Abstract
The dorsal anterior cingulate cortex (dACC) plays a critical role in cognitive control over different domains of tasks. The dACC activities uniformly represent task-generic intensities of control signals across different tasks. However, it remains unclear whether the dACC activities could also encode task identities of control signals across different tasks. If so, how the two types of control information are coherently organized in the dACC? Decision uncertainty is an internally-generated control signal by retrospective monitoring, namely, metacognition, even with no external feedback. We here investigated neural representations of decision uncertainty accompanying three decision-making tasks in the domains of perception, rule-based inference, and memory using trial-by-trial univariate and multivariate analyses on functional magnetic resonance imaging (fMRI) data acquired on human male and female healthy subjects. Our results demonstrated that the dACC represented decision uncertainty commonly across the three decision-making tasks. Further, the multivariate fMRI analyses revealed a mosaic form of neural representations of decision uncertainty across tasks in the dACC. The identity and intensity information was separately represented in two dissociable components, the high-dimensional pattern and the scalar magnitude, of the dACC multivoxel fMRI activities. Lastly, a follow-up behavioral experiment confirmed that this mosaic form of neural representations of parallelly existing decision uncertainty across different tasks should lead to mutual interferences more on the intensity, but less on the identity of control signals. Thus, our findings suggest that the dACC with the mosaic form of neural representations could provide task-generic and task-specific metacognitive control signals to guide appropriate control on different decision-making tasks.
SIGNIFICANCE STATEMENT Metacognition is a form of cognitive control using internally generated decision uncertainty to guide behavior adjustment with no needs of external feedback. Decision uncertainty as a generalizable control signal is commonly encoded in the human dorsal anterior cingulate cortex (dACC) accompanying different decision-making tasks. It remains unknown whether or not the task-specific control information is represented in the dACC. We here revealed that multivoxel functional magnetic resonance imaging (fMRI) activities associated with decision uncertainty in the dACC concurrently represented the identity and intensity information. The mixtures of neural representations of decision uncertainty across different tasks should cause specific interferences on each other. Hence, the neural representations of control signals in the human dACC should be task-generic and task-specific.
Introduction
Cognitive flexibility is fundamental to human behaviors. Precisely monitoring and adaptively adjusting behaviors are critical for optimal control toward intended goals (Norman and Shallice, 1986; Mansouri et al., 2009). A large body of research has reached a consensus that the dorsal anterior cingulate cortex (dACC) plays a central role in encoding such control signals (Miller and Cohen, 2001; Botvinick et al., 2004; Ridderinkhof et al., 2004; Rushworth et al., 2011; Shenhav et al., 2013; Kolling et al., 2016). In particular, the dACC activities reflect the intensity of need for control and appear task-generic (Miller and Cohen, 2001; Ridderinkhof et al., 2004; Rushworth et al., 2011; Shenhav et al., 2013; Mansouri et al., 2017), while the prefrontal regions have been shown to contain rich information for task-specific control (Sakai, 2008; Mante et al., 2013; Stokes et al., 2013; Rigotti et al., 2013). It remains unclear whether the task-specific control information is also represented in the dACC. Most of previous studies have merely investigated neural representations of control signals in a single task. In real-world situations, we often concurrently cope with multiple control-demanded tasks, while simultaneously keeping on monitoring our performance on these tasks and swiftly making appropriate adjustment when necessary. It thus seems that multiple control signals could coexist in the dACC, and each should clearly elicit task-specific control. If the dACC does so, how the task-generic and task-specific control information should be organized in the dACC? One candidate form is that the task-specific identity information and the task-generic intensity information of control signals were separately represented (Shenhav et al., 2013; Kolling et al., 2016; Kragel et al., 2018), while an alternative is that the task-specific identity and task-generic intensity information might be embedded in a high-dimensional fashion (Hunt et al., 2018).
Decisions are often made in the absence of immediate feedback. A sense of high decision uncertainty in that the decision is likely incorrect internally guides behavioral adjustment. On the contrary, a sense of high decision confidence indicates no needs for further control. Thus, decision uncertainty serves as internal control signals for such self-driven cognitive control, namely, metacognition (Flavell, 1979; Nelson and Narens, 1990; Fleming and Dolan, 2012). While a number of studies have identified neural correlates of decision confidence in a diverse set of cortical and subcortical areas accompanying different decision-making tasks (Kiani and Shadlen, 2009; Middlebrooks and Sommer, 2012; Komura et al., 2013; Rutishauser et al., 2015; Hebart et al., 2016; Lak et al., 2017; Miyamoto et al., 2017), the dACC has been consistently shown positively correlated with decision uncertainty in different decision-making tasks (Fleming et al., 2012; Wan et al., 2016; Qiu et al., 2018).
However, overlapping activities in the same region do not necessarily indicate shared neural representations. For example, the dACC region shows overlapping activities in physical pain and social rejection (Eisenberger, 2012). However, the neural representations of different psychological features in the two domains are actually separate (Woo et al., 2014). Most of previous studies have focused on generalizable or separable neural representations of different mental states across different domains. On the contrary, it is thus far unclear whether the neural representations of the same mental state across different domains in the dACC are separable. While the dACC does not have a generalizable representation of control signals across a wide range of cognitive control (Kragel et al., 2018), we here specifically investigated the generalizable neural representations of metacognitive control signals accompanying decision-making tasks across different domains using functional magnetic resonance imaging (fMRI). Our behavioral and neuroimaging results demonstrate that the multivariate neural representations of metacognitive control signals in the dACC across different tasks could be decomposed into two dissociable and complementary components: the multivoxel patterns of neural activities representing the task-specific identity information and the scalar magnitudes of neural activities representing the task-generic intensity information. Thus, such mosaic neural representations in the dACC may convey both identity and intensity information for cognitive control over different tasks.
Materials and Methods
Subjects
Twenty-six subjects (12 males, 24.5 ± 1.5 years old) participated the experiments conducted in the current study. The sample size in the current study was roughly determined by following our previous study using a similar task paradigm (n = 18; Qiu et al., 2018). All subjects were right-handed, had normal or corrected-to-normal vision, and no personal or family history of neurologic or psychiatric disorders. Informed consent was obtained from each individual subject in accordance with a protocol approved by Beijing Normal University Research Ethics Committee.
Sudoku task
In a 4 × 4 grid matrix, each digital number from 1 to 4 should be filled in once and only once in each column, each row, and each corner within four grids. The task used in the current study required the subjects to fill in a target grid with a digital number from 1 to 4 in a partially completed Sudoku puzzle. Each problem had a unique solution. A Sudoku generator (custom codes) was used to create thousands of different Sudoku problems. Problem difficulty was classified into different levels according to the minimum number of logic operation steps required to arrive at the solution. To fairly compare with the memory task, all the presented Sudoku problems had only two possible options (1 or 2). In the control condition, the presented problem had only one possible choice (1 or 2), so that the subject could easily choose the correct option.
Before the fMRI experiment, the subjects were trained to improve their skills to solve the 4 × 4 Sudoku puzzles under experimenters' guidance for at least 2 h/d over a continuous span of 4 d. The subjects practiced solving problems with no time constraints first in two to four runs of 40 problems at an easy difficulty level. Once the average accuracy of a session crossed 90%, she/he then practiced solving problems at the same level within 2 s. Once the average accuracy over the run in a 2 s time frame reached 70%, the participant repeated these steps at the next level of difficulty. After the 4-d intensive training program, each participant had attained high-level proficiency in solving 4 × 4 Sudoku problems in 2 s. The Sudoku problems used in the learning and practicing sessions were different from those used in the behavioral and fMRI experiments.
Random dots motion (RDM) task
In an aperture with a radius of 3° (visual angle), three hundred white dots (radius: 0.08°, density: 2.0%) on a black background were displayed which were moving in different directions at a speed of 8.0°/s. The time span of the movement of each dot lasted for three frames. A portion of dots was moving in the same direction (left or right), while the others were moving in different random directions. The subjects were required to discriminate the net motion direction. According to the proportion of coherently moving dots, discrimination difficulty was classified into different levels, for which the movement coherence varied from 1.6% to 51.2%; the coherence of moving dots in the control condition was 100%.
Memory task
A total of 450 Chinese words each composed of two characters were selected as the materials for the memory test (Hu et al., 2010). Among these words, 360 words that had similar word frequency, stroke number, concreteness and familiarity were used in the task conditions, and 90 words were used in the control condition, of which a half were Chinese numerical characters, while another half were unfamiliar words. The task words were randomly divided into halves. One set of words was used to memorize during the encoding phase (old words), and another set was then used as new words during the retrieving phase. The assignments of the two sets of words were counterbalanced across the subjects. For the words used in the control condition, the numerical words and the unfamiliar words used as the new and old words were also counterbalanced across the subjects. During the encoding phase, 225 words were sequentially and randomly presented for 2 s, and each was repeated twice and the subjects were required to memorize these words as much as possible. This encoding process took 15 min outside the scanner and was completed 30 min before the memory judgment task inside the scanner, in which the subjects was asked to judge whether the presented word was new or old.
Task sequences
The sequences of the three decision-making tasks with two-forced alternative choices were identical. Each trial started with a green cross cue to indicate that the task stimulus would be presented 1 s later. The stimulus was presented randomly for 1, 2, or 3 s, then two options for the choice were presented, and the subject made the choice of an option within 2 s. After the choice was made, four confidence ratings from 1 (lowest) to 4 (highest) were presented, and the subject reported the confidence rating within 2 s. A cue was immediately presented for 1 s to indicate whether the same problem should be decided again or not. In the redecision condition (Fig. 1A), the same stimulus was immediately presented again for 3 s, and the subject made a choice and reported the confidence rating again. In the no_redecision condition, an easy control problem was instead immediately presented for 3 s, and the subject likewise made a choice and reported the confidence rating (always at the level four without decision uncertainty). Each trial lasted for 15 s. In the control condition, the problem used in each task was quite easy and would not evoke the subjects' decision uncertainty. The control trials with the identical task sequence as in the redecision task condition were pseudo-randomly intermingled with the task trials. The three tasks (Sudoku, RDM, and memory) were sequentially conducted in different runs and were counterbalanced across the subjects. Each task consisted of three runs, and each run consisted of 40 task trials and 10 control trials. In the Sudoku and RDM tasks, the task difficulty of each trial was adjusted by a staircase procedure (Levitt, 1971), through which one level was upgraded after two consecutive correct trials with higher confidence (3 and 4), and downgraded by one level after two consecutive incorrect trials with lower confidence (1 and 2), and kept at the same level otherwise, so that the mean accuracy would be controlled at a stable level. Before each experiment, two runs were conducted to allow each subject to practice and stabilize his/her task performance.
Intermixed task
To examine the interference effects on metacognition between two intermixed tasks, all the subjects but one (n = 25) also participated another behavioral task outside the scanner after the fMRI experiment. In this task, the RDM stimuli with three different difficulty levels were randomly inserted into the interval between the first-round decision-making and the second-round decision-making in the Sudoku task. The three RDM difficulty (coherence) levels were calibrated by fitting psychometric function for each individual subject, so that the corresponding performance accuracy was controlled as high as 30%, 50%, and 90%, respectively. Different from the tasks used in the fMRI experiment, here there were four-forced alternative options for both tasks. In total, there were ten runs, and each run consisted of 30 trials.
fMRI parameters
The fMRI experiment was conducted using a 3-T Siemens Trio MRI system with a 12-channel head coil (Siemens) after the 4-d Sudoku training. Functional images were acquired with a single shot gradient echo T2* echoplanar imaging (EPI) sequence with volume repetition time (TR) of 2 s, echo time (TE) of 30 ms, slice thickness of 3.0 mm and in-plane resolution of 3.0 × 3.0 mm2 [field of view (FOV): 19.2 × 19.2 cm2; flip angle (FA): 90°]. Thirty-eight axial slices were taken, with interleaved acquisition, parallel to the anterior commissure-posterior commissure (AC-PC) line.
Behavioral data analyses
A nonparametric approach was employed to assess each individual subject's decision uncertainty sensitivity. The receiver operating characteristic (ROC) curve was constructed by characterizing the incorrectness of the first-round decision-making with different levels of decision uncertainty as the judgment criteria. The level of decision uncertainty was calculated by subtracting the confidence rating from 4. The area under the curve (AUC) was calculated to quantitatively measure how well the subject rated the level of decisional uncertainty matching the actual decision outcome (behavioral AUC; Qiu et al., 2018). We also used the other approaches to measure this quantity (Fleming and Lau, 2014; Fleming, 2017). However, these parameters (e.g., meta-d', HMeta-d) did not show stable correlations across the three tasks. Importantly, as we would like to relate the behavioral uncertainty sensitivity to the neural uncertainty sensitivity (see neural uncertainty sensitivity) measured by the same approach, we then preferred to use the AUC as the metric of individual subject's uncertainty sensitivity.
fMRI data analyses
The basic analyses were conducted with FMRIB's Software Library (FSL; Smith et al., 2004). To correct for the rigid head motion, all EPI images were realigned to the first volume of the first run. Data sets in which the translation motions were larger than 2.0 mm or the rotation motions were larger than 1.0° were discarded. It turned out that no data were discarded in the fMRI experiments. The EPI images were first aligned to individual high-resolution structural images, and were then transformed to the Montreal Neurologic Institute (MNI) space by using affine registration with 12 degrees of freedom and resampling the data with a resolution of 2 × 2 × 2 mm3. A spatial smoothing with a 4 mm Gaussian kernel (full width at half-maximum) and a high-pass temporal filtering with a cutoff of 0.005 Hz were applied to all fMRI data.
Each trial was modeled with three regressors: the first regressor represented the first-round decision-making process, which was time-locked to the onset of the first stimuli presentation with the summation of the presentation period (1, 2, or 3 s) and the differential RT deviated from the mean RT of control trials as the event duration; the second regressor, represented the metacognition process, which was time-locked to the onset of second presentation of the stimuli with the summation of the presentation period (3 s) and the differential RT deviated from the mean RT of control trials as the event duration; the third regressor represented the baseline during the intertrial intervals (ITIs), which was time-locked to the onset of ITI with the ITI duration as the event duration. The level of decision uncertainty accompanying the first-round decision-making and the corresponding RT were implemented as modulators of the second regressor by demeaning the variances of the level of decision uncertainty and RT, and orthogonalizing the level of decision uncertainty with the RT. These regressors were then convoluted with the canonical hemodynamic response function (HRF) with a double-γ function and were linearly superposed together for all the trials to fit with the fMRI signals of each voxel.
For the group-level analyses, we used FMRIB's local analysis of mixed effects (FLAME). Statistical parametric maps for each task were generated by a threshold with Z > 3.1, p < 0.05 with false discovery rate (FDR) correction. Those for the conjunction analysis across the three tasks were generated by Z > 2.6, p < 0.05 with cluster-level family-wise error (FWE) correction.
Region of interest (ROI) definitions
The dACC ROI used for trial-by-trial univariate and multivariate analyses were defined by the voxels that were significantly activated during the second-round decision-making phase in the task trials including both the redecision ansd no_redecision task conditions, in comparison to those during the same phase in the control condition, across the three tasks using the conjunction analysis (Z > 2.6, p < 0.05, cluster-level FWE correction; Fig. 2A).
Trial-by-trial neural activity estimation
For the voxel-based and ROI-based trial-by-trial analyses, the neural activities were estimated at the trial-by-trial basis. That is, each phase of each trial could have an independent β parameter to be convoluted with the canonical HRF and was linearly superposed together to fit with the preprocessed fMRI time series (Rissman et al., 2004; Mumford et al., 2012). To remove the confounding effects from the different response times (RTs) in different tasks (Yeung et al., 2011), we first regressed out the trial-by-trial RTs from the original fMRI time series, and then deconvoluted the residual time series with the trial-by-trial models to obtain the trial-by-trial neural activities (β values). The distances of the trial-by-trial multivariate neural activity vectors (e.g., trial i and trial j) were calculated by the representational dissimilarity [
Neural uncertainty sensitivity
As the dACC trial-by-trial activities were significantly and positively correlated with the levels of decision uncertainty across the three tasks (Fig. 2B), we then used the trial-by-trial estimated neural activities during the second-round decision-making phase as the different criteria to judge whether the decision was incorrect or not (Britten et al., 1992), likewise using the levels of decision uncertainty to calculate the behavioral uncertainty sensitivity (behavioral AUC). By this means, we constructed the neural ROC and measured the neural AUC as the neural uncertainty sensitivity in each voxel or the dACC ROI for each subject. The voxel-wise neural AUC across the subjects in the whole brain were statistically tested in comparison to the chance level (0.5) by the threshold of Z > 2.6, p < 0.05 with cluster-level FWE correction.
Multivoxel pattern analyses (MVPAs)
For ROI-based MVPA, the estimated neural activities at all the voxels of the dACC ROI and the selected white matter (WM) with the same ROI size (505 voxels) for each trial of each task were taken as a multivariate vector. These vectors for each task were then divided into three groups on the basis of the conditions in the second-round decision-making phase (redecision, no_redecision, and control). In the redecision and no_redecision task conditions, the data were further divided into two subgroups on the basis of the binary levels of decision uncertainty (high: confidence ratings of 1 and 2; low: confidence ratings of 3 and 4). Therefore, there were six subgroups of data with three task identity categories multiplied by two intensity categories for either condition. We randomly selected the same number of trials from each subgroup (80% of the minimal number in the six subgroups) and pooled these data together as the training data (∼240 trials in total for each subject), to balance the training trials of different categories. Before entering into training the classifiers, the neural activities of each voxel (feature) were normalized by Z-transformation (Chang and Lin, 2011), to balance their contributions on the classifiers, while the testing data were also processed with the same formula of Z-transformation (the decoding results were quite similar when normalizing the data into the range of [0,1]). We separately trained the weights of the support vector machine (SVM) with linear kernels as the identity and intensity classifiers on the same training dataset. We then used the training SVM classifiers to respectively test the identity and intensity categories of the remaining trials in the same condition (cross-validation) or all the trials in the other conditions. As there were three categories for the identity classifier, the multiclass model with the error-correcting output codes (ECOC) was used (matlab: fitcecoc), while a binary SVM model was used for the intensity classifier (matlab: fitcsvm). 100 iterations of the above decoding procedure were repeated and the average decoding accuracies and decoding sensitivities (d′) were obtained and compared between different testing data.
As the three tasks were performed in different runs, it was expected that the fMRI data originated from the physiological artifacts and other global effects specific to each run might also contribute to decode the task identities. To reveal these effects, we labeled the order of runs in each task and balanced them across the three tasks, so that the task identity information was indiscriminate in the same order of runs. We then used the same procedure of decoding the task identities to decode the run categories. The run decoding effects were largely equivalent to the task identity decoding effects in the white matter (WM) ROI. We then corrected the identity decoding efficiencies by subtracting the corresponding run decoding efficiencies from the originally corresponding values in the dACC ROI, unless mentioned otherwise.
Component decomposition
The trial-by-trial multivoxel fMRI activities within the dACC ROI could be expressed as a general linear model (GLM) as follows
To test the separable components essentially contributing to representations of the identity and intensity information of metacognitive control signals, we serially removed one of the three components in Equation 1 and used the residual signals for the MVPA classification.
Voxel-wise identity-specific selectivity
To assess the spatial organization of neural representations of task identity information, we extracted the SVM weights of all the voxels for each binary classifier used in the multiclass models of the identity classifier trained by the data in the redecision task condition, and calculated the identity-specific selectivity index (ISI) for task i in each voxel of the dACC as follows,
Statistical tests
The imaging statistical tests were specifically described above. The behavioral and decoding results were tested by two-tailed paired t test or one-way repeated ANOVA. For the correlations between two variables, the correlation coefficients were z-transformed and tested using the two-tailed t test. To test for the presence or absence of an effect, we use the two one-sided tests (TOSTs) procedure to test for equivalence and reject the presence of an effect, or remain the presence of an effect (Lakens et al., 2018). We set the equivalence bounds in terms of a standardized effect size (Cohen's d = 0.5). All the reported results remained the presence of an effect, unless specifically mentioned as SE (significant equivalence).
Results
Task paradigm and behavioral results
To examine whether the neural representations of metacognitive control signals in the dACC are task-generic or task-specific, we applied the “decision-redecision” experimental paradigm on three decision-making tasks across the domains of perception, rule-based inference, and memory in the fMRI experiment. At each round of decision-making, the subjects made a choice and immediately reported their confidence (1 for most uncertain and 4 for most certain). The levels of decision uncertainty were calculated by subtracting the confidence ratings from 4. Importantly, there was no external feedback in each task. This paradigm thus allowed the subjects to revise their preceding decisions by metacognition (Fig. 1A). In the perceptual decision-making task, the subjects judged the net motion direction for a patch of random moving dots (the RDM task). In the rule-based decision-making task, the subjects inferred the appropriate digit number filled in a grid of a 4 × 4 Sudoku puzzle (the Sudoku task). In the memory judgment task, the subjects judged whether the presented word was experienced a while ago (the memory task). In all three tasks, there were two stages of decision-making in each trial. In all three tasks, in a half of the trials the same problem that had been presented as the first-round decision-making problem was repeated as the second-round decision-making problem (redecision trials; Fig. 1A). In the other half of trials, the second-round decision-making problem was a different one; it was an easy problem but of the same task type (e.g., 100% RDM stimuli in the RDM task). These are referred to as no_redecision trials. Because they should elicit no uncertainty about the decision taken, no_redecision trials served as one type of control procedure against which to compare the redecision trials. In addition, we used another control procedure; the easy problems, were, on other occasions, presented twice successively. In other words, an easy problem involving no uncertainty was presented twice successively as both the first and second problem. This additional procedure is referred to simply as the “control” procedure. The redecision trials, no-redecision trials, and trials from the control condition version of the same task were pseudo-randomly intermixed. To preclude task switching effects on the neural representations of control signals (Hyafil et al., 2009), the three tasks were separated and randomly intermixed in different runs. The task difficulties were adaptively adjusted by a staircase procedure in the RDM and Sudoku tasks (Levitt, 1971), but not in the memory task because of that the difficulties for memory retrieval were hard to predetermine. Each task consisted of three runs, and totally constituted 60 trials in the redecision task condition and 60 trials in the no_redecision task condition, as well as 30 trials in the control condition.
The task paradigm and behavioral results. A, Three decision-making tasks in different domains (Sudoku, RDM, and memory) were separately conducted in different runs using the “decision-redecision” paradigm, in which the subjects made decisions on the same problem and immediately reported their decision confidence twice in each trial. The level of decision uncertainty was opposite to the reported confidence rating, that is, 4 - the confidence rating. B, The performance accuracy as a function of levels of decision uncertainty for the first-round decision-making, in both the redecision and no_redecision task conditions of the three tasks. C, The correlation of the behavioral AUCs among the three tasks. The ROC curve was constructed by using the levels of decision uncertainty as the judgment criteria for whether the decision would be incorrect, and the AUC was used to measure individual decision uncertainty sensitivity (behavioral AUC). D, Reduction of decision uncertainty by redecision as a function of levels of decision uncertainty accompanying the first-round decision-making. E, The performance accuracy improvement by redecision as a function of levels of decision uncertainty accompanying the first-round decision-making.
Twenty-six subjects participated the fMRI experiment. On average, the performance accuracy in the first-round decision-making was 64.8 ± 12.2% (mean ± SD) in the RDM task, 61.1 ± 7.4% in the Sudoku task, and 70.2 ± 11.4% in the memory task. There was a significant difference of performance accuracy between the memory task and the Sudoku task (two-tailed paired t test, t(25) = 3.84, p = 0.00074, Cohen's d = 0.83), but not in the other two comparisons (ps > 0.05). Nonetheless, the RTs of the three tasks were significantly different in the first-round decision-making [one-way repeated analysis of covariance (ANOVA), F(2,77) = 6.4, p = 0.0027,
The dACC represented metacognitive control signals commonly across the three decision-making tasks
To examine whether neural representations of decision uncertainty are task-generic or task-specific, we assessed the similarities and dissimilarities of neural activities accompanying the three decision-making tasks. Similar as shown in the previous study (Qiu et al., 2018), the same regions in the frontoparietal control network including the dACC were commonly activated during the second-round decision-making phase in both the redecision and no-redecision task conditions, in comparison to the activities during the same phase in the control condition across the three tasks (Z > 2.6, p < 0.05, cluster level FWE correction; Fig. 2A). Further, the dACC activities during the second-round decision-making phase were also parametrically correlated with the levels of decision uncertainty accompanying the first-round decision-making in both the redecision and no-redecision task conditions (Fig. 2B). The dACC activities during the second-round decision-making phase were stronger in the redecision task condition than the no_redecision task condition (Fig. 2C), commonly across the three tasks. Thus, the dACC was a crucial brain area that was associated with metacognition, in particular, encoding decision uncertainty across the decision-making tasks in different domains.
The dACC was associated with metacognition. A, Significant activations during the second-round decision-making phase of the task trials in both redecision and no_redecision conditions in comparison to the trials during the same phase in the control condition across the three tasks. B, Significant activations during the second-round decision-making phase of the task trials in both redecision and no_redecision conditions that were parametrically correlated with the levels of decision uncertainty across the three tasks. C, Significant activations during the second-round decision-making phase of the task trials in the redecision condition in comparison to the no_redecision condition. The conjunction maps were obtained by the threshold of Z > 2.6, p < 0.05, cluster-level FWE correction.
The subjects' behavioral uncertainty sensitivities are determined by the consistency between trial-by-trial neural activities and decision outcomes. We then searched for the brain regions across the whole brain, in which the trial-by-trial neural activities during the second-round decision-making could predict the trial-by-trial decision outcomes of the first-round decision-making. To do so, we measured the voxel-wise neural uncertainty sensitivity by using the trial-by-trial neural activities at each voxel as the different criteria for judging whether the first-round decision would be incorrect (neural AUC; Fig. 3A), similar to the calculation of behavioral AUCs. Throughout the whole brain, the voxels whose neural AUCs across the subjects were significantly larger than the chance level (0.5) were concentered in the dACC and precuneus regions commonly in the three tasks (Z > 2.6, p < 0.05, cluster-level FWE correction; Fig. 3B, for each task, Fig. 3C, for the conjunction analysis across the three tasks). Further, in the dACC ROI defined by the conjunction analysis of comparisons between the task trials and the control trials across the three tasks as shown in Fig. 2A, the voxel-wise mean neural AUCs averaged over the subjects were predominantly larger than the chance level (0.5, dotted lines) in each task (Fig. 3D), indicating that the trial-by-trial neural activities in the voxels of the dACC ROI were predominately predictive of the decision outcomes. Indeed, the neural AUC averaged in the dACC ROI was highly correlated with the behavioral AUC across the subjects in each task (r = 0.40, t(24) = 2.22, p = 0.018 in the Sudoku task; r = 0.42, t(24) = 2.36, p = 0.013 in the RDM task; r = 0.65, t(24) = 4.36, p = 0.000091 in the memory task; Fig. 3E). Taken together, the fMRI activities in the dACC showed task-generic neural representations of decision uncertainty accompanying the three decision-making tasks in different domains. In addition, we did not find any brain region throughout the whole brain in which the fMRI activities were significantly correlated with decision uncertainty but prioritized to any of the three tasks (Z > 2.6, p < 0.005, uncorrected).
Neural uncertainty sensitivity (AUC) in the dACC. A, The ROC curves for the three tasks in a representative voxel in the dACC of a representative subject (AUCSudoku: 0.72; AUCRDM: 0.70; AUCmemory: 0.74). The trial-by-trial estimated fMRI activities during the second-round decision-making phase in the redecision and no_redecision task conditions in each voxel could be used to construct the ROC curve by selecting different strengths of fMRI activities as the judgment criteria for whether the decision would be incorrect. The area under the ROC curve was measured as the neural AUC of each voxel. B, The voxels with the neural AUCs across the subjects significantly larger than the chance level (0.5) in each task (Z > 2.6, p < 0.05, cluster-level FWE correction). C, Overlapping voxels with the neural AUCs across the subjects significantly larger than the chance level (0.5) across the three tasks (the conjunction analysis, Z > 2.6, p < 0.05, cluster-level FWE correction). D, The histograms of the mean neural AUCs averaged over the subjects in all the voxels (505 voxels) of the dACC ROI in the Sudoku, RDM, and memory tasks, respectively. The dotted lines indicate the chance level of neural AUC (0.5). The dACC ROI was selected from Figure 2A. E, The correlations between the neural AUCs and behavioral AUCs in each task.
Concurrent neural representations of identity and intensity of metacognitive control signals in the dACC
The univariate analyses described above could not discriminate the task identities possibly encoded in multivoxel activity patterns. We then conducted MVPAs (Kriegeskorte et al., 2006), to reveal the characteristics of high-dimensional neural representations of decision uncertainty in the dACC (Fig. 4A). In particular, we hypothesize that the task identity information of decision uncertainty is represented by the pattern of high-dimensional activities and the intensity information of decision uncertainty is represented by the scalar magnitude of high-dimensional activities in the dACC (Fig. 4B). First, we found that the representational dissimilarities (1 – r) between the trial-by-trial multivoxel fMRI activities in the dACC ROI within the same task were smaller than those between the different tasks (ps < 0.05; Fig. 4C). Further, the multivoxel fMRI activities in the three tasks were largely segregated into three separate clusters using MDS (Fig. 4D).
Mosaic neural representations of multiple metacognitive control signals in the dACC. A, The mosaic neural representation model for the identity information (upper) and the distribution of identity-specific selective voxels in the dACC region of a representative subject (below). B, The schematic model of neural representations of multiple metacognitive control signals accompanying the decision-making tasks in the three domains. The magnitudes of multivariate neural activities represent the intensities of metacognitive control signals, while the vectors of multivariate neural activities represent the identities of metacognitive control signals. C, The trial-by-trial (upper) and mean (below) representational dissimilarity (1 – r) of multivariate fMRI activities of trials within and between the three tasks in the dACC ROI of a representative subject. D, Illustration of separation of multivariate neural activities in the dACC ROI of a representative subject within three dimensions using MDS. E, The identity decoding sensitivities (d′) after subtracting the run decoding sensitivities changed with the ROI sizes of the dACC (black line) and WM (red line) regions. F, The intensity decoding sensitivities (d′) changed with the ROI sizes of the dACC (black line) and WM (red line) regions. The green dots represent the results at the ROI size of the original dACC ROI (505 voxels). G, The identity decoding sensitivities were dependent on the decision uncertainty levels. The training and testing data were from the redecision task condition; ***p < 0.001. Error bars indicate SEM across the subjects.
To further test our hypothesized model, we separately decoded the identities and intensities of decision uncertainty among the three tasks from the defined dACC ROI (505 voxels). For the sake of simplicity, the four-scale levels of decision uncertainty were collapsed into binary categories (high uncertainty: the confidence ratings of 1 and 2; low uncertainty: the confidence ratings: 3 and 4). In order to balance the number of training trials in each condition, we randomly selected the same number of trials in the three categories of identities × the two categories of intensities (80% of the least number of trials in the six subgroups), and used these trials (∼240 trials) to train two independent SVM classifiers for the identity and intensity categories, respectively. We then used the training SVM classifiers to test the remaining dataset of the same subgroup (cross-validation) and the whole dataset of the other subgroups for each subject. The decoding accuracies and decoding sensitivities (d′) were averaged from the results obtained from 100 iterations of the above procedure. We used the SVM classifiers trained by the fMRI activities during the second-round decision-making phase in the redecision and no_redecision task conditions, respectively, and then tested on the following conditions: (1) the same phase in the same conditions (cross-validation); (2) the first-round decision-making phase in the same conditions; (3) the same phase in the control condition (Fig. 8).
However, as the three tasks were separately performed in the different runs, the task identity classifications might be confounded by the coincidences of run orders. To compensate the confounding effects from the run orders, we rectified the identity decoding sensitivities by subtracting the corresponding run decoding sensitivities in the same ROIs (see Materials and Methods). The fMRI signals in the dACC ROI, but not in the WM ROI, considerably decoded both the identity and intensity of decision uncertainty across the three tasks (Fig. 4E,F).
The SVM classifiers trained by the fMRI activities in the redecision and no_redecision task conditions could decode the identity categories from the testing trials in the same conditions, respectively. The original three-category decoding accuracies were 64.2 ± 5.4% in the redecision condition, and 58.9 ± 6.6% in the no_redecision condition (the chance level: 33.3%; Fig. 5A, “total” panel), and the rectified decoding sensitivities (d′) were 0.58 ± 0.16 in the redecision condition (postre-postre) and 0.46 ± 0.18 in the no_redecision condition (postno-postno; Fig. 6A, “original” panel). The collapsed two-category decoding performance that discriminated each individual task from the other two tasks was quite similar across the three tasks (Fig. 5A). However, these classifiers could not decode the task identities from the fMRI activities during the first-round decision-making phase (postre-decino; Fig. 6A). Because of the negative correlation between the neural activities between the two continuous decision-making phases in the same trials, we tested the trials alternately between the redecision and no_redecision task conditions. These results thus suggest that the fMRI activities containing task identity information in the dACC was specific to the metacognition process during the second-round decision-making phase, but not the decision-making process during the first-round decision-making phase. Further, these classifiers could not decode the task identities in the control condition (postre-postctrl and postno-postctrl; Fig. 6A), while the identical control problems were also presented during the same phase in the no_redecision task condition. Thus, the task identity information encoded in the dACC during the second-round decision-making phase in the no_redecision task condition should be relevant to decision uncertainty accompanying the first-round decision-making, rather than the ongoing decision-making task during the second-round decision-making phase.
The identity and intensity decoding accuracies using the same training classifiers for binary task categories. A, The identity decoding accuracies (the chance level: 5/9 × 100 = 55.5% in discriminating binary categories, while 1/3 × 100 = 33.3% in discriminating ternary categories). B, The intensity decoding accuracies (the chance level: 50%). The conventions are the same as in Figure 6. ns, no significance; SE, significant equivalence; **p < 0.01, ***p < 0.001. Each dot represents the data of each subject.
The decoding accuracies and sensitivities of the intensity SVM classifiers trained by the same dataset as used in the identity classifiers in the redecision task condition were also significant, but much weaker (Figs. 5B, 6E). Again, the intensity categories could only be decoded during the second-round decision-making phase, but not during the first-round decision-making phase (postre-decino; Figs. 5B, 6E). However, the decoding accuracies and sensitivities of the intensity SVM classifiers were not significant in the no_redecision condition (postno-postno; Figs. 5B, 6E). Importantly, the intensity SVM classifier trained by the dataset of each task could also decode the intensity categories in the other two tasks (Fig. 7A). These results confirm that the neural representations of decision uncertainty should be task-generic. In addition, the decoding AUC was significantly correlated with the behavioral AUC across the subjects in the tasks except for the RDM task (r = 0.37, t(24) = 1.95, p = 0.030 in the Sudoku task; r = –0.10, t(24) = 0.49, p = 0.31 in the RDM task; r = 0.35, t(24) = 1.91, p = 0.040 in the memory task; Fig. 7B). Altogether, these decoding results by MVPA consistently reveal that the multivoxel fMRI activities in the dACC should contain the identity and intensity information of metacognitive control signals across the different tasks.
Dissociable components of neural representations in decoding the identity and intensity information of metacognitive control signals in the dACC. The decoding sensitivities (d′) of identity and intensity categories in different conditions using trial-by-trial neural activities after removing different components. The training data were normalized by the Z-transformation. A, The original identity decoding sensitivities without removing any component. B, The identity decoding sensitivities after the component of mean activity of all the voxels in each trial was removed. C, The identity decoding sensitivities after the component of mean activity of all the trials in each voxel was removed. D, The identity decoding sensitivities after the component modulated by decision uncertainty was removed. E, The original intensity decoding sensitivities without removing any component. F, The intensity decoding sensitivities after the component of mean activity of all the voxels in each trial was removed. G, The intensity decoding sensitivities after the component of mean activity of all the trials in each voxel was removed. H, The intensity decoding sensitivities after the component modulated by decision uncertainty was removed. postre-postre: the training data from the second decision-making phase in the redecision task condition; the testing data from the same condition; postre-postctrl: the training data from the second decision-making phase in the redecision task condition; the testing data from the same phase in the control condition; postre-decino: the training data from the second decision-making phase in the redecision task condition; the testing data from the first decision-making phase in the no_redecision task condition; postno-postno: the training data from the second decision-making phase in the no_redecision task condition; the testing data from the same condition; postno-postctrl: the training data from the second decision-making phase in the no_redecision task condition; the testing data from the same phase in the control condition; postno-decire: the training data from the second decision-making phase in the no_redecision task condition; the testing data from the first decision-making phase in the redecision task condition. ns, no significance; SE, significant equivalence; *p < 0.05, **p < 0.01, ***p < 0.001. Each dot represents the data of each subject.
The intensity decoding efficiencies across the tasks. A, The intensity classifier trained by one task could decode the intensity categories in the other two tasks (ps < 0.05). B, The decoding AUCs were correlated with the behavioral AUCs in the Sudoku and memory tasks (r = 0.37, t(24) = 1.95, p = 0.030 in the Sudoku task; r = 0.35, t(24) = 1.91, p = 0.040 in the memory task), but not in the RDM task (r = –0.10, t(24) = 0.49, p = 0.31). ns, no significance; *p < 0.05, **p < 0.01, ***p < 0.001. Error bars indicate SEM across the subjects.
Despite that the task identities could be discriminated from the multivoxel dACC activities during the metacognitive processes, the associated fMRI activities might be coincident, but independent of decision uncertainty. Contrastingly, the identity and intensity decoding sensitivities were not dissociated. The identity decoding sensitivities in the trials with high decision uncertainty were significantly higher than those in the trials with low decision uncertainty (two-tailed paired t test, t(25) = 3.89, p = 0.00033, Cohen's d = 0.85; Fig. 4G). These results then provide evidence against the independence between the neural representations of the identity and intensity information. Instead, the identity information should be entrained in the neural representations of decision uncertainty in the dACC (Fig. 4B).
The decoding efficiencies on multivoxel fMRI activities are critically dependent on the number of features or voxels (the ROI sizes). To examine the dependency of identity and intensity decoding efficiencies on the ROI sizes, we selected different ROI sizes in the dACC (from 10 to 1000 voxels, voxel size: 2 × 2 × 2 mm3) by adjusting the different significance thresholds for the contrast of interest (Fig. 2A). The identity decoding sensitivities in the redecision task condition (postre-postre) stably increased with the ROI sizes, considerably larger than those in the WM ROI with the corresponding same ROI sizes (Fig. 4E). On the contrary, the intensity decoding sensitivities on the same datasets in the dACC ROI did not change with the ROI sizes (Fig. 4F). Importantly, there were no systematic spatial organizations of voxels showing significant ISI in the dACC ROI consistently in individual subjects (see Equation 2 in Materials and Methods; Fig. 8). These results thus support our hypothesis on the mosaic form of neural representations of decision uncertainty in the dACC across the tasks in different domains (Fig. 4A).
Spatial distributions of identity-specific selective voxels in the dACC. The colors indicate the tasks to which the voxels preferred.
Dissociating neural representations of identity and intensity of metacognitive control signals in the dACC
To further dissect the fMRI activities specifically representing the identity and intensity information of metacognitive control signals in the dACC, we decomposed the fMRI activities in the dACC into separate components and assessed the impact from each component on the decoding efficiencies by separately removing the corresponding component from the dACC multivoxel fMRI activities. If a component was critical for representing the identity and/or intensity information, then the decoding efficiency should be severely impaired after the component was removed. As shown in Equation 1 in Materials and Methods, the multivoxel fMRI activities in the dACC were decomposed into three components: (1) the mean fMRI activity across all the voxels in each trial, representing the trial-by-trial homogeneous fMRI activities; (2) the mean activity across all the trials in each voxel, representing the basic voxel-based fMRI activities of the same task; and (3) the component modulated by the strengths of decision uncertainty across all the trials in each voxel (Fig. 9).
The schematic showing dissociable components of multivoxel fMRI activities in the dACC. (1) The mean fMRI activity across all the voxels in each trial. (2) The mean activity across all the trials in each voxel. (3) The component modulated by the levels of decision uncertainty across all the trials in each voxel.
First, after the component of mean activity across all the voxels in each trial was removed, the decoding efficiencies for both the identity and intensity categories remained intact (Fig. 6B,F). These results indicate that both quantities were independent of these trial-by-trial mean fMRI activities, but were dependent on the relatively differential neural activities across the voxels in the dACC. Second, after the component of mean activity across all the trials in each voxel was removed, the identity SVM classifier could not further decode the identity categories (Fig. 6C), but the decoding efficiencies of the intensity SVM classifier remained intact (Fig. 6G). Third, after the component modulated by the levels of decision uncertainty across all the trials in each voxel were removed, the decoding efficiencies of the identity SVM classifier were intact (Fig. 6D), but the intensity SVM classifier could not further decode the intensity categories (Fig. 6H). This double dissociation of dependencies of decoding efficiencies for the identity and intensity categories on the latter two independent components strongly suggests that the two components respectively encoding the identity and intensity information were dissociated and complementary. In particular, the identity information was represented by the pattern of multivoxel fMRI activities, while the intensity of decision information was represented by the magnitudes of multivoxel fMRI activities, consisting with our proposed model (Fig. 4B).
Interferences from concurrent metacognitive control signals
We so far revealed a mosaic form of neural representations of metacognitive control signals in the dACC accompanying the decision-making tasks in different domains. Accordingly, the linear superposition of mosaic neural representations of multiple metacognitive control signals in the dACC might mutually interfere the readout of the intensity and identity information of each control signal that coexists in the dACC (Fig. 4B). To test this prediction, we conducted a follow-up behavioral experiment, in which the RDM task was inserted in the interval between the first-round and second-round decision-making in the Sudoku task. The Sudoku task difficulties were controlled by a staircase procedure, and the RDM stimuli with three different difficulty levels (coherences) were randomly intermixed (Fig. 10A). Such an experimental design allowed us to specifically measure the interference effects from the different levels of decision uncertainty induced by the interrupted task (RDM) on the behavioral performance on the main task (Sudoku) by metacognition during the second-round decision-making (Fig. 11A), while the Sudoku performance accuracies during the first-round decision-making were controlled among the three difficulty levels (Fig. 11B), and between the two decision uncertainty levels (Fig. 11C), of the RDM task.
Interferences from concurrent metacognitive control signals. A, The task sequence of the behavioral experiment to examine interferences from the concurrent decision-making task. The RDM task with only one-round decision-making (within red broken line) was inserted in the interval between the first-round and second-round decision-making in the Sudoku task. B, The performance accuracy changes by redecision in the Sudoku task affected by the levels of difficulty in the RDM task. C, The RT changes by redecision in the Sudoku task affected by the levels of difficulty in the RDM task. D, The confidence changes by redecision in the Sudoku task affected by the levels of difficulty in the RDM task. E, The slope of interference effects on the confidence changes by redecision in the Sudoku task changing with the RDM difficulty levels were positively correlated with the behavioral uncertainty sensitivity (behavioral AUC) in the Sudoku task across the subjects (r = 0.37, t(23) = 2.03, p = 0.029). F–H, The same effects were observed according to the levels of decision uncertainty in the RDM task. I, The slope of interference effects on the confidence changes by redecision in the Sudoku task changing with the RDM decision uncertainty levels were positively correlated with the behavioral uncertainty sensitivity (behavioral AUC) in the Sudoku task across the subjects (r = 0.35, t(23) = 1.90, p = 0.037). The insets in E, I show the interference effect size by the levels of difficulty and decision uncertainty in the RDM task, respectively. ns, no significance; SE, significant equivalence; *p < 0.05, **p < 0.01, ***p < 0.001. Each dot represents the data of each subject.
The behavioral performance in the first-round Sudoku decision-making at the different levels of difficulty and decision uncertainty of the RDM task. A, The performance accuracies, RTs and reported confidence at the three levels of difficulty in the RDM task (easy: 90%; middle: 50%, and difficult: 30%, the chance level: 25%). B, C, The performance accuracies, RTs and reported confidence in the first-round Sudoku decision-making were not differential across the three levels of difficulty (B) or the binary levels of decision uncertainty (C), of the RDM task. ns, no significance; ***p < 0.001. Each dot represents the data of each subject.
The same group of subjects (n = 25, one subject dropped out) also participated this behavioral experiment. The Sudoku performance accuracy change by redecision was likewise dependent on the levels of decision uncertainty accompanying the first-round decision-making (Fig. 12). Critically, both the Sudoku performance accuracy changes and the corresponding RT changes were not significantly affected by the RDM task difficulty levels [one-way repeated ANOVA, F(2,74) = 1.13, p = 0.33,
Both the accuracy changes and the confidence changes by redecision were dependent on the levels of decision uncertainty accompanying the first-round Sudoku decision-making at each level of difficulty (A, B) and decision uncertainty (C, D). low: low uncertainty; high: high uncertainty; ns, no significance; ***p < 0.001. Each dot represents the data of each subject.
The (high or low) levels of decision uncertainty accompanying the first-round Sudoku decision-making also interfered the following RDM task on the reported confidence (A), but not on the performance accuracy (B) or the RTs (C). low: low uncertainty; high: high uncertainty; ns, no significance; ***p < 0.001. Each dot represents the data of each subject.
Further, the interference effects from the concurrent decision uncertainty might be also dependent on the subject's uncertainty sensitivity (i.e., behavioral AUCSudoku) that reflected the robustness of representations in resisting the noise influences. The higher the behavioral AUCSudoku was, the lower the interference effect on the decision uncertainty change was [r = 0.37, t(23) = 2.03, p = 0.029 (Fig. 10E); r = 0.35, t(23) = 1.90, p = 0.037 (Fig. 10I)].
Discussion
In the current study, we investigated the task-specific neural representations of generalizable metacognitive control signals in the dACC across multiple decision-making tasks in different domains using trial-by-trial univariate and multivariate analyses on the fMRI data. The recent fMRI studies have revealed spatially separable neural representations of different mental states and psychological features in the dACC, although the fMRI activities were heavily overlapping with one another (Woo et al., 2014; Kragel et al., 2018). It still lacks such evidence for separable neural representations of the same mental state or psychological feature, particularly control signals, across different tasks. The current study, for the first time, demonstrates that neural representations of the same mental state of metacognitive control signals, that is, decision uncertainty, were separable in high-dimension neural space, but not in the physical space of the dACC.
Consistent with the general idea that metacognition is a form of cognitive control (Flavell, 1979; Nelson and Narens, 1990; Fleming and Dolan, 2012), the dACC was first found to be commonly associated with metacognition, particularly decision uncertainty as metacognitive control signals accompanying three decision-making tasks in the domains of perception, rule-based inference, and memory. The univariate analyses showed that the trial-by-trial fMRI activities in the voxels of the dACC were predictive of the trial-by-trial decision outcomes in each task although there was no immediate feedback, in that the voxel-wise neural uncertainty sensitivities (neural AUCs) in the dACC were predominantly larger than the chance level. Further, the neural AUCs in the dACC were also predictive of the subjects' behavioral uncertainty sensitivities (behavioral AUCs) in each task. These results thus confirm that neural representations of metacognitive control signals in the dACC should be task-generic. The task-generic neural representations of metacognitive control signals in the dACC provide a neural account for the behavioral phenomenon of confidence influences or “confidence leak” between interleaved tasks (Gardelle and Mamassian, 2014; Rahnev et al., 2015), and even between the tasks across different domains, as demonstrated in the current study.
On the other hand, our multivariate analyses further revealed that neural representations of metacognitive control signals in the dACC might be also task-specific. The multivoxel fMRI activity patterns of metacognitive control signals in the dACC were differential and could be discretized among the tasks in different domains. The dACC has been shown to broadly represent a variety of quantities, such as task rules (Johnston et al., 2007; Sakai, 2008; Matsuzaka et al., 2012), action values (Hayden and Platt, 2010; Cai and Padoa-Schioppa, 2012), action outcomes (Matsumoto et al., 2003; Rangel and Hare, 2010; Rushworth et al., 2011), pains (Eisenberger, 2012; Woo et al., 2014), and so on. Critically, the dACC also simultaneously represents multiplexed separable signals (Kolling et al., 2016; Hunt et al., 2018). Hence, it seems not surprising that the fMRI activities in the dACC could specify the task identity. However, our current results provided multifaceted evidence opting for that the task identity information was intrinsically entrained in the neural representations of metacognitive control signals. First, the task identity information could only be readout from the fMRI activities during the second-round decision-making phase, rather during the first-round decision-making phase, suggesting that the task identity information was specific to the postdecision activities in the dACC, which were highly associated with metacognitive control signals. Second, it is of interest to note that retrospectively monitoring decision uncertainty also occurred in the no_redecision condition. This evidence suggests that metacognitive monitoring might be automatically driven (Wan et al., 2016). Despite of no demand for metacognitive control, neural representations of decision uncertainty in this condition remained yet task-discriminant. Third, the readout of task identities was dependent on the levels of decision uncertainty. The decoding sensitivities (d′) of task identities in the dACC were higher at the high level of decision uncertainty than those at the low level of decision uncertainty. Fourth, the identity discriminations were, but the intensity discriminations were not, dependent on the number of features used for decoding. This evidence indicates that the identity information should be rooted on the heterogeneity of multivoxel activities, while the intensity information should be dependent on the homogeneity of multivoxel activities. Fifth, the selective deficits on decoding efficiencies of task identities caused by deleting the corresponding component from the multivoxel fMRI activities in the dACC further demonstrated that discriminations of task identities were crucially dependent on the multivoxel representation pattern (the unit vector), while the intensities of metacognitive control signals were complementarily dependent on the scalar magnitudes of the voxel-wise dACC activities.
These pieces of evidence thus argue against that the representations of task identity information were dissociated from the representations of generalizable metacognitive control signals in the dACC. Instead, the identity and intensity information of metacognitive control signals might be registered together in the dACC with a mosaic form of high-dimension neural representations. Further, the follow-up behavioral experiment also validated the specific predictions by the mosaic neural representation model that concurrent neural representations of metacognitive control signals in the dACC should cause interferences on metacognitive control (i.e., performance change), although the effects should be much less than those on metacognitive monitoring (i.e., confidence change).
Separate neural systems should be involved in the decision-making processes in the different domains of perception, rule-based inference, and memory. Therefore, the metacognitive control signals encoded in the dACC should have distinct origins from these diverse decision-making processes. It remains unclear whether the distributed activities representing multiple metacognitive control signals in the dACC detected by fMRI are predominantly input-dependent or output-dependent, as well as layer-specific (Sajad et al., 2019). Nonetheless, given that the dACC is extensively innervated with the prefrontal regions and the associated cortical and subcortical areas (Vogt et al., 2004; Beckmann et al., 2009; Morecraft et al., 2012). The differential neural representation patterns of metacognitive control signals in the dACC could have differential anatomic and functional connectivity with the prefrontal regions in supporting task-specific metacognitive control in different domains of tasks. The control signals in the dACC have been suggested to be associated with the theta-band neural activities (Womelsdorf et al., 2010; Cavanagh et al., 2012), a characteristic feature of local neural computations used for information communications between long-range regions (Voytek et al., 2015), which may allow the task-specific control information in the dACC to be precisely transmitted to the prefrontal regions for task-specific metacognitive control.
Several previous behavioral and imaging studies have also addressed the relevant issue on whether metacognition is task-generic or task-specific (Baird et al., 2013; McCurdy et al., 2013; Faivre et al., 2018; Morales et al., 2018). But the accumulating behavioral and neural evidence remain controversial, suggesting both task-generic and task-specific control signals might coexist in different regions of the human brain. We here used trial-by-trial univariate and multivariate approaches with higher sensitivity and specificity to show that the dACC, previously believed to represent task-generic metacognitive control signals (Fleming et al., 2012; Wan et al., 2016; Morales et al., 2018; Qiu et al., 2018), might also encode the task-specific information of metacognitive control signals. While we here do not intend to argue against the coexistence of task-specific neural representations in the other brain regions, the mosaic neural representations in the dACC alone could provide sufficient information about the necessity, identity, and intensity of control signals for cognitive control.
Cognitive control is to optimize the cognitive processes motivated by the intended goals (Friston, 2010; Botvinick and Braver, 2015). One form of optimization is driven by intrinsic motivation to reduce uncertainty (Wan et al., 2016; Mansouri et al., 2017; Qiu et al., 2018), as also illustrated in the current study. Another form of optimization is driven by external motivation of utility expectation maximization (Montague et al., 2004; Shenhav et al., 2013; Fouragnan et al., 2019), so that the processes of cognitive control consist of monitoring the discrepancy between the outcome and the expected utility (i.e., prediction errors) and subsequently adjusting the cognitive strategy to increase the outcome or decrease the cost. This form of cognitive control is thus strongly associated with the dopaminergic reward system, which also heavily projects to the dACC (Rushworth et al., 2011; Kolling et al., 2016). As the neural activities encoding the prediction errors as control signals in the midbrain dopaminergic neurons are task-generic (Schultz and Dickinson, 2000), it is speculatively believed that the task identity information might be likewise encoded in the dACC, which has been shown to represent both the task and value relevant information (Matsumoto et al., 2003; Johnston et al., 2007; Hayden and Platt, 2010; Rangel and Hare, 2010; Cai and Padoa-Schioppa, 2012; Hunt et al., 2018). Meanwhile, the control signals of the prediction errors might be used for “gating” cognitive control in the prefrontal regions via the dorsal striatum (Collins and Frank, 2014; Botvinick and Braver, 2015), which has also rich task-specific connectivity with the dACC and the prefrontal regions (Haber and Knutson, 2010). In short, the two types of cognitive control driven by different motivations could share similar neural processes in the dACC, concurrently representing both the identity and intensity information of control signals.
Footnotes
This work was supported by The Key Program for International S&T Cooperation Projects of China, MOST, Grant 2016YFE0129100 (to X.W.), the National Natural Science Foundation of China Grant 31471068 (to X.W.), and partial supported by The Fundamental Research Funds for the Central Universities Grant 2017EYT33 (to X.W.). We thank X. Zhang and Z. Xiao for technical assistance.
The authors declare no competing financial interests.
- Correspondence should be addressed to Xiaohong Wan at xhwan{at}bnu.edu.cn
