Abstract

Although there are numerous theories regarding anterior cingulate cortex (ACC) function, most suggest that it is involved in some form of action or outcome processing. The present study characterized the dominant patterns of ACC activity on a task in which actions and outcomes could vary independently. Patterns of activity were detected using a modified form of principal component analysis (PCA), termed constrained PCA in which a regression procedure was applied prior to PCA to eliminate the contribution of nontask-related activity. When trials were grouped according to outcome, a PC was found in all subjects and sessions that had large fluctuations during actions but only differentiated correct versus error trials prior to the end of the delay and again at time of the outcome. Another PC was always present that separated right from left lever presses, but only around the time of the actual lever press. Individual neurons exhibited significant selectivities for trials involving different actions and/or outcomes. Of the ACC neurons that exhibited significant outcome selectivity, the majority fired more on error trials. The present study revealed separate as well as integrated action and outcome monitoring in the ACC, especially, although not exclusively, under conditions when an error is likely.

Introduction

Numerous theories have attempted to explain the function of the medial prefrontal cortex (mPFC) and, in particular, the anterior cingulate cortex (ACC). Historically, the ACC was thought to be primarily involved in some aspect of movement planning or initiation. ACC neurons project directly to motor areas such as the striatum and spinal cord (Dum and Strick 1991; Devinsky et al. 1995), and the most obvious deficits following ACC damage in humans are akinetic mutism, motor neglect, and impaired movement initiation (Mesulam 1981; Devinsky et al. 1995; Cohen et al. 1999). Accordingly, single-unit activity in the primate and rat mPFC is selective for preparation and execution of movements, especially self-paced movements or in the requisite attentional functions that are necessary to select the correct actions or sequences of actions (Picard and Strick 1996; Shima and Tanji 1998; Pratt and Mizumori 2001; Matsumoto et al. 2003; Isomura and Takada 2004). ACC neurons also respond to visual cues that indicate different future actions (Isomura et al. 2003) or to cues that signal particular action–reward combinations (Matsumoto et al. 2003).

More contemporary theories hold that the ACC, especially the rostral midcingulate portion, is involved in some sort of outcome or error evaluation. The error signal produced by the ACC may be the result of a mismatch between intended and actual behaviors (Bernstein et al. 1995) or when the appropriate response becomes active just as the inappropriate response is executed (Carter et al. 1999; Botvinick et al. 2001). The neural basis for these functions may be mediated in part by the ACC neurons that fire in response to errors (Niki and Watanabe 1976; Ito et al. 2003; Hyman et al. 2011). Holroyd and Coles (2002) expanded on this basic idea by suggesting that dopamine inputs modulate the error signal in ACC according to the valence of the error prediction signal produced when outcomes do not match expectations. A related idea is that ACC neurons generate signals reflecting a learned prediction of the probability of errors (Brown and Braver 2005) or, more generally, the probability of different outcomes (positive or negative) that will result from an action (Alexander and Brown 2011). These types of theories are consistent with recording data, showing that ACC neurons fire when a reward is delivered but not expected or when a reward is not delivered but expected (i.e. prediction errors) (Amiez et al. 2006; Matsumoto et al. 2007). Thus, while numerous theories emphasize that the ACC is involved in some form of mapping between actions and expected or intended outcomes, it is presently unclear how actions and outcomes are represented by ACC networks and how integrated or distinct such representations are at a neural level.

The present study sought to examine the dominant neuronal firing rate patterns in the ACC as rats performed an operant delayed alternation task that is sensitive to mPFC lesions (Izaki et al. 2001). This task has the property that the identical action yields different outcomes on alternative trials, and therefore it is possible to dissociate outcomes from specific actions. Furthermore, the task is sufficiently challenging that errors occur even in well-trained rats, thereby allowing for a comparison of neural activity on correct and error trials. In order to examine the dominant patterns of activity within ACC networks, we used a method well suited for this purpose, principal component analysis (PCA) (Chapin and Nicolelis 1999; Laubach et al. 1999; Paz et al. 2005; Narayanan and Laubach 2009; Machens 2010). To reduce any influence of “off-task” activity, a modified version of PCA was used, termed constrained PCA (CPCA), in which a regression procedure is performed prior to PCA. CPCA also allows differential grouping of trials based on different actions with similar outcomes (all correct trials with right vs. left lever presses) or for different outcomes independently of action (correct vs. errors following both right and left lever presses). Finally, CPCA allowed us to examine exactly when during the time course of a trial differences in activity appeared. In addition to ensemble analysis using CPCA, the selectivity of individual neurons for actions and/or outcomes was investigated. Evidence for both action and outcome encoding was found at the network and single neuron level.

Materials and Methods

Surgery and Electrophysiology

All experimental procedures were approved by the University of British Columbia Institutional Animal Care and Use Committee. A total of 5 separately housed adult male Long-Evans rats participated in this study. For implant surgery, animals were deeply anesthetized under isoflurane gas and implanted with 16 moveable tetrodes in a hyperdrive. Tetrodes were made of 15 μm coated tungsten wire (California Fine Wire, Grover Beach, continuous alternation) and were initially lowered approximately 680 μm into cortex during surgery (AP: +2.7 mm from bregma and ML: ±0.5 mm from the midline). The hyperdrives were mounted to skull screws using dental acrylic (2 grounding screws were placed posterior to lambda, above the cerebellum) and shielded in plastic sheeting. After 1–2 weeks of postsurgical recovery, animals were attached to the tether, and tetrodes were slowly advanced ventrally into the mPFC over a 5–10-day period. Once all tetrodes were placed into ACC (∼2.5 mm from cortical surface) according to lowering records and atlas coordinates, small adjustments were made to maximize the number of neurons recorded.

Electrodes were connected to an EIB-36TT (Neuralynx Inc., Bozeman, MT, USA), which was plugged into HS-36 headstages and tether cables (Neuralynx Inc.). Signals were sent to a Digital Lynx 64 channel system (Neuralynx Inc.), converted from analog to digital, and sent to a PC workstation. Electrophysiological, positional, and behavioral data were read into Cheetah 5.0 software (Neuralynx Inc.) and recorded for later analysis. Files were then read into Offline Sorter (Plexon Inc., Dallas, TX, USA) for spike sorting. Only spikes with amplitudes over 1.5 times the background noise were used for spike sorting. Spikes were initially sorted using visually dissociable clusters in 3D projections along multiple different axes for each individual electrode of a tetrode (peak amplitude, valley amplitude, peak-to-valley ratio, PCs, and area). Sorting was confirmed by examining autocorrelations, cross-correlations, and analysis of variance (ANOVA) conducted from 2D and 3D projections of the different axes used for visual cluster sorting in Offline Sorter (Plexon Inc.). Spike timestamps were then read into MATLAB (Mathworks Inc., Natick, MA, USA) for all further analyses. For all analyses, fast spiking units with mean firing rates >10 Hz (putative interneurons) were removed to ensure that no biases due to cell type skewed the results (in total 19 of 404 recorded units met these criteria).

Histology

After each animal had completed the experiment, they were deeply anesthetized under halothane gas, and electrolytic lesions were created at each electrode location. Animals were then perfused with a solution containing 25 parts 10% buffered formalin, 1 part glacial acetic acid, and 10 g of potassium ferrocyanide. This solution causes a Prussian blue reaction, which marks with blue the location of iron particles deposited by the electrode lesion. The brains were then removed and stored in a 10% buffered formalin solution for about 1–2 weeks. After this time, the brains were then sliced, mounted, and stained with neutral red to determine precise electrode locations. Since multiple sessions were recorded from individual animals, the precise recording locations could not be derived from electrode lesions merely that all electrode tracks were in the ACC and sometimes prelimbic cortex and did not extend into the more ventral infralimbic area.

Behavior

Alternation tasks were performed in a high-walled square enclosure with 2 panels containing the manipulanda. One panel had 2 retractable levers (6′′ apart) with a reinforcement spout in between and a second panel, on the opposite side of the environment, featuring a nose-poke port and a stimulus light 2′ above the port, and the 2 panels were 10′′ apart (Fig. 1A). For the first trial of each session, both levers were extended, and depressing either lever resulted in reinforcement delivery (Fig. 1B). Following the lever press, both levers retracted, and for continuous alternation sessions, the stimulus light above the nose-poke port was illuminated, indicating that the nose-poke port was active. For delayed alternation sessions, a 10 s delay period was inserted between lever presses and the nose-poke stimulus light turning on, and nose pokes made before the delay had no effect. After the delay, animals needed to nose poke (at the panel opposite to the lever panel), which resulted in the stimulus light turning off and both levers extending. The animals then needed to depress the opposite lever from the one depressed on the first trial to receive a reinforcement pellet (if the right lever was depressed on trial 1, then the left lever was the correct lever for trial 2). Pressing either lever caused both levers to retract and the stimulus light over the nose-poke port to illuminate during continuous alternation or the 10 s delay to commence for delayed alternation task. For a given trial, the correct lever was always the opposite one that was pressed on the preceding trial, even if an error was made on the preceding trial. Thus if the error lever was depressed (e.g. the right lever on trial 2), the opposite lever from this error response was the correct lever on the following trial (e.g. the left lever on trial 3). Trials continued in this alternating fashion until the animal completed at least 30 correct trials.

Figure 1.

Alternation task apparatus and structure and behavioral results. (A) Illustration of the alternation task chamber and manipulanda. (B) Flow chart of the delayed alternation trial structure. From left to right, after a lever press is made on the right lever, both levers retract and a 10 s delay interval begins. After the delay, a light turns on over the nose-poke port on the opposite end of the chamber, indicating that the port is active. Following a nose poke, both levers extend and a response on the left lever leads to the delivery of a reinforcement pellet. (C) Cumulative performance during delayed alternation sessions. The number of correct trials is on the y-axis and the number of total trials is on the x-axis. Different sessions are indicated by line color. (D) Mean intertrial interval for correct and error trials. Average time between lever presses in seconds on y-axis (error bars = SEM). (E) Movement paths for a series of correct trials during delayed alternation. The light-blue-shaded box on the blue line in plot C shows the trials plotted. The plots show the paths taken after a correct response (red dot) and during the subsequent delay period (red lines), and the paths taken on the ensuing trial between the nose-poke (green dots) and lever press panels (response period; blue lines). Shown in each plot are the trial number and the time between the last lever press and the nose poke (i.e. the actual delay interval for the rat on that trial). (F) and (G) Probability density histograms of the direction of movements during delay and response periods. These histograms are taken from the paths shown in (E). They show that during response periods (left plots; blue), the animal traveled a quite consistent and direct path, moving primarily in one direction. In contrast, during the delay period (right plots; red), the animal moved about in almost every direction and the distribution of angles was very uniform. This indicates that the rat did not use a path- or rotational-based strategy as a mnemonic aid during the delays.

Figure 1.

Alternation task apparatus and structure and behavioral results. (A) Illustration of the alternation task chamber and manipulanda. (B) Flow chart of the delayed alternation trial structure. From left to right, after a lever press is made on the right lever, both levers retract and a 10 s delay interval begins. After the delay, a light turns on over the nose-poke port on the opposite end of the chamber, indicating that the port is active. Following a nose poke, both levers extend and a response on the left lever leads to the delivery of a reinforcement pellet. (C) Cumulative performance during delayed alternation sessions. The number of correct trials is on the y-axis and the number of total trials is on the x-axis. Different sessions are indicated by line color. (D) Mean intertrial interval for correct and error trials. Average time between lever presses in seconds on y-axis (error bars = SEM). (E) Movement paths for a series of correct trials during delayed alternation. The light-blue-shaded box on the blue line in plot C shows the trials plotted. The plots show the paths taken after a correct response (red dot) and during the subsequent delay period (red lines), and the paths taken on the ensuing trial between the nose-poke (green dots) and lever press panels (response period; blue lines). Shown in each plot are the trial number and the time between the last lever press and the nose poke (i.e. the actual delay interval for the rat on that trial). (F) and (G) Probability density histograms of the direction of movements during delay and response periods. These histograms are taken from the paths shown in (E). They show that during response periods (left plots; blue), the animal traveled a quite consistent and direct path, moving primarily in one direction. In contrast, during the delay period (right plots; red), the animal moved about in almost every direction and the distribution of angles was very uniform. This indicates that the rat did not use a path- or rotational-based strategy as a mnemonic aid during the delays.

Trajectory Analysis

To illustrate the nature of the behavioral patterns employed, we plotted the paths during a period with 100% accuracy during the session shown in Figure 1E. This period comprised a series of 9 consecutive correct trials that were divided into 4 groups: (1) nose poke to left lever press, (2) nose poke to right lever press, (3) left lever press to nose poke, and (4) right lever press to nose poke. A given trajectory consisted of a series of coordinates, where the ith time point had a position vector ri = (xi, yi). However, the distance between consecutive points on the trajectory was not uniform. In order to uniformly weigh the direction from each part of the trajectory, the trajectory was subdivided into equal length segments. The tangent vector of the jth segment, tj = (tx,j, ty,j), was calculated at its midpoint and the angle of the direction of motion defined to be θj = a tan 2(ty,i, tx,i). The angles were histogrammed to produce the 4 angular distributions for the 4 trajectory categories.

Data Analysis

Constrained PCA

For a description of the CPCA methodology, see Takane and Shibayama (1991) for its original proposal. In Metzak et al. (2011a, 2011b), the CPCA application for functional magnetic resonance imaging data is described and compared with related methodologies. This implementation provided the basis for the model used here. The MATLAB code for this analysis is available from T.S.W. The basic process of CPCA involves applying a multivariate multiple regression to produce a matrix of scores predictable from the independent variables reflecting the task timing. PCA is then performed on this matrix of predicted scores, and the components are related back to the independent (predictor) variables to facilitate interpretation. CPCA is subject to all the limitations and assumptions associated with multiple regression and PCA, such as assuming linear relationships in both, and rotational indeterminacy in PCA. The results of a PCA depend on the scaling of the variables and the variables submitted to the PCA. In this case, these were the predicted scores from a multivariate multiple regression, and it should be noted that the variables most highly predicted from the task timing will dominate the PCA solution, this being the objective of the CPCA. In addition, the PC can be dominated by a small subgroup of neurons. This can be evaluated in our data set in Figure 2C,B by examining the component loadings. In addition, the direction of change in a PC does not always map directly onto the direction of change in firing rate. Rather, the weightings provide information about the degree of similarity of the prevailing pattern at a given time point to the pattern that constitutes the PC.

Figure 2.

CPCA reveals a robust outcome PC. (A) Mean predictor weights (averaged across all subjects and sessions) of a PC that exhibited significant separation during the late delay and outcome periods on correct (black) versus error (gray) trials (error bars = SEM; *P< 0.001). (B) Predictor weights from an example delayed alternation session. Predictor weights from correct (black) and error trials (gray) separated over the course of the delay period, but overlapped during the nose-poke and the lever press epochs. The inserted text in (B) reports the percentage of task-related variance in the ensemble activity for this PC. (C) Component loadings and changes in neuronal firing rates for the PC shown in (B). Each neuron in the recorded population is shown along the x-axis. Component loadings (dotted line) for this PC are plotted along the right y-axis, and the difference in mean iFR during delay (gray) and outcome (black) periods is shown on the left y-axis. (D) Mean predictor weights (averaged across all subjects and sessions) for the outcome PC from continuous alternation sessions (black, correct trial and gray, error trial; *P< 0.01).

Figure 2.

CPCA reveals a robust outcome PC. (A) Mean predictor weights (averaged across all subjects and sessions) of a PC that exhibited significant separation during the late delay and outcome periods on correct (black) versus error (gray) trials (error bars = SEM; *P< 0.001). (B) Predictor weights from an example delayed alternation session. Predictor weights from correct (black) and error trials (gray) separated over the course of the delay period, but overlapped during the nose-poke and the lever press epochs. The inserted text in (B) reports the percentage of task-related variance in the ensemble activity for this PC. (C) Component loadings and changes in neuronal firing rates for the PC shown in (B). Each neuron in the recorded population is shown along the x-axis. Component loadings (dotted line) for this PC are plotted along the right y-axis, and the difference in mean iFR during delay (gray) and outcome (black) periods is shown on the left y-axis. (D) Mean predictor weights (averaged across all subjects and sessions) for the outcome PC from continuous alternation sessions (black, correct trial and gray, error trial; *P< 0.01).

The CPCA was performed separately on each data set. The matrix of the originally recorded data contained 1 column for every neuron and 1 row for every bin in which the neural firing rate was measured. For example, if the average action-potential rate was estimated for a series of 200 ms time windows for a 15 min experimental session, and for 40 neurons, the original data matrix would contain 4500 rows and 40 columns. We will refer to this as the Z matrix. There was also a finite-impulse-response model matrix, referred to hereon as the G matrix, containing information about the timing of all task stimuli and responses. The number of rows in the G matrix is the same as the number of rows in the Z matrix because timing of action-potential rate changes is modeled. The number of columns in G is derived from the number of predictor variables of interest. Within each column of G, we placed a 1 in each row representing a time point that fits into a condition of interest (e.g. time point 1, 2, … of a correct response trial) and a 0 in every other row.

We wanted detailed information about the time course of activity within each network; therefore, we included a separate predictor variable for every time window in a typical trial. The delay period was covered by 40 time windows (40 × 200 ms = 8000 ms = 8 s), the nose-poke period was covered by 10 time windows, and the lever press and reinforcement period was covered by 30 time windows. As a result, the task accuracy G matrix, which coded for correct and incorrect response trials separately, had 160 columns (40 time windows × 2 response types for delay, 10 time windows × 2 response types for nose poke, and 30 time windows × 2 response types for lever press and reinforcement). For the stimulus selectivity G matrix, the same time windows were used, but only correct trials were included, and left lever location trials were coded in columns separate from the right lever location trials.

After both the Z and G matrices were normalized to Z scores, the original data (Z) were then regressed onto our model (G) via the following equation: 

formula
 where 
formula
where C is a matrix of beta weights (1 row per column of G and 1 column per column of Z), and E is a matrix of residual scores. Both the GC matrix and the E matrix have the same dimensions as the original Z matrix. Thus, multivariate multiple regression allows us to split every single measure in the recorded data (a single score in Z) into a predicted score (in the corresponding row and column in GC) and a residual score (in the corresponding row and column of E).

In order to identify neural networks with task-related variations in activity, we performed PCA on the GC matrix. This involves carrying out singular value decomposition in MATLAB Software (Mathworks Inc.) and reducing the dimensionality of the data set by retaining a small number of components (determined by way of a scree plot): 

formula
where U is the left singular vector, V the right singular vector, and D the singular values in the solution with reduced dimensionality. The right singular vectors were rescaled to component loadings and rotated via varimax rotation, and the left singular vectors were rescaled to component scores and rotated in accordance with the right singular vectors. The singular values were rescaled to eigenvalues for the permutation test.

In order to check whether each component we extracted was significant, we used a permutation test with 1000 iterations. For each of the iterations, the permuted data were created via randomly reordering blocks of rows (time bins) within each column (neurons). Block lengths equal to number of time points in a trial were used to preserve autocorrelations up to that length (eighty 200 ms time bins). PCA was then performed on these permuted data. We thus had 1000 sets of eigenvalues derived from random data. For each component, we compared the eigenvalue from the actual GC matrix with the distribution of eigenvalues from randomly permuted GC matrices. If the eigenvalue from the actual GC matrix was greater than 99% of the eigenvalues from the permuted GC matrices, it was considered significant. All reported components were found to be significant by permutation testing, in that their eigenvalues were higher than 100% of the eigenvalues derived from permuted data.

Computation of “predictor weights” is essential for interpreting the timing of the components with respect to the experimental manipulations. This is achieved by computing Pearson correlations between the component scores and the model matrix G. This analysis results in 1 predictor weight for each variable of interest in the task design (each column of G). Since all trials of a particular type are coded into a specific column of the G matrix, predictor weights combine information across all trials (a trial refers to individual sequences of 10–20 s each, involving a single delay followed by a single nose poke followed by a single lever press) within a given experimental run (a given 10–15 min recording session). The predictor weights depict how each component's activity timing corresponded to the timing of trial types (e.g. correct vs. incorrect). The predictor weights also allow inspection of how the component score timing maps onto the delay, nose poke, and lever press trial phases, as time points that cover all 3 phases are coded into the G matrix as separate columns.

Continuous Alternation Data

The analysis of the continuous alternation data sets was essentially the same as the analysis of the delayed alternation data sets, except that there was no delay period. Thus, the task accuracy G matrix had 120 columns: 30 time windows for nose-poke correct, 30 time windows for nose-poke error, 30 time windows for lever press correct and reinforcement, and 30 time windows for lever press error. The stimulus selectivity G matrix was constructed identically as for delayed alternation, except with the time windows used for continuous alternation task accuracy.

Individual Neuron Selectivity Indices

To determine whether individual cells robustly distinguish different periods of interest during trials (e.g. left and right lever presses), we examined the degree of “selectivity” in firing rates during these epochs for these 2 trial types. A selectivity index for each unit i with respect to the type of trial (left or right), outcome of trial (correct or incorrect), and type of lever press (left or right) was obtained by grouping the firing rates into 2 classes (A and B) corresponding to the examined behavioral conditions and computing: 

formula
 where 〈·〉 denotes the “mean.” Sets A and B refer to the 2 behavioral conditions compared, that is, left or right lever trial delay activity. Firing rates for the time bins defining these sets were collected from the 2 s periods in the middle of the second half of the delay period (between 5 and 3 s before the nose poke) and from the 2 s periods surrounding lever presses. Unlike firing rates, we have found that selectivity indices tend to be normally distributed. Therefore, to determine statistical significance of firing rate differences, 2-way ANOVAs were performed with trial type (correct vs. incorrect; and left vs. right lever press) and behavioral epochs (prenose poke/delay; nose poke; lever press; and trial outcome). Follow-up tests were performed if a significant interaction was found.

Results

Behavioral Analysis

A complete description of the delayed alternation task is provided in Figure 1A,B. Briefly, the rat presses a lever and the delay interval begins. After the delay interval has ended, the nose-poke port becomes available at the other side of the chamber and the rat must make a nose poke and then go back to press the lever different from the one it pressed prior to the nose poke. After the lever press is made, the next delay interval begins and so on. Figure 1C shows the cumulative performance from 7 delayed alternation sessions and shows that in each session there were multiple long stretches of correct trials in a row (>5 trials). In total, 80% of all the correct trials from all delayed alternation subjects and sessions occurred during stretches of at least 5 straight correct trials.

One way to solve the task might be to employ a movement or path-based strategy, in which body position or rotational direction could be used to guide the next response (Euston and McNaughton 2006; Cowen and McNaughton 2007). If this were true, one would expect that movement paths and/or the direction of rotation would be uniform and consistent for a given trial type. Figure 1E is a representative example of the locomotor paths taken over 9 consecutive correct trials from a single session. What is plotted is the paths between the nose-poke port and lever press panels (response period; blue lines) and the path taken during the ensuing delay period as the rat moved back toward the nose-poke panel (red lines). The figure illustrates that during the delay interval on the trip from the lever press to the nose-poke panel, the paths were highly convoluted and featured multiple turns that were not consistent between trials. To quantify the movement trajectories, we performed an analysis on the angles of movement. For this analysis, trials were divided into 4 groups based on lever type (left/right) and task epoch (response/delay). The 4 probability density histograms are presented in Figure 1F,G. These plots show a wide distribution in the angles of motion during delay periods, indicating that there were no preferred direction paths or rotational biases being employed. This is in contrast to the response period plots in which the paths were much shorter, direct, and consistent on each trial as the animal went directly from the nose-poke port to the lever press. This analysis illustrates that the animals did not use any obvious body position or simple rotational strategy during the delay to solve the task and therefore implies that the information about which lever to press after the delay interval was maintained mnemonically.

Multiple Single-Unit Recordings

A total of 11 sessions were recorded from 5 animals during which at least 15 error trials were committed, but with overall performance >65% (7 delayed alternation sessions from 4 subjects and 4 continuous alternation sessions from 3 subjects) (Table 1). A total of 385 mPFC neurons were recorded during these sessions, and the mean ensemble size per session was 35 neurons. Histology and tetrode-lowering records confirmed that all neurons were recorded from dorsal regions of the mPFC (either ACC or dorsal prelimbic). Since all recordings were made using moveable tetrodes that were advanced over multiple sessions, it was not possible to precisely identify each neuron's recording location. While the vast majority of the recorded neurons were likely in the ACC, there was a small percentage (∼20%) of tetrodes that did advance into the dorsal portion of the prelimbic cortex. Over all sessions, there was an average of 4.75 neurons per tetrode, with a range of 1–12.

Table 1

Behavioral performance and ensemble recording size for all sessions and subjects

Subject Session no. Task No. of trials % Correct No. of cells 
Delay 60 68 30 
Delay 102 80 50 
Delay 81 69 48 
Delay 62 65 20 
Delay 55 65 29 
Delay 155 66 40 
Delay 47 69 31 
Continuous 94 72 29 
Continuous 82 71 57 
Continuous 50 66 26 
Continuous 82 65 52 
Subject Session no. Task No. of trials % Correct No. of cells 
Delay 60 68 30 
Delay 102 80 50 
Delay 81 69 48 
Delay 62 65 20 
Delay 55 65 29 
Delay 155 66 40 
Delay 47 69 31 
Continuous 94 72 29 
Continuous 82 71 57 
Continuous 50 66 26 
Continuous 82 65 52 

A variant of PCA was used to explore the major determinants of task-related variance in neuronal activity. CPCA magnifies task-related variance by first employing a regression model to remove any activity not temporally locked to task epochs. In this case, the model matrix (G matrix) had nonzero values at the time bins around lever presses, nose pokes, delay periods, and at the time of the outcome, but with near zero values for all other times (see Materials and Methods). After the regression procedure, PCA was performed.

The results presented in the sections below should be interpreted based on the following considerations. First, the changes in factor or component scores of a PC throughout a recording session are directly related to the changes in firing rates of the neurons loading strongly onto that PC. Yet while both strongly positive and strongly negative factor scores indicate that the firing rate has deviated substantially from baseline, firing rates cannot be directly mapped onto the component solution due to rotational indeterminacy. In contrast, whatever the factor score of the PC is at a given time point, neurons with positive component (or factor) loadings change their firing rates in the same direction as the factor score and neurons with negative component loadings change in the opposite direction. In the present study, predictor weights are commonly plotted. In CPCA, the predictor weights correlate the factor scores with the model matrix of task events. If a task event occurs consistently when factor scores are positive, the predictor weights will be large and positive. If a task event occurs consistently when factor scores are negative, the predictor weights will be large and negative. If a task event occurs when factor scores are fluctuating around zero, the predictor weights will be near zero.

A PC that Differentiates Correct and Erroneous Trial Outcomes

The first analysis examined differences in the strength of each PC during the task epochs for correct versus incorrect trials, regardless of action type: for this analysis, trials with left and right lever presses were grouped together. Given the explicit interest in error or outcome detection, we searched among the top 3 strongest PCs for any that exhibited a significant difference in strength at the time of the outcome on correct versus error trials. Across all subjects and sessions, we consistently found a PC that separated correct and error trials at the time of the outcome (Fig. 2A). We termed this the “outcome” PC.

The outcome PC accounted for an average of 19.9±2.2% of the task-related variance across all subjects and sessions. While always present, the relative position of this PC varied and, for instance, was the second largest PC in rat 1 and the third largest PC in rat 2. As shown in Figure 2A,B, the PC itself tended to vary throughout the task for both trial types and was especially strong at the time of each action. However, most notably, this PC only exhibited differences in strength on correct versus error trials during the delay period and approximately 1 s after the lever press response in the outcome period. Specifically, the mean predictor weights from all sessions and subjects showed a sustained significant separation (>1.5 s) between correct and error trials at an average of 6 ± 0.6 s before making the nose poke which ended the delay period [Fig. 2A; error bars = standard error of the mean (SEM)] and again after the lever press for 2 s or more. This PC was very consistent across sessions, and for this reason, we were able to compute the absolute difference between correct and error trial predictor weights for each session and then plot these values as the percentage of the mean maximum bin value across sessions. As can be seen in Figure 2A, separation between the different types of trial outcomes peaked approximately 2 s after a lever press. This was confirmed by significant main effects for both trial type (correct vs. error; F1,120 = 50.7; P< 10−11) and time over the course of a trial (800 ms bins; F19,120 = 14.2; P< 10−5), along with significant interaction effects (F19,120= 3.74; P< 10−6). A Tukey's post hoc test revealed significant differences during the delay and trial outcome periods only (asterisk in Fig. 2A). These results suggest that while the outcome PC varied in strength throughout each trial, it only exhibited differences in strength for correct and error trials toward the end of the delay and at the time of the outcome, but not during the actual lever press itself.

We next examined whether delay period correct/error trial separation was influenced by the presence or absence of reward on the previous trial. We created another model of only correct trials that were grouped based on the outcome of the preceding trial (rewarded or nonrewarded). In all of our data sets, we did not find any significant PC with delay period separation between trials that were preceded by a reward or a lack of reward. This analysis suggests that the delay period correct versus error trial separation seen in the PC presented earlier was likely due to processes related to the current trial and not a retrospective memory of the previous trial or reward.

A Trial Outcome PC also Emerged in a Continuous Version of the Task

The outcome PC had predictor weightings that separated during the delay interval for trials that were correct versus incorrect on the delayed alternation task. However, it is possible that this effect was not specific to a delay interval, but rather that such separation occurs seconds prior to a response in association with response or outcome anticipation. One way to disentangle these possibilities was to investigate whether or not a PC could be found that exhibited significant separation prior to the response on a similar task that lacked a delay. The continuous alternation task was identical to the delayed alternation task save for a lack of a delay interval between trials. This allowed us to explore whether any PCs were similar to the outcome PC, in that they diverged in their strength on correct and error trials in the middle of each trial, prior to the response and outcome. We applied the same “task accuracy” grouping used earlier to the continuous alternation task data set. Using the task accuracy grouping, we again found a PC whose predictor weights consistently differed during the trial outcome period and also just prior to the nose poke on correct versus incorrect trials (Fig. 2D). Using a 2-way ANOVA, comparing mean predictor weights from all animals and sessions with trial type (correct vs. incorrect) and time over the course of a trial (800 ms bins) as main factors, there were highly significant effects of the main factors (type: F1,90 = 21.8, P< 10−5; time: F14,90 = 38.1, P< 10−4) and their interaction (F14,90 = 7.1, P< 10−9). Not surprisingly, the separation prior to the nose poke was considerably shorter in duration than on the delayed alternation, but was nonetheless still present (confirmed with Tukey's post hoc tests; P< 0.001). Thus, a PC was always found on a nondelayed alternation task with the same characteristics as the outcome PC described earlier, except that the separation prior to the nose poke was shorter and less pronounced as a result of the lack of an explicit delay interval. This implies that the separation during the delay interval on the delayed alternation task was not related to the forced delay per se, but rather was a consequence of the fact that the delay preceded the response preparation period.

A Separate PC Dissociates Different Actions

To examine whether any PCs separated different actions, we created a second grouping of trials that compared the 2 lever press sides but only for correct trials. When trials were grouped based on whether the rat made a left or right lever press, we consistently found a PC whose predictor weights separated during the actual lever press itself. Again, the number of this PC varied from one rat to the other, but overall this PC accounted for 13.2 ± 0.63% of the task-related variance. Figure 3A shows the lever press component for left (black) and right (gray) trials from an example session in which there were significantly larger predictor weights at the time of right lever presses. In general, across different animals, the predictor weights were always larger for presses on one lever than those on the other. Accordingly, for each session, whichever lever side was associated with the largest change in predictor weights we termed the “dominant” and the other the “secondary” side. There was variability in the timing of predictor weight changes across sessions and subjects, such that during some sessions, this component distinguished left and right responses a couple of seconds before the response but not afterwards. In others, this pattern continued through reinforcement delivery and consumption, and yet in other sessions, left/right differences only appeared tightly locked with the actual lever press. Across all subjects and sessions, the mean dominant (gray) and secondary (black) lever predictor weights were most separated at the time of the lever press (Fig. 3C). A 2-way ANOVA on mean predictor weights from all animals and sessions with trial type (dominant vs. secondary) and time over the course of a trial (800 ms bins) as main factors found highly significant main effects (type: F1,120 = 418.2, P< 10−13; time: F19,120 = 20.1, P< 10−13) and an interaction effect (F19,120 = 8.4, P< 10−12). A Tukey's post hoc test indicated that there were significant differences between dominant and secondary lever press trials, starting just after the nose poke and continuing for 5 s after the lever press (shown by asterisk in Fig. 3C). Thus, this PC strongly differentiated the 2 specific actions, and this separation persisted through the outcome period, even though rewards were given after all trials in this model.

Figure 3.

An action PC differentiates between trials involving different actions. (A) Example of a lever press PC with significant separation in predictor weights at the time of the lever press for trials involving right versus left lever presses. The lever with the greater response strength (dominant) is shown in gray and the other lever (secondary) is in black. (B) Neuronal component loadings and firing rate differences for the PC shown in (A). Unit number is plotted on the x-axis. Component loadings (gray dotted line) are shown on the right y-axis, and the difference in mean lever press iFR between dominant and secondary levers (black line) is plotted on the left y-axis. (C) Mean action PC predictor weights averaged from all delayed alternation sessions and subjects. This PC significantly differentiated left and right trials following the nose poke and for a period after the lever press (*P < 0.001). (D) Mean predictor weightings for an action PC during continuous alternation sessions. The dominant lever is shown in gray and the secondary is in black (error bars = SEM). These aggregate component predictor weights show that separation only appears at the time of the lever press and is maintained for approximately 2 s afterwards (*P< 0.001).

Figure 3.

An action PC differentiates between trials involving different actions. (A) Example of a lever press PC with significant separation in predictor weights at the time of the lever press for trials involving right versus left lever presses. The lever with the greater response strength (dominant) is shown in gray and the other lever (secondary) is in black. (B) Neuronal component loadings and firing rate differences for the PC shown in (A). Unit number is plotted on the x-axis. Component loadings (gray dotted line) are shown on the right y-axis, and the difference in mean lever press iFR between dominant and secondary levers (black line) is plotted on the left y-axis. (C) Mean action PC predictor weights averaged from all delayed alternation sessions and subjects. This PC significantly differentiated left and right trials following the nose poke and for a period after the lever press (*P < 0.001). (D) Mean predictor weightings for an action PC during continuous alternation sessions. The dominant lever is shown in gray and the secondary is in black (error bars = SEM). These aggregate component predictor weights show that separation only appears at the time of the lever press and is maintained for approximately 2 s afterwards (*P< 0.001).

Although there was separation at the time of the response, using this model grouping, we did not find any significant PC whose predictor weights separated during the delay periods on trials in which the rat made a left versus right lever press. To ensure that these negative results were not confounded by correct responses that occurred merely by chance, we isolated only the long strings of correct trials in a row (>5) from each session and then constructed a lever press side model using only these trials. Again, unlike the outcome PC, in none of the sessions were any significant PCs found that differentiated the delay periods on left versus right trials.

Finally, we also applied the “lever press side” trial grouping on the data from the continuous alternation task and likewise found a PC whose predictor weights strongly dissociated the left and right lever presses in all subjects and sessions at the time of the lever press. As can be seen in the mean predictor weight plot in Figure 3D, there was significant separation starting at the time of the lever press and continuing for approximately 2 s afterwards (Tukey's post hoc tests, P< 0.001). This component was very similar to the stimulus selective lever press component found in delayed alternation task sessions, as can be seen in the mean predictor weights from all subjects and sessions (trial type: F1,90 = 263.2, P< 10−14; time: F14,90 = 34.5, P< 10−14; interaction: F14,90 = 12.2, P< 10−14; Fig. 3D). Therefore, across tasks, there is a well-defined pattern of activity around the time of the lever press that specifies the side of the lever press. Again, no PC was found whose predictor weights differed prior to the response for trials in which a left versus right lever press was made.

Single-Unit Analysis

The ensemble analysis mentioned earlier demonstrated that 2 of the top 3 PCs were present across rats and sessions that separated outcomes or actions. Yet, it was unclear how single neurons contributed to action versus outcome encoding. The present analysis focussed on selectivity indices (d') that compared differences in firing rates of a given neuron during 2 intervals relative to their overall firing rate variance. To determine whether d'-values represented significant differences in firing rates, we performed a 2-way ANOVA for each cell with one factor being behavioral epoch and the other being trial type (separately for correct vs. error or left vs. right lever press; see Materials and Methods). For cells with strong d'-values that had a significant interaction between trial type and behavioral epoch, we again used Tukey's post hoc tests to verify that firing rates during the behavioral epoch of interest were different.

We first examined whether individual neuronal firing rates differentiated correct and error trials during the trial outcome period following a lever press. Over all delayed alternation sessions, we found that 55/241 neurons had strong d'-values and mean firing rates that were significantly higher (according to Tukey's post hoc test; P< 0.05) at the time of the outcome for either correct or incorrect trials (Table 2). The majority of these cells had stronger firing rates after error than after correct responses (60%), consistent with our previous single-unit analysis on a related task (Hyman et al. 2011). For the example unit shown in Figure 4A, there was a large increase in activity immediately after the erroneous lever press was made, whereas on correct trials, there was very little activity during the same period regardless of which lever was pressed.

Table 2

Mean firing rates of neurons grouped by significance of behavioral selectivity

Behavior Delay (1) Outcome (2) Both 1 and 2 All not 1, 2, or both Left/right LP (3) All not 3 
% of cells 15 15 77 28 72 
Overall mean iFR 2.05 ± 0.36 2.16 ± 0.36 3.41 ± 0.75 1.86 ± 0.2 2.43 ± 0.31 1.82 ± 0.2 
Delay mean iFR (correct) 2.33 ± 0.46 2.37 ± 0.47 3.2 ± 0.74 1.76 ± 0.19 NA NA 
Delay mean iFR (error) 2.71 ± 0.49 2.51 ± 0.48 3.82 ± 0.88 1.84 ± 0.19 NA NA 
Outcome mean iFR (correct) 1.99 ± 0.4 2.27 ± 0.46 3.08 ± 0.62 2.71 ± 0.49 NA NA 
Outcome mean iFR (error) 2.15 ± 0.42 2.54 ± 0.48 3.56 ± 0.83 2.15 ± 0.42 NA NA 
LP mean iFR (dominant) NA NA NA NA 4.11 ± 0.61 1.99 ± 0.22 
LP mean iFR (secondary) NA NA NA NA 2.5 ± 0.43 1.61 ± 0.19 
Outcome mean component loading 0.17 ± 0.02 0.13 ± 0.01 0.18 ± 0.03 0.11 ± 0.01 NA NA 
Action mean component loading NA NA NA NA 0.13 ± 0.012 0.12 ± 0.09 
Behavior Delay (1) Outcome (2) Both 1 and 2 All not 1, 2, or both Left/right LP (3) All not 3 
% of cells 15 15 77 28 72 
Overall mean iFR 2.05 ± 0.36 2.16 ± 0.36 3.41 ± 0.75 1.86 ± 0.2 2.43 ± 0.31 1.82 ± 0.2 
Delay mean iFR (correct) 2.33 ± 0.46 2.37 ± 0.47 3.2 ± 0.74 1.76 ± 0.19 NA NA 
Delay mean iFR (error) 2.71 ± 0.49 2.51 ± 0.48 3.82 ± 0.88 1.84 ± 0.19 NA NA 
Outcome mean iFR (correct) 1.99 ± 0.4 2.27 ± 0.46 3.08 ± 0.62 2.71 ± 0.49 NA NA 
Outcome mean iFR (error) 2.15 ± 0.42 2.54 ± 0.48 3.56 ± 0.83 2.15 ± 0.42 NA NA 
LP mean iFR (dominant) NA NA NA NA 4.11 ± 0.61 1.99 ± 0.22 
LP mean iFR (secondary) NA NA NA NA 2.5 ± 0.43 1.61 ± 0.19 
Outcome mean component loading 0.17 ± 0.02 0.13 ± 0.01 0.18 ± 0.03 0.11 ± 0.01 NA NA 
Action mean component loading NA NA NA NA 0.13 ± 0.012 0.12 ± 0.09 
Figure 4.

Single-unit examples of outcome delay and action lever press selective neurons. For each unit example, the raster plot is shown on the left and the peri-event histograms are on the right. For all PSTH, the y-axis shows mean frequency (Hz), whereas the time in seconds relative to the nose poke or lever press for (A) is shown on the x-axis. Left (dark gray) and right (light gray) correct and all error (black) trials are plotted. (A) Unit with strong outcome period activity following errors only. Circles show the time of the nose poke. (B and C) Examples of neurons with selective delay period activity for correct versus error trials. Triangles show the time of the lever press. Raster plots of a neuron with strong delay activity on all correct trials but not on error trials. (D) Example of a unit with a strong lever press preference. (E) Frequency distribution histograms of d‘-values between correct and error trials for outcome (black line; mean = − 0.07 ± 0.016) and delay (gray line; mean = 0.014 ± 0.014) periods. Counts per bin are shown on the y-axis and d-values are on the x-axis. Highly selective neurons had absolute d-values greater than 0.3 (shown by the gray dotted lines).

Figure 4.

Single-unit examples of outcome delay and action lever press selective neurons. For each unit example, the raster plot is shown on the left and the peri-event histograms are on the right. For all PSTH, the y-axis shows mean frequency (Hz), whereas the time in seconds relative to the nose poke or lever press for (A) is shown on the x-axis. Left (dark gray) and right (light gray) correct and all error (black) trials are plotted. (A) Unit with strong outcome period activity following errors only. Circles show the time of the nose poke. (B and C) Examples of neurons with selective delay period activity for correct versus error trials. Triangles show the time of the lever press. Raster plots of a neuron with strong delay activity on all correct trials but not on error trials. (D) Example of a unit with a strong lever press preference. (E) Frequency distribution histograms of d‘-values between correct and error trials for outcome (black line; mean = − 0.07 ± 0.016) and delay (gray line; mean = 0.014 ± 0.014) periods. Counts per bin are shown on the y-axis and d-values are on the x-axis. Highly selective neurons had absolute d-values greater than 0.3 (shown by the gray dotted lines).

When the delay period was analyzed, we again found that 55/241 neurons were highly selective (all P< 0.05) for correct versus error trials, yet only a small percentage [12%; 11/91 neurons selective for delay (n = 36), outcome (n = 36), or both (n = 19)] was consistently selective for trial outcome during both the delay and outcome periods (4 for correct and 7 for incorrect trials). While there were more neurons which preferred error when compared with correct trials (60%) during the delay interval, their absolute mean d‘-values were smaller than those for the neurons that preferred correct trial outcomes [t(53) = 1.8; P< 0.04]. Example raster plots and peri-stimulus event histograms (PSTHs) from a correct trial delay selective neuron and an incorrect trial delay selective neuron are shown in Figure 4B,C.

In keeping with the approach mentioned earlier for ensemble analysis, we next looked at action coding. Selectivity indices were calculated for individual neurons for only correct trials in which a lever press was made on the right versus left lever. A total of 67 of 241 neurons expressed significant d'-values for trials involving right versus left lever at the time of the actual lever press (confirmed by 2-way ANOVA and follow-up tests; all P< 0.05). An example raster plot and PSTH from one such neuron are provided in Figure 4D. When trial outcome (correct vs. incorrect) was also considered, 23 out of 67 (34%) lever press selective cells were found to be significantly selective for specific outcomes. Out of this group of 23 neurons that encoded both outcomes and actions significantly, 13 preferred correct outcomes and the rest preferred errors. Thus, there were a considerable number of neurons that respond only to specific outcomes or only to specific actions as well as a good number that responded to specific combinations of outcomes and actions.

Discussion

The present study analyzed multiple single-unit recordings of ACC neurons from rats performing a delayed or continuous alternation task. CPCA revealed 2 distinct and consistent PCs (i.e. activity patterns) across the networks. The first termed the “outcome” PC showed strong variations at the time of each action on all trials, yet showed differences in strength on correct versus error trials prior to the nose poke and during the outcome period. The second termed the “action” PC showed strong fluctuations throughout the trial, but only differed in strength for trials involving a left versus right lever press starting at the time of the lever press itself and continuing into the outcome period. Subsequent single-unit analyses showed that many individual neurons fired more during the prenose-poke or outcome period on error trials relative to correct trials. Groups of single neurons also fired more for one type of lever press than the other. There were also a significant number of neurons that preferred different outcomes involving specific actions (e.g. a correct outcome following a right lever press), and such neurons exemplify the integrated nature of action and outcome encoding in the ACC.

Through CPCA, we were able to detect 2 dominant and consistent patterns of activity within ACC ensembles. This may seem somewhat surprising given that each session involved different animals with different sets of neurons. The fact that there are such consistent patterns of activity, even though only a tiny fraction of the neural space was ever sampled, suggests that the coding scheme in the ACC is multimodal and massively distributed with most neurons having at least some information about relevant task components (see also Duncan and Owen 2000; Isomura and Takada 2004). In the present study, the outcome PC separated correct and error trials prior to the nose poke and again at the time of the outcome. Although it exhibited strong fluctuations surrounding actions consistently on every trial, it did not differentiate between action types (i.e. right/left lever presses). The second PC, termed the action PC, was in a sense the complement of the outcome PC, in that it differentiated between different types of actions at the time of the actions themselves for all correct trials. Thus of the 2 dominant patterns of activity that were consistent across subjects and sessions, one provided differential information about outcomes as well as information about when actions occurred, without being concerned about differences in the actions themselves, whereas the other was concerned with encoding specific actions for a similar outcome. These findings provide evidence for both independent and integrated action and outcome encoding in ACC ensembles.

It was also useful to know how single ACC neurons fired in response to different outcomes or actions. One way to determine how individual neurons contribute to a given PC would be to look at their factor loadings (e.g. Narayanan and Laubach 2008, 2009). However, this approach was not informative in the present case because here the outcome PC varied throughout the trial and thus the factor loadings reflected how much each neuron contributed to this variance rather than how much each neuron contributed to the differentiation at specific points on correct versus error trials. Therefore, a selectivity index-based analysis was used to provide information about the directionality of firing rate changes of single neurons during the critical periods revealed by CPCA.

Out of the total pool of 241 neurons (from delayed alternation sessions), 55 showed statistically significant differences in firing rate on correct versus error trials at the time of the outcome. Of this subgroup, most fired more on error trials relative to correct trials. An identical number of neurons exhibited differential activity for correct and error trials in the period prior to the nose poke, and again a slight majority of these neurons fired more on error trials. These were not usually the same neurons, however, as only 11 neurons exhibited differential firing activity between correct and error trials during both the prenose-poke and outcome periods. Out of this group of 11 neurons, there were 7 that fired more on error trials. When viewed collectively, the present data suggest that not only does a dominant pattern of activity differentiate correct versus error trials prior to the nose poke and at the outcome period, but also that a significant fraction of individual ACC neurons fire more during these periods on error trials relative to correct trials. In contrast, although a small number of neurons fired during these periods on correct trials, the neurons that did fire tended to fire more vigorously (i.e. had higher average firing rates during these periods on correct trials; correct outcome selective correct trial mean iFR = 3.15 ± 0.57 Hz; error outcome selective correct trial mean iFR = 1.55 ± 0.33 Hz).

The predominance of error-tuned neurons is consistent with numerous past studies showing error-locked responses in the human (Carter et al. 1999) and primate (Ito et al. 2003) medial PFC and with theories of the ACC as an error detection system. These data are also consistent with our recent study that focussed on the properties of error-responsive neurons in the ACC on a similar alternation task (Hyman et al. 2011), as well as other recent single-unit analyses of rat mPFC. Narayanan and Laubach (2008, 2009) found that many medial PFC neurons modulated their firing rates after errors in a tone-cued simple reaction time task. They reported that in some cases, the increased posterror firing persisted into the following trial. Likewise, Totah et al. (2009) found that ACC neurons exhibited a lower firing rate on incorrect trials of a 3-choice serial reaction time attention task. However, a significant increase in the firing of neurons occurred after an incorrect response was made. A similar type of dichotomy was observed in the present study (Fig. 4). Thus, it is evident that the rat mPFC and ACC, in particular, like its counterpart in humans and monkeys (Schall et al. 2002; Ridderinkhof et al. 2004; Botvinick 2007), have, as part of its overall function, some roles in processing errors.

Yet, error-locked firing appeared to be only part of the signal in the ACC. In the present study, a majority of outcome-selective neurons fired more on error trials during the nose-poke period prior to the response or the outcome. It is more difficult to view this activity as indicative of strict error detection since they fired well before the error, during the time of the nose poke. In these alternation tasks, the nose poke only indicates a readiness to make a lever press and is not itself ever associated with any type of specific action or outcome. Therefore, the fact that more neurons exhibited increased activity prior to the nose poke on error trials may reflect some equivocation or uncertainty about which choice was the correct response on that trial. The separation in the outcome PC prior to the nose poke is consistent with the sort of error prediction likelihood calculation proposed by Brown and Braver (2005). They argue that the main function of the ACC is to predict the likely outcome of an action. In the present study, during the delay interval on error trials, there may have been uncertainty about the correct response and the activity of the group of prenose-poke error-responsive neurons may have signaled the strong likelihood of an unsuccessful outcome.

The function of the ACC may be even broader than these theories would suggest. At least in the rat ACC, ensembles enter unique activity state patterns not only at the times of outcomes and actions but at each and every epoch of a variety of different types of cognitive tasks (Lapish et al. 2008; Durstewitz et al. 2010; Ma et al. 2011) as well when a rat is simply exploring different environmental contexts (Hyman et al. 2012). It would seem that the ACC monitors the environment for conditions that require adjustments in control over the course of action (Paus 2001; Behrens et al. 2007; Woodward et al. 2008). More specifically, we would argue that the overall function of this region is to monitor all currently relevant aspects of one's behavior and the environment in order to signal when cognitive effort and adjustments are required, especially in situations when unfavorable outcomes appear likely.

Funding

This work was funded by grants from the Canadian Institute for Health Research and NARSAD to J.K.S.

Notes

Conflict of Interest: None declared.

References

Alexander
WH
Brown
JW
Medial prefrontal cortex as an action–outcome predictor
Nat Neurosci
 , 
2011
, vol. 
14
 (pg. 
1338
-
1344
)
Amiez
C
Joseph
JP
Procyk
E
Reward encoding in the monkey anterior cingulate cortex
Cereb Cortex
 , 
2006
, vol. 
16
 (pg. 
1040
-
1055
)
Behrens
TE
Woolrich
MW
Walton
ME
Rushworth
MF
Learning the value of information in an uncertain world
Nat Neurosci
 , 
2007
, vol. 
10
 (pg. 
1214
-
1221
)
Bernstein
PS
Scheffers
MK
Coles
MG
“Where did I go wrong?” A psychophysiological analysis of error detection
J Exp Psychol Hum Percept Perform
 , 
1995
, vol. 
21
 (pg. 
1312
-
1322
)
Botvinick
MM
Conflict monitoring and decision making: reconciling two perspectives on anterior cingulate function
Cogn Affect Behav Neurosci
 , 
2007
, vol. 
7
 (pg. 
356
-
366
)
Botvinick
MM
Braver
TS
Barch
DM
Carter
CS
Cohen
JD
Conflict monitoring and cognitive control
Psychol Rev
 , 
2001
, vol. 
108
 (pg. 
624
-
652
)
Brown
JW
Braver
TS
Learned predictions of error likelihood in the anterior cingulate cortex
Science
 , 
2005
, vol. 
307
 (pg. 
1118
-
1121
)
Carter
CS
Botvinick
MM
Cohen
JD
The contribution of the anterior cingulate cortex to executive processes in cognition
Rev Neurosci
 , 
1999
, vol. 
10
 (pg. 
49
-
57
)
Chapin
JK
Nicolelis
MA
Principal component analysis of neuronal ensemble activity reveals multidimensional somatosensory representations
J Neurosci Meth
 , 
1999
, vol. 
94
 (pg. 
121
-
140
)
Cohen
RA
Kaplan
RF
Zuffante
P
Moser
DJ
Jenkins
MA
Salloway
S
Wilkinson
H
Alteration of intention and self-initiated action associated with bilateral anterior cingulotomy
J Neuropsychiatr Clin Neurosci
 , 
1999
, vol. 
11
 (pg. 
444
-
453
)
Cowen
SL
McNaughton
BL
Selective delay activity in the medial prefrontal cortex of the rat: contribution of sensorimotor information and contingency
J Neurophysiol
 , 
2007
, vol. 
98
 (pg. 
303
-
316
)
Devinsky
O
Morrell
MJ
Vogt
BA
Contributions of anterior cingulate cortex to behaviour
Brain
 , 
1995
, vol. 
118
 
Pt 1
(pg. 
279
-
306
)
Dum
RP
Strick
PL
The origin of corticospinal projections from the premotor areas in the frontal lobe
J Neurosci
 , 
1991
, vol. 
11
 (pg. 
667
-
689
)
Duncan
J
Owen
AM
Common regions of the human frontal lobe recruited by diverse cognitive demands
Trends Neurosci
 , 
2000
, vol. 
23
 (pg. 
475
-
483
)
Durstewitz
D
Vittoz
NM
Floresco
SB
Seamans
JK
Abrupt transitions between prefrontal neural ensemble states accompany behavioral transitions during rule learning
Neuron
 , 
2010
, vol. 
66
 (pg. 
438
-
448
)
Euston
DR
McNaughton
BL
Apparent encoding of sequential context in rat medial prefrontal cortex is accounted for by behavioral variability
J Neurosci
 , 
2006
, vol. 
26
 (pg. 
13143
-
13155
)
Holroyd
CB
Coles
MG
The neural basis of human error processing: reinforcement learning, dopamine, and the error-related negativity
Psychol Rev
 , 
2002
, vol. 
109
 (pg. 
679
-
709
)
Hyman
JM
Hasselmo
ME
Seamans
JK
What is the functional relevance of prefrontal cortex entrainment to hippocampal theta rhythms?
Front Neurosci
 , 
2011
, vol. 
5
 pg. 
24
 
Hyman
JM
Ma
L
Balaguer-Ballester
E
Durstewitz
D
Seamans
JK
Contextual encoding by ensembles of medial prefrontal cortex neurons
Proc Natl Acad Sci USA
 , 
2012
, vol. 
109
 (pg. 
5086
-
5091
)
Isomura
Y
Ito
Y
Akazawa
T
Nambu
A
Takada
M
Neural coding of “attention for action” and “response selection” in primate anterior cingulate cortex
J Neurosci
 , 
2003
, vol. 
23
 (pg. 
8002
-
8012
)
Isomura
Y
Takada
M
Neural mechanisms of versatile functions in primate anterior cingulate cortex
Rev Neurosci
 , 
2004
, vol. 
15
 (pg. 
279
-
291
)
Ito
S
Stuphorn
V
Brown
JW
Schall
JD
Performance monitoring by the anterior cingulate cortex during saccade countermanding
Science
 , 
2003
, vol. 
302
 (pg. 
120
-
122
)
Izaki
Y
Maruki
K
Hori
K
Nomura
M
Effects of rat medial prefrontal cortex temporal inactivation on a delayed alternation task
Neurosci Lett
 , 
2001
, vol. 
315
 (pg. 
129
-
132
)
Lapish
CC
Durstewitz
D
Chandler
LJ
Seamans
JK
Successful choice behavior is associated with distinct and coherent network states in anterior cingulate cortex
Proc Natl Acad Sci USA
 , 
2008
, vol. 
105
 (pg. 
11963
-
11968
)
Laubach
M
Shuler
M
Nicolelis
MA
Independent component analysis for quantifying neuronal ensemble interactions
J Neurosci Meth
 , 
1999
, vol. 
94
 (pg. 
141
-
154
)
Ma
L
Hyman
JM
Durstewitz
D
Phillips
AG
Seamans
JK
Representation of operant–action sequences by prefrontal cortex ensembles
Soc Neurosci Abs
 , 
2011
 
96.02
Machens
CK
Demixing population activity in higher cortical areas
Front Comput Neurosci
 , 
2010
, vol. 
4
 pg. 
126
 
Matsumoto
K
Suzuki
W
Tanaka
K
Neuronal correlates of goal-based motor selection in the prefrontal cortex
Science
 , 
2003
, vol. 
301
 (pg. 
229
-
232
)
Matsumoto
M
Matsumoto
K
Abe
H
Tanaka
K
Medial prefrontal cell activity signaling prediction errors of action values
Nat Neurosci
 , 
2007
, vol. 
10
 (pg. 
647
-
656
)
Mesulam
MM
A cortical network for directed attention and unilateral neglect
Ann Neurol
 , 
1981
, vol. 
10
 (pg. 
309
-
325
)
Metzak
P
Feredoes
E
Takane
Y
Wang
L
Weinstein
S
Cairo
T
Ngan
ET
Woodward
TS
Constrained principal component analysis reveals functionally connected load-dependent networks involved in multiple stages of working memory
Hum Brain Mapp
 , 
2011a
, vol. 
32
 (pg. 
856
-
871
)
Metzak
PD
Riley
J
Wang
L
Whitman
JC
Ngan
ETC
Woodward
TS
Decreased efficiency of task-positive and task-negative networks during working memory in schizophrenia
Schizophrenia Bull
 , 
2011b
 
[Epub ahead of print]
Narayanan
NS
Laubach
M
Neuronal correlates of post-error slowing in the rat dorsomedial prefrontal cortex
J Neurophysiol
 , 
2008
, vol. 
100
 (pg. 
520
-
525
)
Narayanan
NS
Laubach
M
Delay activity in rodent frontal cortex during a simple reaction time task
J Neurophysiol.
 , 
2009
, vol. 
101
 (pg. 
2859
-
2871
)
Niki
H
Watanabe
M
Cingulate unit activity and delayed response
Brain res.
 , 
1976
, vol. 
110
 (pg. 
381
-
386
)
Paus
T
Primate anterior cingulate cortex: where motor control, drive and cognition interface
Nat Rev Neurosci
 , 
2001
, vol. 
2
 (pg. 
417
-
424
)
Paz
R
Natan
C
Boraud
T
Bergman
H
Vaadia
E
Emerging patterns of neuronal responses in supplementary and primary motor areas during sensorimotor adaptation
J Neurosci
 , 
2005
, vol. 
25
 (pg. 
10941
-
10951
)
Picard
N
Strick
PL
Motor areas of the medial wall: a review of their location and functional activation
Cereb Cortex
 , 
1996
, vol. 
6
 (pg. 
342
-
353
)
Pratt
WE
Mizumori
SJ
Neurons in rat medial prefrontal cortex show anticipatory rate changes to predictable differential rewards in a spatial memory task
Behav Brain Res
 , 
2001
, vol. 
123
 (pg. 
165
-
183
)
Ridderinkhof
KR
van den Wildenberg
WP
Segalowitz
SJ
Carter
CS
Neurocognitive mechanisms of cognitive control: the role of prefrontal cortex in action selection, response inhibition, performance monitoring, and reward-based learning
Brain Cogn
 , 
2004
, vol. 
56
 (pg. 
129
-
140
)
Schall
JD
Stuphorn
V
Brown
JW
Monitoring and control of action by the frontal lobes
Neuron
 , 
2002
, vol. 
36
 (pg. 
309
-
322
)
Shima
K
Tanji
J
Role for cingulate motor area cells in voluntary movement selection based on reward
Science
 , 
1998
, vol. 
282
 (pg. 
1335
-
1338
)
Takane
Y
Shibayama
T
Principal component analysis with external information on both subjects and variables
Psychometrika
 , 
1991
, vol. 
56
 (pg. 
97
-
120
)
Totah
NK
Kim
YB
Homayoun
H
Moghaddam
B
Anterior cingulate neurons represent errors and preparatory attention within the same behavioral sequence
J Neurosci
 , 
2009
, vol. 
29
 (pg. 
6418
-
6426
)
Woodward
TS
Metzak
PD
Meier
B
Holroyd
CB
Anterior cingulate cortex signals the requirement to break inertia when switching tasks: a study of the bivalency effect
Neuroimage
 , 
2008
, vol. 
40
 (pg. 
1311
-
1318
)