Metacognition: ideas and insights from neuro- and educational sciences

Operational definitions

In cognitive neuroscience, research in metacognition is split into two tracks. One track mainly studies meta-knowledge by investigating the neural basis of introspective judgements about one’s own cognition (i.e., metacognitive judgements), and meta-control with experiments involving cognitive offloading. In these experiments, subjects can perform actions such as set reminders, making notes and delegating tasks, or report their desire for them. Some research has investigated how metacognitive judgements can influence subsequent cognitive behaviour (i.e., a downward stream from the meta-level to the object level), but only one study so far has explored how this relationship is mapped in the brain. In the other track, researchers investigate EF, also referred to as cognitive control, which is closely related to metacognition. Note however that EF are often not framed in metacognitive terms in the literature (but see ref.). For the sake of concision, we limit our review to operational definitions that have been used in neuroscientific studies.

Metacognitive judgements

Cognitive neuroscientists have been using paradigms in which subjects make judgements on how confident they are with regards to their learning of some given material. These judgements are commonly referred to as metacognitive judgements, which can be viewed as a form of meta-knowledge (for reviews see Schwartz and Nelson). Historically, researchers mostly resorted to paradigms known as Feelings of Knowing (FOK) and Judgements of Learning (JOL). FOK reflect the belief of a subject to knowing the answer to a question or a problem and being able to recognise it from a list of alternatives, despite being unable to explicitly recall it. Here, metacognitive judgement is thus made after retrieval attempt. In contrast, JOL are prospective judgements during learning of one’s ability to successfully recall an item on subsequent testing.

More recently, cognitive neuroscientists have used paradigms in which subjects make retrospective metacognitive judgements on their performance in a two-alternative Forced Choice task (2-AFC). In 2-AFCs, subjects are asked to choose which of two presented options has the highest criterion value. Different domains can be involved, such as perception (e.g., visual or auditory) and memory. For example, subjects may be instructed to visually discriminate which one of two boxes contains more dots, identify higher contrast Gabor patches, or recognise novel words from words that were previously learned (Fig. 2). The subjects engage in metacognitive judgements by rating how confident they are relative to their decision in the task. Based on their responses, one can evaluate a subject’s metacognitive sensitivity (the ability to discriminate one’s own correct and incorrect judgements), metacognitive bias (the overall level of confidence during a task), and metacognitive efficiency (the level of metacognitive sensitivity when controlling for task performance; Fig. 3). Note that sensitivity and bias are independent aspects of metacognition, meaning that two subjects may display the same levels of metacognitive sensitivity, but one may be biased towards high confidence while the other is biased towards low confidence. Because metacognitive sensitivity is affected by the difficulty of the task (one subject tends to display greater metacognitive sensitivity in easy tasks than difficult ones and different subjects may find a task more or less easy), metacognitive efficiency is an important measure as it allows researchers to compare metacognitive abilities between subjects and between domains. The most commonly used methods to assess metacognitive sensitivity during retrospective judgements are the receiver operating curve (ROC) and meta-d′. Both derive from signal detection theory (SDT) which allows Type 1 sensitivity, or d’′ (how a subject can discriminate between stimulus alternatives, i.e. object-level processes) to be differentiated from metacognitive sensitivity (a judgement on the correctness of this decision). Importantly, only comparing meta-d′ to d′ seems to give reliable assessments metacognitive efficiency. A ratio of 1 between meta-d’′ and d’′, indicates that a subject was perfectly able to discriminate between their correct and incorrect judgements. A ratio of 0.8 suggests that 80% of the task-related sensory evidence was available for the metacognitive judgements. Table 1 provides an overview of the different types of tasks and protocols with regards to the type of metacognitive process they operationalise. These operationalisations of meta-knowledge are used in combination with brain imaging methods (functional and structural magnetic resonance imaging; fMRI; MRI) to identify brain regions associated with metacognitive activity and metacognitive abilities. Alternatively, transcranial magnetic stimulation (TMS) can be used to temporarily deactivate chosen brain regions and test whether this affects metacognitive abilities in given tasks.

A recent meta-analysis analysed 47 neuroimaging studies on metacognition and identified a domain-general network associated with high vs. low confidence ratings in both decision-making tasks (perception 2-AFC) and memory tasks (JOL, FOK). This network includes the medial and lateral prefrontal cortex (mPFC and lPFC, respectively), precuneus and insula. In contrast, the right anterior dorsolateral PFC (dlPFC) was specifically involved in decision-making tasks, and the bilateral parahippocampal cortex was specific to memory tasks. In addition, prospective judgements were associated with the posterior mPFC, left dlPFC and right insula, whereas retrospective judgements were associated with bilateral parahippocampal cortex and left inferior frontal gyrus. Finally, emerging evidence suggests a role of the right rostrolateral PFC (rlPFC), anterior PFC (aPFC), dorsal anterior cingulate cortex (dACC) and precuneus in metacognitive sensitivity (meta-d′, ROC). In addition, several studies suggest that the aPFC relates to metacognition specifically in perception-related 2-AFC tasks, whereas the precuneus is engaged specifically in memory-related 2-AFC tasks. This may suggest that metacognitive processes engage some regions in a domain-specific manner, while other regions are domain-general. For educational scientists, this could mean that some domains of metacognition may be more relevant for learning and, granted sufficient plasticity of the associated brain regions, that targeting them during interventions may show more substantial benefits. Note that rating one’s confidence and metacognitive sensitivity likely involve additional, peripheral cognitive processes instead of purely metacognitive ones. These regions are therefore associated with metacognition but not uniquely per se. Notably, a recent meta-analysis suggests that domain-specific and domain-general signals may rather share common circuitry, but that their neural signature varies depending on the type of task or activity, showing that domain-generality in metacognition is complex and still needs to be better understood.

In terms of the role of metacognitive judgements on future behaviour, one study found that brain patterns associated with the desire for cognitive offloading (i.e., meta-control) partially overlap with those associated with meta-knowledge (metacognitive judgements of confidence), suggesting that meta-control is driven by either non-metacognitive, in addition to metacognitive, processes or by a combination of different domain-specific meta-knowledge processes.

Executive function

In EF, processes such as error detection/monitoring and effort monitoring can be related to meta-knowledge while error correction, inhibitory control, and resource allocation can be related to meta-control. To activate these processes, participants are asked to perform tasks in laboratory settings such as Flanker tasks, Stroop tasks, Demand Selection tasks and Motion Discrimination tasks (Fig. 4). Neural correlates of EF are investigated by having subjects perform such tasks while their brain activity is recorded with fMRI or electroencephalography (EEG). Additionally, patients with brain lesions can be tested against healthy participants to evaluate the functional role of the impaired regions.

In a review article on the neural basis of EF (in which they are defined as meta-control), Shimamura argues that a network of regions composed of the aPFC, ACC, ventrolateral PFC (vlPFC) and dlPFC is involved in the regulations of cognition. These regions are not only interconnected but are also intricately connected to cortical and subcortical regions outside of the PFC. The vlPFC was shown to play an important role in “selecting and maintaining information in working memory”, whereas the dlPFC is involved in “manipulating and updating information in working memory”. The ACC has been proposed to monitor cognitive conflict (e.g. in a Stroop task or a Flanker task), and the dlPFC to regulate it. In particular, activity in the ACC in conflict monitoring (meta-knowledge) seems to contribute to control of cognition (meta-control) in the dlPFC and to “bias behavioural decision-making toward cognitively efficient tasks and strategies” (p. 356). In a recent fMRI study, subjects performed a motion discrimination task (Fig. 4c). After deciding on the direction of the motion, they were presented additional motion (i.e. post-decisional evidence) and then were asked to rate their confidence in their initial choice. The post-decisional evidence was encoded in the activity of the posterior medial frontal cortex (pMFC; meta-knowledge), while lateral aPFC (meta-control) modulated the impact of this evidence on subsequent confidence rating. Finally, results from a meta-analysis study on cognitive control identified functional connectivity between the pMFC, associated with monitoring and informing other regions about the need for regulation, and the lPFC that would effectively regulate cognition.

Online vs. offline metacognition

While the processes engaged during tasks such as those used in EF research can be considered as metacognitive in the sense that they are higher-order functions that monitor and control lower cognitive processes, scientists have argued that they are not functionally equivalent to metacognitive judgements. Indeed, engaging in metacognitive judgements requires subjects to reflect on past or future activities. As such, metacognitive judgements can be considered as offline metacognitive processes. In contrast, high-order processes involved in decision-making tasks such as used in EF research are arguably largely made on the fly, or online, at a rapid pace and subjects do not need to reflect on their actions to perform them. Hence, we propose to explicitly distinguish online and offline processes. Other researchers have shared a similar view and some have proposed models for metacognition that make similar distinctions. The functional difference between online and offline metacognition is supported by some evidence. For instance, event-related brain potential (ERP) studies suggest that error negativities are associated with error detection in general, whereas an increased error positivity specifically encodes error that subjects could report upon. Furthermore, brain-imaging studies suggest that the MFC and ACC are involved in online meta-knowledge, while the aPFC and lPFC seem to be activated when subjects engage in more offline meta-knowledge and meta-control, respectively. An overview of the different tasks can be found in Table 1 and a list of different studies on metacognition can be found in Supplementary Table 1 (organised in terms of the type of processes investigated, the protocols and brain measures used, along with the brain regions identified). Figure 5 illustrates the different brain regions associated with meta-knowledge and meta-control, distinguishing between what we consider to be online and offline processes. This distinction is often not made explicitly but it will be specifically helpful when building bridges between cognitive neuroscience and educational sciences.

Training metacognition

There are extensive accounts in the literature of efforts to improve EF components such as inhibitory control, attention shifting and working memory. While working memory does not directly reflect metacognitive abilities, its training is often hypothesised to improve general cognitive abilities and academic achievement. However, most meta-analyses found that training methods lead only to weak, non-lasting effects on cognitive control. One meta-analysis did find evidence of near-transfer following EF training in children (in particular working memory, inhibitory control and cognitive flexibility), but found no evidence of far-transfer. According to this study, training on one component leads to improved abilities in that same component but not in other EF components. Regarding adults, however, one meta-analysis suggests that EF training in general and working memory training specifically may both lead to significant near- and far-transfer effects. On a neural level, a meta-analysis showed that cognitive training resulted in decreased brain activity in brain regions associated with EF. According to the authors, this indicates that “training interventions reduce demands on externally focused attention” (p. 193).

With regards to meta-knowledge, several studies have reported increased task-related metacognitive abilities after training. For example, researchers found that subjects who received feedback on their metacognitive judgements regarding a perceptual decision-making task displayed better metacognitive accuracy, not only in the trained task but also in an untrained memory task. Related, Baird and colleagues found that a two-week mindfulness meditation training lead to enhanced meta-knowledge in the memory domain, but not the perceptual domain. The authors link these results to evidence of increased grey matter density in the aPFC in meditation practitioners.

Summary

Research on metacognition in cognitive science has mainly been studied through the lens of metacognitive judgements and EF (specifically performance monitoring and cognitive control). Meta-knowledge is commonly activated in subjects by asking them to rate their confidence in having successfully performed a task. A distinction is made between metacognitive sensitivity, metacognitive bias and metacognitive efficacy. Monitoring and regulating processes in EF are mainly operationalised with behavioural tasks such as Flanker tasks, Stroop tasks, Motion Discrimination tasks and Demand Selection tasks. In addition, metacognitive judgements can be viewed as offline processes in that they require the subject to reflect on her cognition and develop meta-representations. In contrast, EF can be considered as mostly online metacognitive processes because monitoring and regulation mostly happen rapidly without the need for reflective thinking.

Although there is some evidence for domain specificity, other studies have suggested that there is a single network of regions involved in all meta-cognitive tasks, but differentially activated in different task contexts. Comparing research on meta-knowledge and meta-control also suggest that some regions play a crucial role in both knowledge and regulation (Fig. 5). We have also identified a specific set of regions that are involved in either offline or online meta-knowledge. The evidence in favour of metacognitive training, while mixed, is interesting. In particular, research on offline meta-knowledge training involving self-reflection and metacognitive accuracy has shown some promising results. The regions that show structural changes after training, were those that we earlier identified as being part of the metacognition network. EF training does seem to show far-transfer effects at least in adults, but the relevance for everyday life activity is still unclear.

One major limitation of current research in metacognition is ecological validity. It is unclear to what extent the operationalisations reviewed above reflect real-life metacognition. For instance, are people who can accurately judge their performance on a behavioural task also able to accurately assess how they performed during an exam? Are people with high levels of error regulation and inhibitory control able to learn more efficiently? Note that criticism on the ecological validity of neurocognitive operationalisations extends beyond metacognition research. A solution for improving validity may be to compare operationalisations of metacognition in cognitive neuroscience with the ones in educational sciences, which have shown clear links with learning in formal education. This also applies to metacognitive training.