related potentials are similarly evident in the macaques, documenting that the human EEG findings will be tightly linked to the macaque functional electrophysiology (Godlove et al. 2011). Our results of attention-specific error encoding in the ACC add to this overall picture that implicates the ACC circuitry as entailing a general error-detection system for multiple types of errors (see also Alexander and Brown 2011; Hayden et al. 2011; Hyman et al. 2011; Khamassi et al. 2013).

Content and Function of the Outcome Representation in ACC

We found that up to 30% of cells with error-locked firing modulation in one task epoch were encoding error outcomes in a second task epoch. This “joint error coding” comprised cells that detected errors across all task epoch combinations, suggesting that subpopulations of cells in the ACC integrate specific error outcomes into a more general representation of task outcomes (Hayden et al. 2011; Kennerley et al. 2011). Such an integrative function of classifying outcomes has been suggested recently in 4 different, but largely complementary, conceptual frameworks of ACC functions (Alexander and Brown 2011; Hyman et al. 2011; Khamassi et al. 2013; Shenhav et al. 2013). Most closely related to our findings are predictions from the Expected Value of Control Framework to understand ACC function. According to this EVC framework, the abstract detection of processing errors serves as the key signal to specify the required control intensity for future processing (Shenhav et al. 2013).

We speculate that erroneous outcomes in each of the processing stages of our experiment signal the need to strengthen inhibitory control, mediated by the ACC, of particular control processes as outlined in Figure 10 that are each realized in larger functional networks (Munakata et al. 2011; Hutchison et al. 2012; Medalla and Barbas 2012). In particular, we speculate that enhancing covert attentional focusing (in the Attention Epoch) involves, in particular, the strengthening of inhibitory control of the oculomotor system (e.g., encompassing FEF and superior colliculus) (Baluch and Itti 2011). Secondly, the strengthening of attentional filtering (and interference control in the Filter Epoch) may pose particular demands to inhibit salient bottom-up information arriving via sensory pathways (e.g., in humans via the temporo-parietal junction, see Corbetta and Shulman, 2011). Thirdly, erroneous stimulus-response mapping in the Choice Epoch is calling upon an enhancing the strength of task-rule representations and attentional set information in the lateral prefrontal cortex (Miller and Cohen 2001; Johnston et al. 2007; Tanji and Hoshi 2008; Womelsdorf et al. 2010).

In summary, such a network perspective on error processing in the ACC is consistent with its hypothesized roles of specifying the strength and identity of attentional control signals for efficient task processing (Phillips et al. 2013; Shenhav et al. 2013). Our findings suggest that such an overarching function can rely on error-detection signals that track specific attentionally demanding processing states in spiking activity of single cells in the ACC.

Supplementary Material

Supplementary material can be found at http://www.cercor.oxfordjournals.org/ online.

Funding

This research was supported by grants from the Canadian Institutes of Health Research (CIHR), the Natural Sciences and Engineering Research Council of Canada (NSERC), and the Ontario Ministry of Economic Development and Innovation (MEDI) (T.W.). S.W. was funded by the “Deutsche Akademie der Naturforscher Leopoldina” (LPDS 2012-08). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Notes

Conflict of Interest: None declared.

References

Alexander
WH
Brown
JW
.
2011
.
Medial prefrontal cortex as an action-outcome predictor
.
Nat Neurosci
 .
14
:
1338
1344
.
Amiez
C
Joseph
JP
Procyk
E
.
2005
.
Anterior cingulate error-related activity is modulated by predicted reward
.
Eur J Neurosci
 .
21
:
3447
3452
.
Asaad
WF
Eskandar
EN
.
2008a
.
Achieving behavioral control with millisecond resolution in a high-level programming environment
.
J Neurosci Methods
 .
173
:
235
240
.
Asaad
WF
Eskandar
EN
.
2008b
.
A flexible software tool for temporally-precise behavioral control in Matlab
.
J Neurosci Methods
 .
174
:
245
258
.
Baluch
F
Itti
L
.
2011
.
Mechanisms of top-down attention
.
Trends Neurosci
 .
34
:
210
224
.
Barbas
H
Zikopoulos
B
.
2007
.
The prefrontal cortex and flexible behavior
.
Neuroscientist
 .
13
:
532
545
.
Botvinick
MM
.
2007
.
Conflict monitoring and decision making: reconciling two perspectives on anterior cingulate function
.
Cogn Affect Behav Neurosci
 
7
:
356
366
.
Bryden
DW
Johnson
EE
Tobia
SC
Kashtelyan
V
Roesch
MR
.
2011
.
Attention for learning signals in anterior cingulate cortex
.
J Neurosci
 .
31
:
18266
18274
.
Cavanagh
JF
Grundler
TO
Frank
MJ
Allen
JJ
.
2010
.
Altered cingulate sub-region activation accounts for task-related dissociation in ERN amplitude as a function of obsessive-compulsive symptoms
.
Neuropsychologia
 .
48
:
2098
2109
.
Cheng
MY
Hall
P
.
1998
.
Calibrating the excess mass and dip tests of modality
.
J Roy Stat Soc Ser B Stat Methodol
 .
60
:
579
589
.
Corbetta
M
Shulman
GL
.
2011
.
Spatial neglect and attention networks
.
Ann Rev Neurosci
 .
34
:
569
599
.
Debener
S
Ullsperger
M
Siegel
M
Fiehler
K
von Cramon
DY
Engel
AK
.
2005
.
Trial-by-trial coupling of concurrent electroencephalogram and functional magnetic resonance imaging identifies the dynamics of performance monitoring
.
J Neurosci
 .
25
:
11730
11737
.
Dehaene
S
Kerszberg
M
Changeux
JP
.
1998
.
A neuronal model of a global workspace in effortful cognitive tasks
.
Proc Natl Acad Sci USA
 .
95
:
14529
14534
.
Euston
DR
Gruber
AJ
McNaughton
BL
.
2012
.
The role of medial prefrontal cortex in memory and decision making
.
Neuron
 .
76
:
1057
1070
.
Everling
S
Johnston
K
.
2013
.
Control of the superior colliculus by the lateral prefrontal cortex
.
Philos Trans Roy Soc Lond Ser Biol Sci
 .
368
:
20130068
.
Godlove
DC
Emeric
EE
Segovis
CM
Young
MS
Schall
JD
Woodman
GF
.
2011
.
Event-related potentials elicited by errors during the stop-signal task. I. Macaque monkeys
.
J Neurosci
 .
31
:
15640
15649
.
Grundler
TO
Cavanagh
JF
Figueroa
CM
Frank
MJ
Allen
JJ
.
2009
.
Task-related dissociation in ERN amplitude as a function of obsessive-compulsive symptoms
.
Neuropsychologia
 .
47
:
1978
1987
.
Hartigan
JA
Hartigan
PM
.
1985
.
The dip test of unimodality
.
Ann Stat
 .
13
:
70
84
.
Hayden
BY
Heilbronner
SR
Pearson
JM
Platt
ML
.
2011
.
Surprise signals in anterior cingulate cortex: neuronal encoding of unsigned reward prediction errors driving adjustment in behavior
.
J Neurosci
 .
31
:
4178
4187
.
Hayden
BY
Pearson
JM
Platt
ML
.
2009
.
Fictive reward signals in the anterior cingulate cortex
.
Science
 .
324
:
948
950
.
Henderson
DJ
Parmeter
CF
Russell
RR
.
2008
.
Modes, weighted modes, and calibrated modes: evidence of clustering using modality tests
.
J Appl Econ
 .
23
:
607
638
.
Holroyd
CB
Coles
MG
.
2008
.
Dorsal anterior cingulate cortex integrates reinforcement history to guide voluntary behavior
.
Cortex
 .
44
:
548
559
.
Holroyd
CB
Krigolson
OE
Baker
R
Lee
S
Gibson
J
.
2009
.
When is an error not a prediction error? An electrophysiological investigation
.
Cogn Affect Behav Neurosci
 .
9
:
59
70
.
Holroyd
CB
Yeung
N
.
2012
.
Motivation of extended behaviors by anterior cingulate cortex
.
Trends Cogn Sci
 .
16
:
122
128
.
Hutchison
RM
Womelsdorf
T
Gati
JS
Leung
LS
Menon
RS
Everling
S
.
2012
.
Resting-state connectivity identifies distinct functional networks in macaque cingulate cortex
.
Cereb Cortex
 .
22
:
1294
1308
.
Hyman
JM
Hasselmo
ME
Seamans
JK
.
2011
.
What is the functional relevance of prefrontal cortex entrainment to hippocampal theta rhythms?
Front Neurosci
 .
5
:
24
.
Hyman
JM
Whitman
J
Emberly
E
Woodward
TS
Seamans
JK
.
2013
.
Action and outcome activity state patterns in the anterior cingulate cortex
.
Cereb Cortex
 .
23
:
1257
1268
.
Ito
S
Stuphorn
V
Brown
JW
Schall
JD
.
2003
.
Performance monitoring by the anterior cingulate cortex during saccade countermanding
.
Science
 .
302
:
120
122
.
Johnston
K
Levin
HM
Koval
MJ
Everling
S
.
2007
.
Top-down control-signal dynamics in anterior cingulate and prefrontal cortex neurons following task switching
.
Neuron
 .
53
:
453
462
.
Kaping
D
Vinck
M
Hutchison
RM
Everling
S
Womelsdorf
T
.
2011
.
Specific contributions of ventromedial, anterior cingulate, and lateral prefrontal cortex for attentional selection and stimulus valuation
.
PLoS Biol
 .
9
:
e1001224
.
Kennerley
SW
Behrens
TE
Wallis
JD
.
2011
.
Double dissociation of value computations in orbitofrontal and anterior cingulate neurons
.
Nat Neurosci
 .
14
:
1581
1589
.
Kennerley
SW
Walton
ME
Behrens
TE
Buckley
MJ
Rushworth
MF
.
2006
.
Optimal decision making and the anterior cingulate cortex
.
Nat Neurosci
 .
9
:
940
947
.
Kerns
JG
Cohen
JD
MacDonald
AW
Cho
RY
Stenger
VA
Carter
CS
.
2004
.
Anterior cingulate conflict monitoring and adjustments in control
.
Science
 .
303
:
1023
1026
.
Khamassi
M
Enel
P
Dominey
PF
Procyk
E
.
2013
.
Medial prefrontal cortex and the adaptive regulation of reinforcement learning parameters
.
Prog Brain Res
 .
202
:
441
464
.
Lachenbruch
PA
.
2002
.
Analysis of data with excess zeros
.
Stat Methods Med Res
 .
11
:
297
302
.
Luk
CH
Wallis
JD
.
2013
.
Choice coding in frontal cortex during stimulus-guided or action-guided decision-making
.
J Neurosci
 .
33
:
1864
1871
.
Matsumoto
M
Matsumoto
K
Abe
H
Tanaka
K
.
2007
.
Medial prefrontal cell activity signaling prediction errors of action values
.
Nat Neurosci
 .
10
:
647
656
.
Medalla
M
Barbas
H
.
2012
.
The anterior cingulate cortex may enhance inhibition of lateral prefrontal cortex via m2 cholinergic receptors at dual synaptic sites
.
J Neurosci
 .
32
:
15611
15625
.
Medalla
M
Barbas
H
.
2010
.
Anterior cingulate synapses in prefrontal areas 10 and 46 suggest differential influence in cognitive control
.
J Neurosci
 .
30
:
16068
16081
.
Medalla
M
Barbas
H
.
2009
.
Synapses with inhibitory neurons differentiate anterior cingulate from dorsolateral prefrontal pathways associated with cognitive control
.
Neuron
 .
61
:
609
620
.
Miller
EK
Cohen
JD
.
2001
.
An integrative theory of prefrontal cortex function
.
Annu Rev Neurosci
 .
24
:
167
202
.
Mitchell
DJ
McNaughton
N
Flanagan
D
Kirk
IJ
.
2008
.
Frontal-midline theta from the perspective of hippocampal “theta”
.
Prog Neurobiol
 .
86
:
156
185
.
Morecraft
RJ
Stilwell-Morecraft
KS
Cipolloni
PB
Ge
J
McNeal
DW
Pandya
DN
.
2012
.
Cytoarchitecture and cortical connections of the anterior cingulate and adjacent somatomotor fields in the rhesus monkey
.
Brain Res Bull
 .
87
:
457
497
.
Munakata
Y
Herd
SA
Chatham
CH
Depue
BE
Banich
MT
O'Reilly
RC
.
2011
.
A unified framework for inhibitory control
.
Trends Cogn Sci
 .
15
:
453
459
.
Narayanan
NS
Cavanagh
JF
Frank
MJ
Laubach
M
.
2013
.
Common medial frontal mechanisms of adaptive control in humans and rodents
.
Nat Neurosci
 .
16
:
1888
1895
.
Narayanan
NS
Laubach
M
.
2008
.
Neuronal correlates of post-error slowing in the rat dorsomedial prefrontal cortex
.
J Neurophysiol
 .
100
:
520
525
.
Oostenveld
R
Fries
P
Maris
E
Schoffelen
JM
.
2011
.
FieldTrip: open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data
.
Comput Intell Neurosci
 .
2011
:
156869
.
Passingham
RE
Wise
SP
.
2012
.
The neurobiology of the prefrontal cortex: anatomy, evolution, and the origin of insight
 .
Oxford
:
Oxford University Press
.
Petrides
M
Pandya
DN
.
2007
.
Efferent association pathways from the rostral prefrontal cortex in the macaque monkey
.
J Neurosci
 .
27
:
11573
11586
.
Phillips
JM
Vinck
M
Everling
S
Womelsdorf
T
.
2014
.
A long-range fronto-parietal 5- to 10-Hz network predicts “top-down” controlled guidance in a task-switch paradigm
.
Cereb Cortex
 .
24
:
1996
2008
.
Procyk
E
Goldman-Rakic
PS
.
2006
.
Modulation of dorsolateral prefrontal delay activity during self-organized behavior
.
J Neurosci
 .
26
:
11313
11323
.
Quilodran
R
Rothe
M
Procyk
E
.
2008
.
Behavioral shifts and action valuation in the anterior cingulate cortex
.
Neuron
 .
57
:
314
325
.
Riesel
A
Weinberg
A
Endrass
T
Meyer
A
Hajcak
G
.
2013
.
The ERN is the ERN is the ERN? Convergent validity of error-related brain activity across different tasks
.
Biol Psychol
 .
93
:
377
385
.
Rothe
M
Quilodran
R
Sallet
J
Procyk
E
.
2011
.
Coordination of high gamma activity in anterior cingulate and lateral prefrontal cortical areas during adaptation
.
J Neurosci
 .
31
:
11110
11117
.
Saleem
KS
Miller
B
Price
JL
.
2013
.
Subdivisions and connectional networks of the lateral prefrontal cortex in the macaque monkey
.
J Comp Neurol
 . .
Sallet
J
Quilodran
R
Rothe
M
Vezoli
J
Joseph
JP
Procyk
E
.
2007
.
Expectations, gains, and losses in the anterior cingulate cortex
.
Cognitive, affective & behavioral neuroscience
 .
7
:
327
336
.
Seo
H
Lee
D
.
2007
.
Temporal filtering of reward signals in the dorsal anterior cingulate cortex during a mixed-strategy game
.
J Neurosci
 .
27
:
8366
8377
.
Shenhav
A
Botvinick
MM
Cohen
JD
.
2013
.
The expected value of control: an integrative theory of anterior cingulate cortex function
.
Neuron
 .
79
:
217
240
.
Tanji
J
Hoshi
E
.
2008
.
Role of the lateral prefrontal cortex in executive behavioral control
.
Physiol Rev
 .
88
:
37
57
.
Totah
NK
Kim
YB
Homayoun
H
Moghaddam
B
.
2009
.
Anterior cingulate neurons represent errors and preparatory attention within the same behavioral sequence
.
J Neurosci
 .
29
:
6418
6426
.
Van Essen
DC
Drury
HA
Dickson
J
Harwell
J
Hanlon
D
Anderson
CH
.
2001
.
An integrated software suite for surface-based analyses of cerebral cortex
.
J Am Med Inform Assoc
 .
8
:
443
459
.
Womelsdorf
T
Johnston
K
Vinck
M
Everling
S
.
2010
.
Theta-activity in anterior cingulate cortex predicts task rules and their adjustments following errors
.
Proc Natl Acad Sci USA
 .
107
:
5248
5253
.