Brain mechanisms of metacognition

Listen to audio

The existence of animal models of metacognition opens the door to an examination of its neural mechanisms. Indeed, metacognitive protocols that assess the animal's degree of confidence in its own responses have become simple enough to lead to neurophysiological experimentation. Kiani and Shadlen (2009) used J. David Smith's "escape key" paradigm in a statistical decision-making lecture. The animal is asked to make a binary decision about the direction of movement. On some trials, even before this decision is made, a third response possibility appears ("escape" response). Choosing this target leads to a fixed but smaller reward. This choice can therefore be interpreted as a refusal to respond to the main task, which could indicate that the animal is uncertain and lacks self-confidence. Indeed, behavioral observations indicate that objective performance is better when the animal has the option of refusing to respond than when it does not. This means that, given the same stimulus, the animal correctly discarded trials where it felt unable to respond.

What signals enable the animal to calculate its level of confidence in its responses? Neuronal discharges in the lateral intraparietal area (LIP) reflect not only the upcoming decision, but also the animal's confidence in this response, even before the animal knows whether or not the escape option will be offered. In addition, trial-to-trial fluctuations in neuronal discharges, measured just before the option of refusing to respond is presented, predict the choice of this option: the animal selectively chooses to avoid responding, precisely on trials where its parietal neurons do not sufficiently discriminate between the two responses of the primary task. The rate of increase of discharges during stimulus presentation also contributes independently to the choice of the option to refuse to respond. Taken together, these results are well described by a simple mathematical model that reduces metacognition to the level of a simple higher-level decision.

Thus, the parietal cortex includes not only neural decision signals, but also what we might call "meta-signals" that code for the confidence we can have in the decision to come. It seems likely, however, that the LIP area is only one element of a larger network, probably involving the prefrontal cortex, whose other nodes have so far hardly been the subject of cellular recordings. In fact, only research in rats, using recordings of neurons from the orbitofrontal cortex (Kepecs, Uchida, Zariwala & Mainen, 2008), converges on similar results. They confirm that, among the many metacognitive faculties, the judgment of trust is one of those simple enough to be studied today in animals and modeled at neuronal level.

In humans, brain imaging is making it possible to explore more widely the networks of brain areas associated with metacognition, without however going down to the level of the single neuron. A converging body of research links metacognitive thinking abilities to the most anterior regions of the prefrontal cortex, notably Brodmann's area 10. As early as 2002, a functional MRI study of thefeeling of knowing showed that several regions, all prefrontal, increased their activity in direct proportion to the subjective impression of knowing (Kikyo, Ohki & Miyashita, 2002). Lesioning the mesial prefrontal cortex can selectively decrease the "feeling of effort" associated with a difficult cognitive task (Naccache et al., 2005). Similarly, the subjective feeling of having seen a masked stimulus is selectively impaired by an anterior prefrontal lesion (Del Cul, Dehaene, Reyes, Bravo & Slachevsky, 2009). In healthy volunteers, transcranial magnetic stimulation of the dorsolateral prefrontal cortex affects second-order metacognitive judgment of confidence in one's response, without altering first-order psychophysical performance (Rounis, Maniscalco, Rothwell, Passingham & Lau, 2010).

Perhaps more surprisingly, individual variations in these introspective abilities can be related to variations in brain organization (Fleming, Weil, Nagy, Dolan & Rees, 2010). Indeed, the people who are most effective in making second-order judgments are those with the most grey matter in the right frontopolar cortex. Metacognitive competence is also correlated with participants' white matter density in the rostral region of the corpus callosum, which interconnects the right and left prefrontal regions. Finally, there is a correlation between brain activation itself and metacognitive abilities: when participants in an fMRI examination are classified according to the correlation between their introspection and objective success, this score is strongly correlated with the activity of the right frontopolar cortex during task performance (Yokoyama et al., 2010).

The frontopolar cortex, which overlaps with Brodmann's area 10, would therefore play a very special role in second-order metacognitive decisions, perhaps by performing evidence accumulation on signals from first-level decision mechanisms (Kiani & Shadlen, 2009; Rolls, Grabenhorst & Deco, 2010). Interestingly, this area is of recent evolution, and its expansion particularly differentiates the human species from other primate species (Semendeferi, Armstrong, Schleicher, Zilles & Van Hoesen, 2001).

At the end of this lecture, it is clear that introspection and, more generally, metacognitive abilities have become respectable fields of research, with unquestionable psychological status, and are the subject of in-depth investigations in cognitive neuroscience. Current research demonstrates the complexity of metacognitive functions and the diversity of brain systems involved. As early as the sensory areas, a form of metacognition probably appears, insofar as each brain area represents not only a series of elementary sensory decisions, but also the uncertainty attached to them. Various networks in the prefrontal and parietal cortex then intervene to recode, at a higher level, the conscious and reflective decision, confidence in that decision, anticipation of reward, self-detection of errors, and resumption of metacognitive control. These networks are present in a variety of animal species, enabling them to gauge the confidence associated with each action. Finally, the frontopolar cortex and the parieto-temporal junction seem to form a particularly well-developed network in the human species, enabling us to self-assess our performance and relate this personal assessment to the representation of the thinking of others.