​THE DIFFICULTY OF LEARNING IN A CHANGING WORLD

To be efficient in a changing world, our learning mechanisms must adapt. Indeed, when an upheaval renders what we have learned previously useless, we must quickly relearn from new observations (flexibility). On the other hand, if nothing changes, it is in our interest to combine past and present information to refine our knowledge (stability). This balance, at the heart of adaptive learning, is particularly difficult to find in practice. Mathematical tools such as Bayesian inference offer optimal solutions to this type of problem (see our news of February 2022). Bayesian models also postulate the existence of a confidence weighting mechanism: learning should be modulated by the level of confidence that accompanies predictions. Based on the idea that our sense of confidence allows us to arbitrate between past knowledge (stability) and new observations (flexibility), researchers have designed increasingly powerful algorithms for adaptive learning. However, the neural basis of this confidence is much less known than that of surprise.

IDENTIFICATION OF THE REGIONS INVOLVED IN THE REPRESENTATION OF CONFIDENCE

In this study, researchers used a dynamic probability learning task and high-field functional MRI to identify cortical regions potentially involved in the representation of confidence in predictions during human learning. Their work was based on four representation criteria commonly used in neuroscience: sensitivity, specificity, invariability, functionality. They show that the representation of confidence is computationally distinct from that of surprise and that the effects of the two quantities overlap anatomically, particularly in parietal and frontal cortex, in a manner qualitatively consistent with the Bayesian method and invariant to the predicted item. By showing that the activity patterns of the localized regions correspond to both a confidence and a surprise effect in accordance with the confidence weighting principle, the authors suggest that the confidence representation has a functional role in learning.

Further studies should clarify the mechanisms of confidence-weighted learning and the extent to which representations of confidence are subjective or stimulus-based (and possibly ideal), consciously accessible or implicit, and invariant across task and domain. The precise localization of effects enabled by surface-based analysis and 7T fMRI could guide more targeted recordings in the future, such as electrophysiology.

Contact : Florent Meyniel (florent.meyniel@cea.fr)   florent.meyniel@gmail.com