When a subject faces conflicting situations, decision-making becomes uncertain. The human dorsal anterior cingulate cortex (dACC) has been repeatedly implicated in the monitoring of such situations, and its neural activity is thought to be involved in behavioral adjustment. However, this hypothesis is mainly based on neuroimaging results and is challenged by animal studies that failed to report any neuronal correlates of conflict monitoring. This discrepancy is thought be due either to methodological or more fundamental cross-species differences. In this study, we eliminated methodological biases and recorded single-neuron activity in monkeys performing a Stroop-like task. We found specific changes in dACC activity during incongruent trials but only in a small subpopulation of cells. Critically, these changes were not related to reaction time and were absent before any incorrect action was taken. A larger fraction of neurons exhibited sustained activity during the whole decision period, whereas another subpopulation of neurons was modulated by reaction time, with a gradual increase in their firing rate that peaked at movement onset. Most of the neurons found in these subpopulations exhibited activity after the delivery of an external negative feedback stimulus that indicated an error had been made. These findings, which are consistent with an executive control role, reconcile various theories of prefrontal cortex function and support the homology between human and monkey cognitive architectures.
When we make a decision, we sometimes face ambiguous, competing, or conflicting situations that render our choice uncertain. These situations are critical since we must adapt our actions very quickly in order to succeed in goal-directed behavior. Several hypotheses of how the brain regulates cognitive aspects of motor control propose that the dorsal anterior cingulate cortex (dACC) plays a prominent role. Among them, the “conflict-monitoring” theory postulates that the dACC detects the occurrence of conflict or interference between incompatible response tendencies, whereas other structures actually implement the adjustments to resolve conflict.
This theory, which is supported by event-related potentials (ERPs) and functional magnetic resonance imaging (Carter et al. 1998; Botvinick et al. 1999; Van Veen and Carter 2002), is challenged by the lack of direct evidence of conflict-monitoring correlates within the dACC in animal studies (Ito et al. 2003; Nakamura et al. 2005), as well as the weakness of the few single-unit recording studies in human (Davis et al. 2005; Sheth et al. 2012). Conflict monitoring is thought to involve neural processes that are likely to be detected by fMRI but not by single-unit recording (Cole et al. 2009). Another proposed explanation is based on the discrepancy between data in humans and monkeys (Mansouri et al. 2007, 2009; Cole et al. 2009) and posits that the previously accepted homology between structures in these species (Picard and Strick 1996, 2001) is less obvious than initially thought (Cole et al. 2009). However, an alternative explanation involves some critical methodological biases. First, since studies in monkeys do not use the exact analog of human tasks, it may be that the dACC is not engaged in a similar manner (see, e.g., Badgaiyan and Posner 1998). Second, the few studies that assessed the conflict-monitoring hypothesis at the neuronal level all employed oculomotor tasks (Ito et al. 2003; Nakamura et al. 2005), and there is little evidence to what degree these insights into circuits controlling eye movements can be generalized to motor control in other systems, such as the arm. To offset these methodological differences, we recorded single-neuron activity within the more rostral part of the cingulate motor area (CMAr), which is thought to be functionally homologous to the dACC human division involved in conflict monitoring (Picard and Strick 1996, 2001; Bush et al. 2000; Dum and Strick 2002; Rushworth et al. 2007), whereas 2 monkeys performed an arm-reaching task analogous to the classical Stroop task (CST) (Stroop 1935; MacLeod 1991; Michelet et al. 2007, 2009).
The Stroop task has the advantage of assessing interference both at the target-detection and response level (Kornblum et al. 1990; van Veen et al. 2001). In this task, subjects have to name the color in which a word representing a color is printed. In the case of incongruence between the color and the word (e.g., the word “GREEN” printed in red, see Fig. 1A,B), subjects face competition between reading the word and naming the color. In contrast, in the congruent condition, the color and the word are mapped onto the same response and lead to the so-called Stroop-facilitation effect. Our Stroop-like task (SLT) is based on learned associations (see Materials and Methods) between shapes and colors (banana and yellow, apple and red, pear and green). By permuting color and shape, we were able to present 3 cognitive conditions similar to the conditions of the CST (Fig. 1A). Animals were trained to choose the correct color target in response to a shape that was simultaneously provided on a video touch screen (Fig. 1B).
Materials and Methods
The animals (2 female monkeys (Macaca mulatta), weighing 5 and 6 kg, respectively) were housed in individual primate cages, and their care was supervised by veterinarians skilled in the health care and maintenance of nonhuman primates, in strict accordance with the European Community Council Directive for experimental procedures in animals.
We designed an experimental paradigm based on the original Stroop test, which demands resolution of an interference between 2 competing tendencies. This SLT is an associative task in which a subject learns to associate the shape of a fruit (banana, apple, or pear) with a corresponding color (yellow, red, and green, respectively; size 5 × 9 cm). When the 3 colors were presented on the lower part of a video touch screen, monkeys had to touch the one that corresponded to the fruit shape simultaneously presented above (Fig. 1B). Using different combinations of color and shape, we were able to present 3 cognitive situations, corresponding to control, congruent, and incongruent conditions, comparable with the CST conditions (Fig. 1A).
During experimental recording sessions, monkeys were seated in a primate chair, their heads immobilized and their left hands restrained, facing a video touch screen. Animals were trained to keep their right hands on the hand rest, which was equipped with a position sensor, until the shape and 3 colors appeared on the screen. Each trial comprised the same succession of events (Fig. 1C): A 3-s rest period during which the screen remained white and the monkey was instructed to remain still; 2) a 500-ms warning period (presentation of a warning stimulus in the form of a black circle, diameter 5 cm); 3) after 500 to 1000 ms, presentation of the task itself (SLT), that is, the simultaneous presentation of 1 fruit shape and the 3 different colors; 4) the response period during which the monkey took its hand off the hand rest, thus releasing the position sensor and was supposed to touch the correct color on the screen; 5) the evaluation period which began either with delivery of the reward (a drop of orange juice) in the case of success or with the appearance of a negative visual feedback in the form of a black screen after an incorrect response. During each session, we used a video device to monitor the trajectory of a monkey′s hand from the position sensor (hand at rest) to the choice target (touch screen) in order to detect error self-correction. A careful frame-by-frame examination was conducted on each trial and when an actual change in movement direction was observed (vacillation), we classified the trial as “self-corrected” (but see Supplementary Video for examples).
The prior training of each monkey took approximately 6 months. This began with the learning of a simple motor task, which was subsequently used during recording sessions as a motor control test, in which monkeys had to touch a black target, displayed randomly on the left, in the middle or on the right of a video touch screen. Once they had mastered this stage, we confronted them with the basic control situation, in which the 3 colors were presented with a colorless fruit shape. When a monkey′s overall performance reached >50% correct responses (chance = 33%), we presented the congruent situation, in which the color of the fruit presented was its normal color. Once this stage was mastered according to the same criteria, we moved on to the incongruent condition.
Because the strength of conflict behavioral effect is thought to depend on the relative proportion of congruent and incongruent trials (Botvinick et al. 2001), we used 30% of incongruent trials in order to maintain a high level of interference and 35% of control and congruent trials.
The different possible trials were presented randomly within each recording session with the color order of targets (6 possibilities, e.g., from left to right: red, green, yellow; see Fig. 1B, lower panel) being changed for each recording session.
Human Validation of Stroop-Like Task
In the CST, subjects had to name the color of the ink in which a word representing a color was printed without reading the word. In the case of incongruence between the ink color and the word (e.g., the word “BLUE” is printed in red), subjects were slower to respond, whereas congruency (e.g., the word “BLUE” written in blue) induced faster reaction time (RTs). The human experiments were run on a PC computer (CST) and a Macintosh (SLT). The tasks described later were administered in a random order: CST first and then SLT or SLT first and then CST. We used 30% of incongruent trials and 35% of control and congruent trials for both tasks in order to have exactly the same proportion of each condition in human and monkey experiments. Ten healthy right-handed subjects (5 female; age 23 to 38, mean = 27.2) performed the 2 tasks. During the classical Stroop Test, the subject was presented with color names (RED, YELLOW and GREEN) printed in a contrasting ink (e.g., YELLOW printed in red). The instruction was to name the ink color and to ignore the meaning of the word. The task comprised 98 trials. After each response given by the subject, the experimenter pressed 1 of 3 keys to code the response as correct, incorrect, or invalid. The code response given by the experimenter initiated the next trial. Each participant began with a practice block of 12 trials. A control task was also administered consisting of 24 trials in which a colored “XXXX” image, red, yellow, or green, appeared in the center of the screen. The subject was told to name the color of the “XXXX” presented. Naming (vocal response) latencies on the CTS were collected with a microphone that was connected to the computer. Each participant was tested individually. The subject was seated in front of the computer, and the experimenter was seated beside him. The experimenter allowed pauses between the tests. Before each task, the instructions were given to the subject, and a pretest was performed to ensure that the instructions were understood.
The SLT experiment was run on a Macintosh computer equipped with a touch screen (but also see Behavioral procedures in Materials and Methods). A position sensor was connected to the computer and placed in front of it. During the test session, the subject was asked to always keep his right hand on the sensor before giving the motor response and to replace it immediately after. This equipment made it possible to collect the subject′s RTs (latency to leave his hand from the start position after stimuli onset) and the time spent in reaching the target on the screen (Movement time). Human participants learned the modified task thanks to verbal instructions given at the very beginning of the experimental session where they were told to associate each shape with a single particular color (banana, apple, or pear associated with yellow, red, and green color, respectively), whatever the color in which the shape could eventually be filled in.
Surgery and Electrophysiological Recordings in Monkeys
A stainless steel recording chamber (diameter 19 mm) was implanted in the skull of each monkey under general anesthesia. The central axis of the cylinder was stereotaxically positioned at A26 and L0 in both monkeys (Fig. 9A,B). A head holder was embedded with dental cement around the chamber for immobilization during neuronal recording. Antibiotic (ampicillin, 100 mg/kg) and analgesic (paracetamol, 30 mg/kg) treatments were given for 1 week after surgery. Extracellular single-unit activity was recorded using tungsten microelectrodes insulated with epoxy (impedance, 0.5–1.0 MΩ at 1 kHz). Neuronal activity was amplified (10–20 K), filtered (300–3 KHz), and displayed on an oscilloscope. A window discriminator was used to select spikes from background activity. These were then processed through an analog–digital converter interface and stored on-line in a microcomputer. Most of the neuronal activity was also stored using the chart system software (Chart 5.0, ADI instruments). We then systematically compared isolation from our window discriminator (which allowed us to monitor online the response of the recorded neurons to the different events of the SLT) and the offline spike discrimination method based on shape recognition.
Intracortical microstimulation (ICMS) was performed after almost each neuronal recording. Characterization of the motor response induced by ICMS in the dACC allowed us to localize the arm region in the CMAr. A train of cathodal pulses (width, 0.2 ms; train duration, 50–150 ms at 300 Hz) was applied through a constant-current stimulator, and the same electrode was used for extracellular recording. Movements were recorded only if they were evoked repeatedly and were clearly identified by 2 investigators.
Peri-event time histograms, raster displays, and all the analyses were produced on a MATLAB platform (Mathworks, Inc.). Analyses of neuronal activity were performed on successive 20-ms data bins. To exclude the possibility that the neuronal response to the negative feedback signal could be a simple visual response, we randomly interspersed other trials in which we presented a colored screen (pink, blue, or brown) immediately after the initial rest period. We also checked that the neuronal activity related to obtainment of the reward was not attributable to simple appetitive mechanisms by occasionally giving drops of orange juice immediately after the rest period. Finally, because each experimental session was video recorded, the monkey′s orofacial movements could be carefully inspected during the SLT. Inspection revealed no association between the activity of these neurons and movements of the mouth.
We also charted the activity of the biceps brachii and triceps brachii muscles and used a custom-made video apparatus to record oculomotor activity. Because conflict and error monitoring could theoretically occur at different levels of the decision-making period (van Veen et al. 2001), we defined 9 epochs (see Fig. 1C) related either to the target presentation, movement initiation, target selection (i.e., selecting a target on the touch screen), and performance feedback stimuli (i.e., juice reinforcement as a positive feedback or “black screen” negative feedback). Relative to target presentation, firing rate was measured during 3 epochs: E1 correspond to the RT period (from stimuli onset to movement initiation, and therefore correspond exactly to the RT), E2 (0–200 ms after target presentation), and E3 (200–400 ms after target presentation). Relative to movement initiation, firing rate was measured during 4 epochs: E4 (−400 to −200 ms before movement initiation), E5 (−200–0 ms before movement initiation), E6 corresponding to the whole movement period (from movement initiation to target selection), and E7 (0–200 ms after movement initiation). Relative to target selection, firing rate was measured for one epoch, E8 (−200–0 ms before target selection). Relative to feedback stimuli, firing rate was measured for one epoch, E9 (0–500 ms after feedback) encompassing the feedback error-related negativity epoch (200–300 ms) found in human ERP experiments. Event-related modifications of activity were detected by comparing (paired t test) neuronal activity related to a specific event with activity during the initial 3-s rest period. A change in neuronal discharge frequency was considered significant if P-value was <0.01. Neurons exhibiting significant changes in neuronal activity for at least one condition in at least one epoch were considered as task related and further analyzed.
Statistical analyses were based on ANOVAs performed for each of the 9 epochs with mean firing rate of each trial as dependent variable. For the incongruent-related analysis, 3 factors were considered: task condition (control, congruent, or incongruent), behavioral performance (correct or error), and spatial location of target (left, middle, or right). For the error-related analysis, 1 or 2 factors were considered: behavioral performance (correct or error) and, when a sufficient number of correct and error trials allowed statistical comparison for the separate conditions (Kass et al. 2005), the trial task condition factor (control, congruent, or incongruent).
ANOVAs and t-test with Bonferroni–Dunn correction for post hoc analysis were used. We set the significance levels for the ANOVAs to P < 0.01 and for the post hoc t-test to P < 0.05. All data are given as means ± SEM.
To test whether a potential enhanced neuronal activity was caused by interference per se or was a consequence of more common long RTs in incongruent trials, we compared neuronal response magnitude in the 3 conditions for trials with similar RT distributions. Four criteria were necessary for a neuron to be considered as incongruent related: In order to analyze the timing at which an incongruent-related cell signaled interference, neuronal modulation trial-by-trial was determined by using 2 different methods. The first analysis determined the onset of the activation of the neuronal response (Fig. 4A). These times were determined by using Poisson spike train analysis, as described in Hanes et al. (1995). It has been shown that a distribution of interspike intervals approximates a Poisson distribution and thus provides a reliable null hypothesis to detect changes in neuronal modulation (Legéndy and Salcman 1985; Hanes et al. 1995). Poisson spike train analysis determines how improbable it is that the number of action potentials within a specific time interval is a chance occurrence. This is achieved by comparing the actual number of spikes within a studied epoch with the number of spikes predicted by the Poisson distribution derived from the mean discharge rate during the entire time period in which deviations from randomness are sought. The probability P that a specific elevation in activity is a chance occurrence is determined by Poisson′s formula:
Significant difference in firing rate between incongruent and control trials.
Significant difference in firing rate between incongruent and congruent trials.
Exhibited no difference in firing rate between congruent and control trials.
The above-mentioned conditions also had to be met when matched RT trials were used (see Supplementary Fig. 1). In practice, we first extracted all correct incongruent trials. For each of these trials, we selected the trial exhibiting the minimum difference in the duration of RTs from the congruent trials group. The same procedure was performed for control trials. A trial could not be selected twice. Hence, we obtained a reduced set of trials, with the same number of incongruent, control, and congruent trials with close RT distributions. Moreover, because correct incongruent trials are always less frequent than correct congruent and control ones, this can cause a bias in firing rate frequency analysis. The comparison of the same number of trials for the 3 conditions allowed us to check whether the different number of trials across conditions did not affect our results.
One cell was removed from this analysis because the onset of activation was not possible to determine for most of its trials, so it was impossible to calculate a corresponding P-value. The second analysis assessed the maximum activation (maximum spike rate) reached during a given epoch. The relationships between the RTs of incongruent trials and both measures of neuronal modulation for the corresponding trials could then be studied.
We also used another method (DiCarlo and Maunsell 2005) in order to determine the onset of activation of the neuronal response (also called here neuronal latency or NL) and consequently at which level of processing these neurons operate (i.e., where in the neuronal “processing chain,” the neurons responsible for the sensory-motor transformation from stimulus to behavioral response operate).
Two different values (λ and β) obtained from a single neuron indicate whether this neuron exhibited “sensory,” “sensorimotor,” or “motor” responses. λ is the neuron′s normalized mean NL and is calculated by dividing its mean NL by the mean RT. It has been used in many studies to determine the position of a neuron along the sensory-motor pathway. However, such a measure cannot distinguish neurons that are “on” or “off” the processing chain so additional information is provided by a statistical measure of the trial-by-trial association of RT and NL. Since the interesting information is not contained within the absolute covariance per se but rather within the fraction of the RT variance that is associated with the NL, DiCarlo et al. defined a normalized measure of association β: the covariance of NL and RT divided by the RT variance: β is a unit-less value that progresses from near zero for neurons that have little correlation with RT (neurons involved in the early stage of information processing) to near one for neurons that have activity closely correlated with the timing of the behavioral response. This analysis is performed for all the trials of a recording session and not for a reduced set of trials such as incongruent-related trials, since it must take into account the variance of the RT of the complete set of trials.
Finally, to provide a more detailed quantitative analysis of the temporal evolution in the firing rate of these neurons, we studied the trial-by-trial relationship between changes in the slope of the firing-rate-versus-time and RT. Hence, like Roitman and Shadlen (2002), we studied the slope of the straight line (ramp) that best explained the average firing rate (calculated during the epoch E1; see above). This ramp indicates the best fit (obtained under maximum likelihood estimation) of a linear rate function to the spike train and provides an estimate of the slope of the “firing-rate-versus-time” function. For each given (correct) trial, the slope of the “firing-rate-versus-time” function is calculated and plotted against the corresponding RT. This makes it possible to study the relationship (coefficient of correlation) between RT and the slope of the “firing-rate-versus-time” function of a given neuron (Fig. 6A–C, right part). This analysis is repeated for each neuron and then the distribution of correlation coefficients obtained from all the neurons of the studied neuronal population is plotted (Fig. 7A).
In order to validate our analog of the Stroop task, we first performed a validation study in humans by comparing performances in both the classical and analog Stroop task.
Human results are shown in Figure 2A,B. Because of the extremely low error rate, only reaction times (RTs) for correct trials were considered in the analyses. The RTs were subjected to an analysis of variance, with condition (control, congruent, and incongruent) as independent variable. For the naming task (Fig. 2B), this showed a clear influence of task condition on RT in all subjects (F2,875 = 42; P < 0.0001). This was confirmed by post hoc comparison, which showed that RTs were significantly longer in the incongruent condition (811.3 ± 10.4 ms) than in the congruent (674.2 ± 13.9 ms) and control conditions (683.4 ± 14.6 ms; P < 0.0001). The slight difference between control and congruent conditions was not significant. The same analysis performed for the SLT (Fig. 2A) also showed a clear influence of task condition on RT for all subjects (F2,962 = 16.1; P < 0.0001). This was confirmed by post hoc comparison, which showed that RTs were significantly longer in the incongruent condition (573 ± 6.3 ms) than in the congruent (514.2 ± 8.8 ms) and control conditions (535.7 ± 8.9 ms; P < 0.0001). The slight difference between control and congruent conditions was still not significant.
Analyses of monkeys′ behavior (Fig. 2C,D and Supplementary Table 1) during this task showed a clear influence of conditions on RT (ANOVA, main effect of condition: F2,306 = 67.7, P < 0.00001) as confirmed by post hoc comparison, which showed that the mean RT was significantly increased in incongruent trials (1061.8 ± 29.2 ms) in comparison with congruent (691.1 ± 16 ms) and control (849.5 ± 20.5 ms) ones (Fig. 2C; t test, P < 0.00001). This behavioral “Stroop effect” also present in humans performing the same SLT [see Fig. 2A, but see also Stroop (1935); MacLeod (1991)] is thought to reflect the competition between the 2 incompatible responses induced either by color or shape of the cue. This hypothesis was further confirmed by the higher error rate in the incongruent condition (40.8%, Fig. 2D) and by the fact that errors were mainly directed toward targets colored as the incongruent cue (63%, χ2 = 1252, df = 1, P < 0.0001). Furthermore, acts of error self-correction distinguished incongruent from congruent trials (see Supplementary Fig. 2 and Supplementary Video clips for details). Indeed, the percentage of self-corrected trials was significantly greater in the incongruent condition (6.3%, χ2 = 57.86, df = 2, P < 0.0001) and lesser in the congruent condition (2.9%) than in the control one (4.4%). Moreover, we also found significant shortening of RTs (t test, P < 0.00001) and improved performance (χ2 = 1026.3, df = 2, P < 0.0001) in the congruent condition compared with the control condition (error rate: 13.5% vs. 28.6%; Fig. 2D), as expected in a valid analog of the Stroop task. Taken together, these results indicate that our SLT is well designed to study conflict processing.
Based on the conflict-monitoring hypothesis, Botvinick et al. (2001, 2004) tried to explain the Gratton effect, a behavioral effect showing that the influence of interference is not limited to the current trial but also affects performance in the subsequent trial (Gratton et al. 1992). To test this prediction, we compared RTs and error rate in compatible trials (i.e., congruent and control) and incongruent trials depending on the condition of the previous trials (Fig. 2E,F). Four categories were defined: cC: compatible trials preceded by a compatible trial, iC: compatible trials preceded by an incongruent trial, iI: incongruent trials preceded by an incongruent trial, cI: incongruent trials preceded by a compatible trial. A one-way ANOVA showed a clear influence of the preceding trial on the RT (F3,12524 = 246.23, P < 0.00001), confirmed by post hoc comparison, which showed that RTs (expressed in ms) for an incongruent trial preceded by another incongruent trial (iI) were shorter (988.2 ± 24.8) than when the incongruent trial was preceded by a compatible one (CI; 1072.8 ± 12) (Fig. 2E; t test, P < 0.00001). Similarly, error rates were lower in iI than in cI trials (31.9 vs. 42.8, χ2 = 563.5, df = 3, P < 0.0001) (see Fig. 2F).
To study neuronal responses to interference during the entire decision period, we investigated 3 different periods of the SLT: the RT period between onset of the task cues and movement initiation, the movement period defined as the time between movement initiation and target selection on the touch screen, and the evaluation period just following positive (i.e., juice delivery) or negative (i.e., “black screen”) external feedback. These periods were either considered as a whole or subdivided into smaller epochs (see Materials and Methods and Fig. 1C). We recorded 403 neurons in the left hemisphere whereas monkeys performed right arm-pointing movements. The analysis below includes the 103 task-related neurons (mostly recorded in the ventral bank of the cingulate sulcus) with a sufficient number of trials to allow statistical comparison in all conditions (i.e., cells that contained sufficient correct and error trials in control, congruent, and incongruent conditions). About nineteen percent (19/103) of neurons were incongruent related. They exhibited a significantly enhanced activity during incongruent trials in comparison with control and congruent ones (Fig. 3A–C). Forty-two percent of incongruent-related neurons (8/19) discharged after stimulus presentation (Fig. 3A) whereas 21% (4/19) exhibited enhanced incongruent activity just before (Fig. 3B) and 37% (7/19) immediately after (Fig. 3C) movement initiation.
These neurons were not influenced by the direction of movement and were thus not involved in a direction-selective conflict-monitoring process (Hoshi et al. 2005; Nakamura et al. 2005). This enhanced stimulus-detection or response-related activity in incongruent trials was not present during error trials in which monkeys chose incorrect targets (black dotted lines in Fig. 3A–C).
To test whether this enhanced neuronal activity was caused by interference per se or was a consequence of more common long RTs in incongruent trials, we compared neuronal response magnitude in the 3 conditions for trials with similar RT distributions (see Material and Methods). The response property of incongruent-related neurons was virtually identical in such analysis (see Supplementary Fig. 1), which ruled out an incidental “incongruent” effect arising from an RT bias and also confirmed that the different number of trials across conditions did not affect our results. We almost never found any trial-by-trial relationships between the onset of neuronal modulation and RT. This indicates that the timing of incongruent-related activity had no impact on the timing of behavioral adaptation (Fig. 4).
The activity of each of the incongruent-related neurons was compared for cC, iC, cI, and iI trials in their corresponding incongruent-related epochs. Individual and population neuronal analyses were performed. In the latter case, firing rates (F) were normalized (7) by using the range of responses for each cell (maximum rate minus minimum rate, Fmax − Fmin), according to the formula (Fi − Fmin)/(Fmax − Fmin). Mean firing rates across these 4 conditions were compared by one-way ANOVA and paired t-tests. However, there was no significant difference in neuronal activity for iI trials compared with cI trials (t test, P > 0.05; Fig. 5). Hence, behavioral performance but not neuronal firing rate was modulated by previous trial.
We then analyzed the activity of each neuron after separation of trials into fast, medium, and slow RTs. Indeed, an alternative hypothesis posits that the dACC is more active during difficult tasks. In accordance with Paus (Paus et al. 1998), we defined difficulty by using a simple behavioral index based on the idea that longer RTs mean more difficulty. Figure 6C shows a typical neuron exhibiting a ramp-like increase in activity to a threshold value. An earlier rise in spike rate is found for shorter RT trials. In contrast, other examples of cell activity during the decision-time epoch illustrate a more phasic (Fig. 6A) or tonic (Fig. 6B; sustain activity) pattern that cannot discriminate between the different trials. To provide a more detailed quantitative analysis of the temporal evolution in firing rate of these neurons, we studied the trial-by-trial relationship between changes in slope of the firing-rate-versus-time and RT (e.g., Fig. 6 right panels, but see Materials and Methods section for details). Figure 7A summarizes this relationship with negative values indicating that firing rate is inversely correlated with RT, so these neurons exhibit a gradual-ramping activity and the movement seems triggered at a constant level of neuronal activity. Taken together, these results indicate that a subpopulation of neurons is modulated by incongruent trials, another subpopulation is directly influenced by RT (5% of the neurons shared both properties), and a third subpopulation is characterized by sustained activity.
Because these neurons are modulated during the whole decision time period, we sought to determine at which level of processing these neurons operate. Statistical tests using neuronal latency (NL, i.e., the time for neuronal modulation in each trial) were performed to determine whether neurons were related to stimulus identification or to movement initiation. In accordance with the RT–NL method originally described by DiCarlo and Maunsell (2005), we postulate that neurons involved in the early stage of information processing respond with mean NL close to stimulus presentation (e.g., neurons in Fig. 6A and Supplementary Fig. 1A), whereas neurons late in processing become active shortly before movement initiation (e.g., neurons in Fig. 6C and Supplementary Fig. 1C). Figure 7B plots normalized mean NL (λ) versus normalized covariance of NL and RT (β) values, the distribution of which indicates whether neurons exhibited “sensory,” “sensory-motor,” or “motor” responses (DiCarlo and Maunsell 2005). These data show that incongruent-related and gradual-ramping neurons are dispersed and intermingled throughout the whole sensory-motor process involved in decision-making.
During the third (evaluation) period, 89% (92/103) of CMAr neurons exhibited significant activity in relation to performance feedback stimuli. The activity of 25% (23/92) of them was modified by positive but not by negative feedback. In contrast to the first group, 11% (10/92) of them increased their firing rate after negative but not after positive feedback. The third type was the most frequent (59/92 or 64%). Most of these neurons showed a sharp activation to negative feedback, as well as a smaller but significant increase in activity after positive feedback. Among the incongruent-related neurons, a majority was also feedback-related (73.6% n = 14, χ2 = 4.22, df = 1, P = 0.03; Fig. 8A,B but see Fig. 8C for a counterexample of this rule and Fig. 9C). These activities were not modulated in relation to the cognitive conditions of the task.
Several hypotheses have been proposed to account for the role of the dACC in cognitive control. Theoretically, the conflict hypothesis should account for variations in performances (RT and error rate) on line and on subsequent trials (Botvinick et al. 2001, 2004). Thus, RTs are expected to be longer for incongruent (i) trials than for congruent ones (c) because of competition at the stimulus detection or response level. Our finding that RTs during incongruent trials were significantly longer than during control or congruent ones indicates that incongruent trials clearly influenced the monkey′s online behavior (Fig. 2C). Similarly, iI trials (incongruent trials following incongruent trials) are thought to be faster than cI trials (incongruent trials following congruent trials) because the preceding incongruent trial results in greater cognitive control. In our task, we also found clear behavioral evidence of such an adaptation effect (see Fig. 2E). This is consistent with other experiments using incongruent stimuli in monkeys (Washburn 1994; Lauwereyns et al. 2000; Ito et al. 2003; Mansouri et al. 2007). However, until now, conflict-related neurons have been found in the supplementary eye field (SEF) during eye movement tasks but not in the anatomically adjacent ACC (Stuphorn et al. 2000; Ito et al. 2003). The presence of cells monitoring conflict during saccadic tasks within the SEF (Stuphorn et al. 2000), a brain region involved in oculomotor control, suggests a close link between the brain region involved in cognitive control and the motor system on which the behavioral response relies. The results of our arm-reaching task provide evidence that the nonhuman primate dACC exhibits neuronal activity in response to incongruent stimuli and could be involved in conflict processing. It seems clear that the involvement of ACC in cognitive control depends on the motor system involved in the response (Turken and Swick 1999) and that somatotopical organization of CMAr is dominated by, and helps to control, the arm region (Hoshi et al. 2005).
Interference arising from incongruent stimuli is believed to result from competition between mutually exclusive potential responses. During correct incongruent trials, the shape of the stimulus activated one (correct) learned response (e.g., selecting the yellow target on the touch screen if the cue is a banana), whereas the incongruent color simultaneously activated an incompatible (incorrect) response alternative (e.g., selecting the red target if the banana is colored in red), yielding incongruence between the 2 potential options.
According to the conflict-monitoring theory, error trials also exhibit a transient period during which both correct and incorrect responses are simultaneously activated and should consequently be associated with dACC conflict detection (Botvinick et al. 2004). Strictly speaking, these 2 criteria (error and conflict) should be fulfilled to validate the conflict-monitoring theory. In opposition to this hypothesis, we did not find any enhanced activity in erroneous incongruent trials, which exhibited similar activity to congruent and control trials. Furthermore, we surprisingly found incongruent-related responses at both the stimulus-identification and response levels. In all cases, we almost never found any trial-by-trial relationships either between the onset of neuronal activation and RT (Fig. 4A) or the maximum of activation and RT (Fig. 4B). This indicates that the timing of incongruent-related activity had no impact on the timing of behavioral adaptation. We also failed to reproduce an important feature of the simulation results (Yeung et al. 2004) showing that response conflict should be only slightly greater in incongruent trials than in congruent trials when the conditions are matched for RT (see Supplementary Fig. 1). Finally, we failed to find any significant differences either at individual cell or population level (Fig. 5) for neuronal responses on incongruent trials following congruent trials (cI) and on incongruent trials following incongruent trials (iI). Assuming that incongruent-related neurons directly monitor conflict, they should have exhibited a stronger activity during erroneous trials or cI trials (Botvinick et al. 1999). It is thus possible that these incongruent-related signals are related to an “attention for action” process linked to focal attention that is directed toward a specific feature of the external stimulus (here, the shape of the cue) for selecting an appropriate action (Isomura et al. 2003). This attention process, which is known to improve behavioral performance by increasing processing of pertinent information, could be implemented by an executive control network in which the dACC plays a central role (Petersen and Posner 2012). Moreover, because the percentage of self-corrected trials was significantly greater in the incongruent condition and lower in the congruent condition, one could hypothesize that this incongruent-related activity is directly related to an on-line error detection process serving as a triggering signal that indicate the need of an enhanced attentional control for an optimal task processing (Shen et al. 2015). This hypothesis is in line with conceptual frameworks of ACC function emphasizing the role of dACC in the tracking and identification of the processes leading to erroneous task outcomes (Shen et al. 2015) and could additionally explain why incongruent-related activity was not present during incorrect incongruent trials. In all cases, incongruent-related activities occurred before the movement was completed. It is therefore possible for the subject to influence the outcome of the current trial by providing top-down control until the very last moment of the decision period.
Another central function of executive control system is outcome error detection (Posner and Rothbart 1998; Petersen and Posner 2012). Error-related activities are now well characterized and have been studied and observed in different situations: following overt response errors in choice RT tasks (Gehring et al. 1993), following feedback about response accuracy (Miltner et al. 1997; Holroyd and Coles 2002), and following late responses in deadline RT tasks (Luu et al. 2000). Here, we found that “decision-related” neurons could also be activated by the negative external feedback stimulus (on error trials), so they could then participate in triggering adjustments for the next trial. In our SLT, feedback-related activities occur, by definition, after movements, when errors are recognized by their consequences (i.e., here, a negative feedback) (Ullsperger and von Cramon 2003). This may explain why we could not identify any of the error-related activity at the time of the motor response (that the conflict-monitoring hypothesis predicts). Such activity would be more likely in a speeded RT task when errors correspond to slips resulting from premature responses (Ullsperger and von Cramon 2003). Indeed, in our task, there was virtually no time constraint because the monkey had a maximum of 4000 ms in which to touch the correct target. Erroneous anticipated responses are therefore unlikely to be seen during this task, in contrast to speeded RT tasks with a drastic time constraint for responses in which participants are instructed to respond as fast (before a deadline) and accurately as possible. This result consequently complements other studies addressing the error-monitoring role of dACC neurons in monkeys and humans in both arm and eye movement tasks (Shima and Tanji 1998; Ito et al. 2003; Ridderinkhof et al. 2004; Emeric et al. 2008; Michelet et al. 2009).
Another functional hypothesis regarding the involvement of ACC in cognitive processes posits that dACC activation reflects task difficulty. This hypothesis was preliminary based on a meta-analysis showing that difficulty (defined by the presence of longer RTs in some experimental situations relative to others) plays a major role in modulating blood-flow response in the ACC (Paus et al. 1998). This finding is even more firmly supported by a recent study showing a strong correlation of ACC fMRI response and RTs (Thielscher and Pessoa 2007). The gradual-ramping neurons we found could account for this result. Indeed, because they reflect the temporal accumulation of information relative to the task up to a fixed threshold, the longer the trial, the longer these neurons exhibited an increase in firing rate significantly different from its “rest” activity. This was not the case of the other type of decision-period-related neurons exhibiting a more phasic pattern. The question of the relationship between single neuronal activity and blood-flow response is out of the scope of our paper. However, given the logical inference that increasing difficulty caused an increase in RT, which in turn caused longer neuronal activation, we suggest that the ramp-like and sustained patterns of response we found could be related to the trial-to-trial correlation of ACC blood-flow and RTs found in studies in which the longer the RT, the stronger the ACC blood-flow response (Thielscher and Pessoa 2007). This sustained activity could reflect the maintenance of task parameters across the trials and thus be directly related to an executive control function (Dosenbach et al. 2006). Because a large percentage of the corticospinal neurons in the arm representation of the CMAr project to lower cervical segments (Dum and Strick 1991; He et al. 1995), it is also conceivable that such activities participate in the priming of lower motor structures. Taken as a whole, our results cannot definitely establish whether activity in the dACC is associated with detecting or resolving the conflict. They indicate that the CMAr neurons possess all the characteristics required by a supervisory attention or executive control system: They contribute to monitoring the consequences of (arm) movements and are also more activated when situations require greater (focal) attentional control. These gradual-ramping-, Incongruent-, and feedback-related activities are often implemented by the same individual neurons and occupy largely overlapping loci in the monkey dACC. This finding can be directly related both to studies emphasizing the global performance-monitoring function of the human dACC, in which largely overlapping conflict and negative feedback clusters of activation are found (Ridderinkhof et al. 2004) and to several other studies in monkeys focusing on the adaptation of behavior under the guidance of feedback (Shima and Tanji 1998; Procyk et al. 2000; Ito et al. 2003; Amiez et al. 2005, 2006; Matsumoto et al. 2007; Michelet et al. 2007). This consequently minimizes the hypothesis of fundamental differences between humans and monkeys in ACC functions (Cole et al. 2009, 2010; Schall and Emeric 2010) and may have important implications for present and future neurophysiological research, since the pertinence of primate studies is strongly based on the supposed homology between human and monkey neuronal structures.
We thank P. Drapeau for his help in initial matlab code development and James J. DiCarlo for sharing the matlab code for the RT–NL analysis; R. Cooke for improving the English; P. Cisek, J. Schall, P. Middlebrooks, D. Godlove, M. Bosc, and M. Guthrie for their thoughtful comments on a previous version of the manuscript, and M. Bosc for her help with the anatomical reconstruction. Conflict of Interest: None declared.