Abstract
Healthy individuals readily adjust their behavior in response to errors using learning mechanisms. This raises the question of how error-related neural mechanisms underlie the learning process and its progress. In this study, 21 healthy participants performed a challenging functional magnetic resonance imaging (fMRI) task to answer this question. We assessed the evolution of error-related neural response as a function of learning progress. We tested the hypothesis that the dorsal anterior cingulate cortex (dACC) and anterior insula, key regions of the error monitoring neural circuitry, reflect both the performance of an action and its improvement. Given the nature of trial-and-error learning, we also expected an involvement of the striatum, particularly the putamen. We found that error-related neural activity (in the dACC and anterior insula) was similar following correct responses and errors in an initial learning period. However, as learning progressed, the activity continuously decreased in response to correct events and increased after errors. In opposition, during the initial learning phase, the putamen activity was modulated by errors, but, as it progressed, this region became unaffected by response outcomes. In sum, our study provides neural evidence for an interaction between the mechanisms underlying error monitoring and learning, contributing to clarifying how error-related neural responses evolve with learning.
1 Introduction
Error monitoring enables one to detect, process, and signal errors prior to, during, and following actions (Ullsperger et al., 2014). It facilitates more efficient actions by helping minimize subsequent mistakes, which makes it crucial for adaptive behavior and learning (Ninomiya et al., 2018; Ullsperger, 2017; Ullsperger et al., 2014).
Electroencephalographic (EEG) research has identified the error-related negativity (ERN), a midfrontal negative event-related potential (ERP) estimated to originate in the dorsal anterior cingulate cortex (dACC) (Neta et al., 2015; Ullsperger et al., 2014) that arises 50-100 ms after erroneous actions (Bhattacharyya et al., 2017; Chavarriaga et al., 2014; Estiveira et al., 2022). According to the reinforcement learning theory of the ERN (Holroyd & Coles, 2002), this ERP reflects reward prediction errors, which are dopaminergic signals that encode discrepancies between expected and actual outcomes to trigger appropriate adjustments and optimize behavior (Alexander & Brown, 2011; Ullsperger et al., 2014). In line with this, functional magnetic resonance imaging (fMRI) studies link the dACC to an extensive range of functions, including salience, reward-based decision making, conflict, and error monitoring. The dACC has also been hypothesized to monitor external and internal processes which contribute to making predictions about action outcomes and providing action feedback to downstream circuits. Such activity is employed to update predictions and optimize future behavior. Therefore, it is believed that error and conflict are tracked by the dACC and signaled when additional control and cognitive resources are needed (Heilbronner & Hayden, 2016; Weiss et al., 2018).
Even though most error monitoring studies have focused on the activity of the dACC, several other regions have been linked to error monitoring processes, such as the anterior insula, pre-supplementary area (pre-SMA), basal ganglia, and lateral prefrontal cortex (Neta et al., 2015; Ninomiya et al., 2018; Ullsperger et al., 2014). The anterior insula, in particular, has been consistently shown to have an important role in error awareness, as previous studies reported increased activity in perceived compared to unperceived errors (Dali et al., 2023; Harsay et al., 2012; Klein et al., 2007). It has been suggested that the dACC contributes to error awareness as well (Dali et al., 2023; Orr & Hester, 2012). In accordance with this, both regions have been proposed to be involved in interoceptive awareness, and, more precisely, to integrate autonomic information with salient events such as errors (Dali et al., 2023; Estiveira et al., 2022; Harsay et al., 2012; Klein et al., 2007, 2013).
The anterior insula and dACC are thus frequently co-activated for a variety of cognitive processes. These structures have been suggested to form the salience network, which is vital for monitoring important stimuli that require autonomic regulation (Dali et al., 2023; Harsay et al., 2012; Klein et al., 2013). Once sensory areas recognize a salient event, this information is transmitted to the salience network. This network, in turn, triggers a signal to engage brain regions that mediate attentional, working memory, and action selection processes, while disengaging the default mode network. Errors can be seen as salient events due to their occasional occurrence and usefulness in guiding behavior and learning (Harsay et al., 2012).
Although the neural processes underlying error monitoring have been extensively studied, their relationship with learning mechanisms remains unclear. Kennerley et al. (2006) demonstrated that the ACC is vital for learning: the authors found that lesions in the ACC lead to impairments in integrating past feedback information, causing monkeys to use only the immediate previous feedback as a guide for subsequent choices. Additionally, previous studies have found that error-related dACC activity predicts post-error adaptations (Danielmeier et al., 2011) and error correction (Hester et al., 2008, 2009). Moreover, Mars et al. (2005) suggested that dACC activity changes during learning of stimulus-response associations by trial and error. In their study, before learning, error information was not available until the delivery of external performance feedback, but after learning, it was available earlier from internal sources. Accordingly, the authors reported that error feedback-related dACC activity decreases as learning proceeds, while error response-related activity increases. Their results revealed a neural response shift as a function of learning, from external sources provided by error feedback to internal sources linked to the error response itself. Similarly, using EEG, Bellebaum and Daum (2008) found that feedback-related responses decrease with learning, while internal error-related responses increase.
Most of these studies used tasks with external performance feedback to analyze the neural bases of learning from errors. Nonetheless, most decisions in daily life are not followed by feedback. Reward prediction errors are conveyed to the striatum (Ullsperger et al., 2014), and Daniel and Pollmann (2012) demonstrated that this dopaminergic region is activated not only by explicit external rewards, such as food and money, but also by internally generated signals on perceived correctness (Daniel & Pollmann, 2012, 2014).
In this study, we aimed to clarify how error-related neural mechanisms progress with learning, and how learning is affected by error monitoring, during a challenging functional magnetic resonance imaging (fMRI) task without external feedback. We hypothesized that the dACC, anterior insula, and striatal regions such as the putamen would play a crucial role in learning from errors. We anticipated that the activity in these neural regions would be modulated by the relationship between error monitoring and learning. Both error-related brain activity and connectivity were investigated at different learning periods.
2 Methods
2.1 Participants
Twenty-one healthy volunteers (11 females, mean age 27.38 ± 5.43 years) participated in this study. Participants with neurological or psychiatric illnesses were excluded, and all participants had normal or corrected-to-normal visual acuity. All volunteers were right-handed, and the mean laterality index was 82.61 ± 20.35 (Oldfield, 1971). Every participant completed the magnetic resonance imaging (MRI) safety questionnaires and provided written informed consent in accordance with the Declaration of Helsinki prior to participation. The study followed the safety guidelines for research on humans and was approved by the Ethics Committee of the Faculty of Medicine of the University of Coimbra.
2.2 Task
Our stimulation sequence was based on three types of visual cues: a facial expression, a frame shape, and a frame color cue (Fig. 1). The facial expression cues were planned to raise the participants’ simultaneous attention to the eyes and mouth (happy and sad faces looking to the right or the left) to achieve high performance on the proposed task. The combination of emotional expressions and gaze directions resulted in four different facial cues (Fig. 1A): happy left (1), happy right (2), sad left (3), and sad right (4). The information from both the facial expression and frame colour cues (green or red, Fig. 1B) was needed to infer the direction of action required. If the frame was green, a happy face indicated to act in the same direction of the face’s eye gaze, and a sad face indicated to act in the opposite direction of the face’s eye gaze; if the frame was red, the instructions were inverted (i.e., a happy face instructed to act in the opposite direction of the face’s eye gaze). The frame shape (Fig. 1B)—a rectangle or a diamond—instructed participants to either perform a button press or a saccade, respectively. A correct response required the correct interpretation of the three cues (facial expression, frame shape, and frame color). Participants were informed about the task rules through verbal explanation before entering the MRI scanner.
The experimental paradigm followed a fast event-related design and included five periods (Fig. 2). Each sequence contained an instruction and action period, and three gaps. Gap periods consisted of a black background with a white fixation cross in the middle of the screen. The first one occurred before the instruction and lasted for 500 ms. Then, the instruction was presented during 500 ms. The instruction order was pseudo-random mixing saccade with button press cues and happy with sad faces. The frame color of the facial cue changed after every three trials. The instruction presentation period was followed by a gap of variable duration (1000 ms, 2000 ms, or 3000 ms). Participants should then give their response after this when the fixation cross disappeared. They had 1000 ms to perform their button press or saccade response. Lastly, there was another gap period varying from 1000 ms to 3000 ms. We did not provide feedback (see rationale in the Introduction). This paradigm was adapted from Estiveira et al. (2022), as it has been previously shown to be an effective approach to studying error-monitoring neural circuitry.
Each participant performed 7 experimental runs of 48 trials each. Button press and saccadic trials, happy and sad, and green and red frame cues were counterbalanced. Each run included 3 baseline periods that occurred at the end of every 16 trials (white noise with a duration of 10 s, 11 s, or 12 s) and lasted approximately 6 min.
The stimuli were programmed using the Psychophysics Toolbox 3.0 for MATLAB (R2019b) and presented on a 1920 × 1080 resolution monitor with a refresh rate of 60 Hz, 156 cm away from participants. Facial expression images were sized as 6° × 5.20°, while their frames were 10° × 8.6° (height x width, visual angles from the centre to periphery). The fixation cross had a length of 2.57° × 2.57°. The facial expression images, frames, and fixation cross were displayed at the center of the screen. A gray grid designed to help homogenize saccade behavior included intersections at a horizontal visual angle of 19.13° from the center of the screen (participants were instructed to look at the intersection if they aimed to perform a saccade). The facial expression images were based on a young adult male and obtained from the Radboud Faces Database (Langner et al., 2010).
2.3 Data acquisition
MRI was performed using a 3T Siemens MAGNETOM Prisma Fit scanner (Siemens, Erlangen) using a 64-channel head coil. Functional images were acquired using a 2D multi-band (MB) gradient-echo echoplanar imaging (GE-EPI) sequence, with the following parameters: TR/TE = 1000/30.2 ms, voxel size = 2.5 × 2.5 × 2.5 mm3, 42 axial slices (whole-brain coverage), FOV = 192 × 192 mm2, FA = 90°, phase encoding in the anterior-posterior direction, and MB factor = 3. A 3D whole-brain anatomical T1-weighted MPRAGE (TR = 2530 ms, TE = 3.50 ms, 192 interleaved slices with isotropic voxel size of 1 mm3) was collected for subsequent image registration. Respiratory and cardiac data were recorded at 125 Hz and 500 Hz, respectively, with a respiratory belt and pulse oximeter from the physiological monitoring unit of the MRI system.
Eye-tracking data were recorded using EyeLink 1000 (SR Research, Canada) at a sampling rate of 1000 Hz and accuracy of 0.25°–0.5°. Each run started with an eye-tracker 9-point calibration and validation session. Button press information (response and response latency) was registered using MATLAB version R2018b (*.mat).
2.4 Behavioral data analysis: participants’ performance
To assess participants’ performance, we analyzed whether their responses were correct or not. A correct response required an accurate integration of all cues (facial expression, frame color, and frame shape) at the correct timing, that is, during the Action period (Fig. 2). Errors included incorrect action direction (e.g., pressing the right button instead of the left button), incorrect action (e.g., performing a saccade instead of a button press), omission errors (i.e., not responding), action direction indecisions (e.g., looking at the right and then at the left), action indecisions (e.g., looking at the right and then pressing the right button), simultaneous incorrect action and incorrect action direction, and anticipated errors (including all error types mentioned previously). Anticipated correct responses (i.e., correct responses given before the Action period), which corresponded to 13.87 ± 17.12% of all trials, were excluded from further analysis to avoid ambiguity.
Saccades were defined as horizontal movements superior to 4.27° of visual angle (Poletti & Rucci, 2016) from the screen center. In addition, a minimum duration of 15 ms was considered to identify a saccade. We visually inspected some trials of each participant and adapted these parameters for participants with unusual patterns of ocular movements. The detection of saccadic responses was made in a custom-made MATLAB script (version 2018b, MathWorks).
Participants’ error rate (i.e., the ratio between the erroneous responses and the sum of correct and erroneous responses) was evaluated according to distinct types of instruction with distinct difficulty levels (happy green, happy red, sad green, and sad red) and learning periods (run 1 to 7). We considered each run a distinct learning period since repeated practice leads to learning (Toppino & Gerbier, 2014).
Additionally, we analyzed how different cues reflected distinct task difficulty levels using a post-fMRI session questionnaire. We asked participants which color cue (green or red) and facial expression cue (happy or sad) were the most difficult (with two-choice questions). We measured the percentage of participants that selected each condition.
2.4.1 Behavioral data: statistical analysis
To assess the impact of difficulty and learning period on participants’ error rate, we conducted a linear mixed-effects model with instruction and run/learning period as factors, and error rate as the dependent variable. We also included the interaction between instruction and run, but removed fixed factors and interactions without significance to simplify the model (Seltman, 2018). The participant ID was defined as random variable, and we included a random intercept for each participant. Results were considered significant for p < 0.05. The linear mixed-effects modeling was performed using IBM SPSS Statistics 25 software.
2.5 fMRI data processing
MRI data were preprocessed using custom scripts in MATLAB (version R2019b, MathWorks), based on the SPM12 software with PhysIO toolbox (Kasper et al., 2017), and FMRIB Software Library (FSL), adapted from Soares et al. (2022). The pipeline included: (1) slice timing correction; (2) realignment of all fMRI volumes relative to the first volume; (3) correction of geometric distortions caused by magnetic field inhomogeneity, with FSL tool TOPUP (Andersson et al., 2003); (4) bias field correction; (5) image registration (functional to structural); (6) estimation of nuisance regressors (with PhysIO toolbox) such as cardiac and respiratory signals—RETROICOR regressors (Glover et al., 2000), heart rate variability convolved with the cardiac response function, and respiratory volume per time convolved with the respiratory response function—and head motion (6 motion parameters and framewise displacement censoring with threshold = 0.5 mm); (7) segmentation of the T1 structural image (for the CONN standard denoising pipeline); (8) functional and anatomical images normalization to Montreal Neurological Institute (MNI) space; and (9) spatial smoothing with a 7.5 mm full-width-at-half-maximum (FWHM) isotropic Gaussian kernel - 3 times the voxel size (Chen & Calhoun, 2018).
2.6 fMRI data analysis
2.6.1 Brain activity analysis
To map the regions involved in error monitoring and learning, we built a general linear model (GLM) comprising four regressors for each run: instruction, correct response, error, and other (i.e., excluded responses). The saccade event was set at the end of the horizontal eye movement, as error-related signals are temporally more similar between keypresses and saccades when the saccade event is defined at the end of the eye movement, probably due to greater impulsiveness during saccadic responses (Estiveira et al., 2022). The button press event was set in the moment during which the button was pressed. The onset of omission errors was defined based on the mean onset value for the expected response action (saccade or button press) for each participant’s run. The durations were set to zero, except for the instruction condition, in which it was set to 500 ms. The design matrix for the GLM included a high-pass filter with a cutoff period of 128 s, and regressors were convolved by the canonical hemodynamic response function. Cardiac, respiratory, and head motion regressors were added to the design matrix as confounding signals.
Group activation maps for the contrast between correct events and errors (performance-based modulation: error > correct, computed by errorR1 + errorR2 + errorR3 + errorR4 + errorR5 + errorR6 + errorR7 > correctR1 + correctR2 + correctR3 + correctR4 + correctR5 + correctR6 + correctR7) and the contrast between the first two runs’ responses and the last two runs’ responses (learning-based modulation: late learning period > initial learning period, computed by correctR6 + errorR6 + correctR7 + errorR7 > correctR1 + errorR1 + correctR2 + errorR2) were then generated. SPM12 was used to build the GLM and generate the group activation maps. For visualization purposes, we used BrainNet Viewer (Xia et al., 2013) and MRIcroGL (https://www.nitrc.org/projects/mricrogl).
Moreover, we investigated how the learning period (measured by run number), performance (coded as correct and erroneous responses), and difficulty (assessed by instruction type) interact at the neural response level. With that in mind, we extracted the percent of signal change per event at the level of our regions of interest (ROIs), the dACC, anterior insula, and putamen. The single-response percent signal change was estimated by dividing the single-response beta coefficient by the run model intercept (β0) (Mayer et al., 2012). The single-response beta coefficient was computed based on a Least-Squares All approach (Abdulrahman & Henson, 2016) using SPM 12. Following this approach, each trial response (correct or erroneous) was modeled as a separate regressor in the GLM. However, for instruction and other (i.e., excluded responses) conditions, all repetitions of the same type were collapsed into the same regressor, given that, for the statistical analysis, we were only interested in the percent signal change of correct response and error conditions.
We used the Brainnetome Atlas (Fan et al., 2016) to functionally define our ROIs, based on which we then extracted brain response values to include in the linear mixed-effects models analysis. Regarding the error monitoring analysis, we hypothesized that the dACC and anterior insula would reveal an increased activation during errors (compared to correct responses). Note that, in previous studies, the dACC has also been termed anterior midcingulate cortex (aMCC) (Ninomiya et al., 2018; Ullsperger, 2017; Ullsperger et al., 2014; Vogt, 2016), and the definition of this region has been variable (Vogt, 2016). Hereinafter, the dACC is used to refer to the pregenual area 32 (Beckmann et al., 2009; Fan et al., 2016; Palomero-Gallagher et al., 2008). Accordingly, we selected the Brainnetome regions corresponding to the dACC and bilateral anterior insula, which were the pregenual area 32 (A32p) and the conjunction of the ventral and dorsal agranular insula (vIA and dIA). Moreover, we also included the bilateral putamen in our analysis to explore learning-related brain response modulation, as suggested by previous literature (Ashby et al., 2010; Baladron & Hamker, 2020; Brovelli et al., 2011; Patterson & Knowlton, 2018; Seger & Spiering, 2011; Tricomi et al., 2009). This ROI was defined as the ventromedial putamen (vmPu) Brainnetome region of the atlas.
2.6.2 Functional connectivity analysis
To better understand the brain networks involved in error monitoring and learning, we performed a generalized psychophysiological interaction (gPPI) connectivity analysis (McLaren et al., 2012) using bivariate regression in the CONN toolbox (Whitfield-Gabrieli & Nieto-Castanon, 2012). Preprocessed (but unfiltered—since the CONN standard denoising pipeline includes filtering) data were imported into the toolbox, as well as the onsets and durations of the conditions (baseline, instruction, correct response, error, and other - excluded responses). We then implemented CONN standard denoising pipeline (Nieto-Castanon, 2020). It combines two steps: linear regression to remove confounding effects in the blood oxygen level-dependent (BOLD) signal and temporal filtering. Potential confounding effects include noise components from white matter and cerebrospinal areas, estimated subject-motion parameters, identified outlier scans, constant and first-order linear BOLD trends, and task effects. Moreover, data were high-pass filtered with a cut-off frequency of 0.01 Hz.
To assess the neural networks implicated in error monitoring and learning, we conducted seed-to-voxel analyses for the contrasts error > correct and late learning period > initial learning period. The seeds were our ROIs: dACC and bilateral anterior insula for the first contrast, and bilateral putamen for the second.
2.6.3 fMRI data: statistical analysis
For the group activation maps concerning the contrast between correct and erroneous responses, as well as between initial and late learning periods, the significance threshold was set at p < 0.05 with family-wise error (FWE) correction for multiple comparisons. Moreover, to measure the impact of learning, performance, and difficulty on the activity of our ROIs (the dACC, anterior insula, and putamen), we conducted a linear mixed-effects model with run/learning, performance, and instruction/difficulty as factors and single-response percent of signal change of each ROI as the dependent variable. We also included all possible interactions between factors but removed fixed factors and interactions without significance to simplify the models (Seltman, 2018). The participant ID was defined as random variable, and we included a random intercept for each participant. Using linear mixed-effects modeling, we accounted for within and across subjects’ variability and included multiple factors in the analysis while considering an unequal number of repetitions (Walker et al., 2019). The linear mixed-effects modeling was performed using IBM SPSS Statistics 25 software. Due to multiple testing (one model for each ROI), we used Bonferroni correction for multiple comparisons by multiplying the p-value by the number of ROIs. For the seed-to-voxel analyses, the connection threshold was set at p < 0.001 (uncorrected), and the cluster threshold was set at p < 0.01 with FWE correction for multiple comparisons.
3 Results
3.1 Behavioral results: participants’ performance
In total, we identified 4604 correct responses and 1210 errors (mean 219.24 ± 90.27 correct responses and 57.62 ± 35.46 errors per participant). A linear mixed-effects model analysis revealed a significant impact of run (F(6, 562.97) = 49.83, p = 4.15 × 10-49), which is consistent with the learning curve shown in Figure 3A, and instruction (F(3, 563.02) = 9.02, p = 8.00 × 10-6) on error rate (Fig. 3B), and no significant interaction between both. To compare the error rate between all learning periods and instructions, we ran post-hoc pairwise comparisons. The error rate in the first two runs was higher than the error rate in the remaining runs, and the error rate in the third run was higher than the error rate in the last three runs. Moreover, fewer errors occurred following the happy green instruction compared to the sad green and sad red instructions, and the number of errors following the happy red instruction was lower than the number of errors that followed the sad red instruction. Supplementary Table S1 details these results.
Regarding the participants’ reports on task difficulty levels, 84.21% considered the sad cue more difficult than the happy cue, and 89.47% rated the red cue as more demanding than the green cue.
3.2 Neurophysiological results
3.2.1 Brain activity modulation by error monitoring and learning
We started by mapping the regions involved in error monitoring. A group statistical map for the contrast between correct responses and errors (error > correct) was generated (Fig. 4). Clusters identified with significant brain activity for this contrast (detailed in Supplementary Table S2) included the dACC, anterior insula, and other regions also known to be involved in error monitoring mechanisms, namely the pre-supplementary motor area (pre-SMA), supplementary motor area (SMA), paracingulate gyrus, inferior frontal gyrus (IFG), and orbital gyrus (Cieslik et al., 2024; Neta et al., 2015; Ninomiya et al., 2018; Ullsperger, 2017; Ullsperger et al., 2014). The statistical map for the contrast correct > error is shown in Supplementary Figure S1 (with details provided in Supplementary Table S2).
Along with error monitoring, we were interested in mapping the regions involved in learning. Therefore, we generated group activation maps concerning the contrast late learning period > initial learning period (Fig. 5). Only the ventral putamen region revealed a significant response for this contrast (details provided in Supplementary Table S3). We did not find significant clusters for the contrast initial learning period > late learning period.
3.2.2 Relationship between error monitoring and learning
In addition to the whole-brain contrast statistical maps, we tested the impact of performance, learning period, and difficulty level on the responses of our regions of interest, specified in our hypothesis: dACC and anterior insula, and putamen. As mentioned in section 2.6.1, we used the Brainnetome Atlas (Fan et al., 2016) to functionally define our ROIs. Supplementary Figure S2 illustrates the overlay between the clusters that resulted from the previous activation maps (for the contrasts error > correct and late learning period > initial learning period) and the Brainnetome Atlas-based ROIs.
The dACC and anterior insula responses showed a similar pattern. A linear mixed-effects modeling analysis revealed a significant effect of performance on the activity of dACC (F(1, 5581.62) = 85.94, p-corr. = 7.77 × 10-20) and anterior insula (F(1, 5149.33) = 141.13, p-corr. = 1.19 × 10-31). It also revealed a significant interaction between learning period (run number) and performance on dACC (F(12, 5769.74) = 4.38, p-corr. = 1.58 × 10-6) and anterior insula (F(12, 5773.50) = 3.65, p-corr. = 5.10 × 10-5). No significant effect was found related to task difficulty (measured by instruction type).
Given that the mixed-effects analyses revealed significant interactions between learning period and performance, we ran post-hoc tests to compare the BOLD activity in response to correct trials and errors per experimental run/learning period (Fig. 6). In the first run/learning period, there was no difference between correct and erroneous responses for the regions of the salience network (dACC and anterior insula), but it increased with the progress of learning: the activity of dACC and anterior insula decreased for correct responses and increased for errors along the runs. The statistical results are detailed in the Supplementary Table S4.
We also found a significant impact of learning period (F(6, 5766.21) = 38.04, p-corr. = 4.62 × 10-45) and performance (F(1, 5722.17) = 10.87, p-corr. = 0.003) on the activity of bilateral putamen, as well as a significant interaction between learning period and performance (F(6, 5769.24) = 2.87, p-corr. = 0.024). Given the significant interaction, we ran post-hoc tests to compare the BOLD activity in response to correct answers and errors for each run/learning period (Fig. 7). Putamen activity differed between correct and erroneous responses in the first run, but it became equivalent between conditions with the effect of learning, featuring an opposite effect to the one seen in the ACC and anterior insula. The putamen response increased in both cases over time. The statistical results are detailed in the Supplementary Table S5.
3.2.3 Neural correlates of error monitoring at the functional connectivity level
To evaluate the neural networks implicated in error monitoring and learning, we conducted seed-to-voxel analyses for the contrasts error > correct (using the dACC and anterior insula as seed regions) and late learning period > initial learning period (with the putamen as seed region). Figure 8 illustrates the results from the seed-to-voxel analysis concerning the first contrast, using the dACC as seed region, and all the statistical details are provided in Supplementary Table S6. We did not find significant results for the remaining analyses.
4 Discussion
In this study, we investigated how neural processes of error monitoring interact with learning when the integration of different facial cues is required. We found that activity in the dACC and anterior insula was not modulated by participants’ performance in the initial learning period, but increasingly differed between correct and erroneous responses with learning progress. The results showed an opposite pattern for the putamen: its activity was initially modulated by errors, but then similarly increased over time for correct and erroneous responses.
4.1 Behavioral results: a clear learning curve and distinct difficulty levels
The performed task led to a well-defined learning curve: in the first run, participants had a mean error rate of above 50%, and in the last, of less than 15%. Additionally, the different types of conditions led to distinct error rates, confirming the existence of different difficulty levels, as also perceived and reported by the participants. There are few error monitoring studies with clear learning curves (Daniel & Pollmann, 2012; Maurer et al., 2019; McDougle, 2022), which reveals that this task is appropriate to study the relationship between performance monitoring and learning while accounting for potential difficulty effects. It was relevant to measure the influence of difficulty on our results, and to disentangle the effects of error monitoring and task difficulty, since previous literature proposed that the dACC is modulated by choice difficulty (Shenhav et al., 2014).
4.2 Differences between correct responses and errors in key performance monitoring regions increase with learning
As hypothesized, the dACC and anterior insula showed increased BOLD signal during erroneous than correct responses (although only after an initial learning period). These two regions have been described as key regions of the error monitoring neural circuitry and linked to error awareness (Dali et al., 2023; Orr & Hester, 2012). The dACC has been hypothesized to signal the need for additional cognitive control resources following errors (Heilbronner & Hayden, 2016; Weiss et al., 2018).
When assessing how the dACC response to error changed with the learning progress, we found that, at the beginning of the task, the response to correct and erroneous events was similar, and, over time, the differences between these conditions became more evident. Along the learning periods, for correct responses, the dACC activity decreased; for errors, it increased. Kelly and Garavan (2005) have argued that decreases in the extent or intensity of activations are expected with task practice due to increased neural efficiency. Activation in attentional and control regions, for instance, may return to baseline on account of practice. Matsumoto et al. (2007), using intracranial recordings in monkeys, demonstrated that a subgroup of ACC neurons responds to positive visual feedback, and the magnitude of these responses decreases throughout adaptation. The authors reported that bigger responses were evoked by positive feedback when the monkey was not confident with its action, that is, when the feedback was not predicted (and thus useful for learning). Here, on one hand, the dACC activity in correct trials decreased with practice, which may be related to increased neural efficiency and/or task confidence, due to learning. On the other hand, the dACC increased with time in response to errors, which may be related to the fact that, as learning proceeds, erroneous responses become more unexpected. In the first run, errors are not surprising since participants are still learning the task, but with practice erroneous responses become uncommon.
These results are compatible with the predicted response outcome (PRO) model (Alexander & Brown, 2011), according to which the ACC and surrounding medial prefrontal cortex predict the likelihood of future outcomes and compare it with the actual outcome. When the expected and actual outcomes do not match, a reward prediction error is produced and communicated to the prediction units. The predictions are then updated for future reference. Therefore, according to this theory, the ACC is primarily sensitive to the unpredictability of outcomes. The reinforcement learning model (Holroyd & Coles, 2002, 2008) may also explain the increase in dACC activity during erroneous responses with learning and practice, as it similarly suggests that the dACC activity reflects reward prediction errors. This theory proposes that the dACC selects the action plan for a particular task based on reward prediction error signals carried to the dACC by the midbrain dopaminergic system (with massive projections to the striatum) (Daniel & Pollmann, 2014).
The response to error in the anterior insula had similar progress as the dACC, in line with previous studies reporting co-activation of these neural regions as part of the salience network (Harsay et al., 2012; Klein et al., 2007; Ullsperger et al., 2010). The anterior insula has mainly been linked to error awareness (Dali et al., 2023; Harsay et al., 2012; Klein et al., 2013), suggesting that participants became more aware of their mistakes with learning. Moreover, the anterior insula has been previously linked to learning (Horing & Büchel, 2022; Kornhuber et al., 1995; Palminteri et al., 2012). For instance, Palminteri et al. (2012) demonstrated that damage to the anterior insula impairs punishment avoidance and hence affects the learning process.
Additionally, we evaluated the impact of task difficulty level (measured by the different instruction types) on the dACC and anterior insula activity during error monitoring progress with learning but did not find significant results. This finding is at odds with theories of dACC function that suggest that the dACC codes for choice difficulty (Shenhav et al., 2014). The dACC functions are still a matter of debate, and several models have been proposed and discussed in the literature (Vassena et al., 2017). Some of them accommodate our findings—for instance, the PRO model (Alexander & Brown, 2011) and the reinforcement learning model (Holroyd & Coles, 2002, 2008) —but not the one proposed by Shenhav et al. (2014).
4.3 The role of putamen in learning from errors
The putamen revealed to have increasing BOLD activity over time during our task. Its response progressed differently from the dACC and anterior insula: it differed between correct and erroneous responses in the beginning and then similarly increased for both correct and erroneous responses with learning.
According to the reinforcement learning theory, the reward prediction error is communicated to the ventral striatum (including the ventral putamen), where it is used to improve outcome predictions and compare them to actual outcomes (Daniel & Pollmann, 2014; Ullsperger et al., 2014). This may happen during the initial learning periods, leading to differences in putamen activity between correct and erroneous responses, as found in our results. With learning, this differentiation is taken over by the dACC and anterior insula for performance monitoring. Similarly, Daniel et al. (2012) demonstrated increased activation for correct than erroneous responses only before extensive training. Afterward, the anterior insula, medial prefrontal areas, and midbrain revealed higher activity for errors than correct answers.
The putamen has been linked to habit learning and automaticity (Ashby et al., 2010; Baladron & Hamker, 2020; Brovelli et al., 2011; Miyachi et al., 1997, 2002; Patterson & Knowlton, 2018; Reading et al., 1991; Seger & Spiering, 2011; Tricomi et al., 2009; Yin et al., 2006, 2004), which is consistent with the increase in putamen response with learning progress found in our results. While habit formation is sensitive to changes in outcome value, actions become more automatic with practice and, if a stimulus-response habit has been formed, a cue might evoke its associated response despite the outcome (Dickinson & Balleine, 2002; Tricomi et al., 2009). This may explain the difference in putamen activity between correct and erroneous responses found at the beginning of our session only.
4.4 Other regions modulated by errors
In addition to these regions, the paracingulate gyrus, pre-SMA, SMA, IFG, and orbital gyrus revealed an increased activity during errors compared to correct responses. Previous studies have shown that adjacent areas to the dACC, such as the paracingulate gyrus, pre-SMA, and SMA, are also modulated during response errors, conflict, negative feedback, and surprise (Neta et al., 2015; Ullsperger, 2017). The pre-SMA is one of the most reported regions in this regard (Fu et al, 2023; Ninomiya et al., 2018; Ullsperger, 2017; Ullsperger et al., 2014). Some studies suggested that the pre-SMA activity is mainly linked to conflict monitoring and showed a spatial dissociation between conflict and error monitoring, located in the pre-SMA and dACC, respectively (Garavan et al., 2003; Ullsperger & Von Cramon, 2001). Others suggested that the neural activation in response to errors located in the dACC extends to the pre-SMA (Iannaccone et al., 2015; Jahn et al., 2016). Fu et al. (2023) have described the existence of two frontal regions with error monitoring-related activity: the supplementary motor complex, within the superior frontal gyrus (SFG) and including the SMA, pre-SMA, and supplementary eye field; and the ACC. Moreover, the pre-SMA and IFG, which seem involved in response inhibition, have also been related to post-error behavior adjustments, namely to post-error slowing (Ullsperger, 2017). The orbitofrontal cortex, in turn, appears to encode outcome valence, outcome expectancies, and information about rewards and choices (Ullsperger et al., 2014).
4.5 How do error-related and learning-related brain regions communicate?
To clarify how error-related and learning-related neural regions interact during errors, we conducted seed-to-voxel analyses that revealed augmented functional connectivity during errors (compared to correct responses) between the dACC and the IFG, supramarginal gyrus, and middle temporal gyrus. These regions have been previously linked to error monitoring (Barke et al., 2017; Cieslik et al., 2024; Dali et al., 2023; Gauvin et al., 2016; Neta et al., 2015; Ullsperger et al., 2014), and Du et al. (2020) have also shown that the dACC has strong functional connectivity with the IFG and supramarginal gyrus.
The IFG has been linked to inhibitory control (Aron et al., 2004, 2007; Ullsperger et al., 2014), and previous literature has suggested that the middle temporal gyrus is associated with the processing of awareness of discrepancies (Van Kemenade et al., 2019). Besides, the supramarginal gyrus has been proposed to integrate a system that detects and direct attention to relevant stimuli during periods of internally focused mental activity (Corbetta & Shulman, 2002; Reniers et al., 2012). We suggest that, when an error is committed, the middle temporal gyrus detects the discrepancy between the expected and obtained outcome, the supramarginal gyrus directs attention to the erroneous action, and the dACC signals the error to other regions to trigger adjustments and optimize behavior, namely to the IFG, that inhibits future mistakes.
Apart from the IFG, we did not find these regions (the middle temporal gyrus and supramarginal gyrus) in the group map for the contrast between correct responses and errors, due to the differences between the GLM and gPPI analyses (Satake et al., 2024). While the GLM analysis detects the BOLD response related to neural activation during the task compared with the resting period, the gPPI functional connectivity analysis evaluates the synchronization of neuronal fluctuations influenced by the task. Gerchen and Kirsch (2017) demonstrated that task-evoked activation and connectivity effects reflect separable and complementary information.
Interestingly, only the dACC showed regions of significantly higher connectivity following errors compared to correct responses, suggesting a functional differentiation between the dACC and anterior insula. This distinction had been previously proposed by Harsay et al. (2012), who argued that the anterior insula is more specialized in processing multimodal sensory input, whereas the dACC, which is connected to action selection and execution systems, triggers adaptive actions in response to this input (Estiveira et al., 2022).
4.6 Limitations and future work
The main limitation of our study is the sample size (n = 21); therefore, future research should replicate this work. Moreover, the clusters we obtained from the statistical maps concerning the contrasts between, on the one hand, correct and erroneous responses and, on the other, initial and late learning periods, were smaller than the ROIs defined based on the Brainnetome atlas. In the future, we aim to study the functional parcellation of our ROIs during error monitoring.
Given that detecting, processing, and signaling errors is vital for regulating behavior, preventing future mistakes, and learning, less efficient error monitoring mechanisms may lead to difficulties in everyday life (Goldberg et al., 2011; Hüpen et al., 2016). Therefore, it would be relevant to study the evolution of error-related neural processes throughout learning in populations with altered error monitoring mechanisms, such as autism, attention-deficit/hyperactivity disorder, and obsessive-compulsive disorder (Bellato et al., 2021).
5 Conclusions
In this study, we aimed to understand how neural processes of error monitoring relate to the learning progress. Our findings provide evidence for the interplay between error monitoring and learning neural processes. The activity in the dACC, anterior insula, and putamen was found to be modulated by the interaction between performance monitoring and learning.
The modulation of dACC and anterior insula response by errors had a similar evolution with learning: in the beginning, the neural responses for correct events and error did not differ but, with learning, the differences became evident, with a simultaneous decrease in correct responses and increase following errors. This seems to be related to prediction error progress with learning. Importantly, our results also revealed two learning phases that modulate the putamen activity. During the initial phase—habit formation through trial-and-error (reinforcement) learning—, the putamen activity following correct and erroneous responses was distinct. Its activation increased with practice-based learning, and, during the second phase—habit consolidation—, actions became more automatic with practice (and insensitive to changes in outcome value).
We found a temporal distinction in error-related neural responses between the putamen (which showed an explicit relation with the learning curve) and the dACC/anterior insula. Correct and erroneous responses were differentiated in the putamen in the initial learning period, and later in the dACC and anterior insula, suggesting a chronometric relationship between circuits underlying learning and error monitoring in the basal ganglia and salience network.
Data and Code Availability
The dataset used here will be made available once the project, within which this study is integrated, is completed. The group activation and connectivity maps are available at https://doi.org/10.5281/zenodo.13742605, and all the code used for pre-processing and analysis is available at https://github.com/CIBIT-UC/errormonitoring_learning.git.
Author Contributions
All authors conceived and designed the experiments. C.D. and T.S. performed the experiments and acquired the fMRI data. All authors contributed to analysis and interpretation of the data. C.D. wrote the first draft of the manuscript and prepared figures and tables. All authors reviewed drafts of the paper and approved the final manuscript.
Funding
This research work was funded by the FCT Portuguese National Funding Agency for Science, Research and Technology [grants UI/BD/150832/2021, UIDP&B/4950/2020, PTDC/PSI-GER/30852/2017, PTDC/EEI-AUT/30935/2017, CEEC: 2021.01469.CEECIND], and BIAL Foundation [project PT/FB/BL-2018-306].
Declaration of Competing Interest
None.
Acknowledgments
We are very grateful to the participants who enrolled in this study. We also thank João Estiveira for his help in designing the experiment, Sónia Afonso and Tânia Lopes for their help with the MRI setup and scanning, and Júlia Soares for her help in the fMRI data processing.
Supplementary Materials
Supplementary material for this article is available with the online version here: https://doi.org/10.1162/imag_a_00343.