​OUR BRAIN, A CODE-CREATOR

This study explores a mechanism that remains poorly understood in cognitive science : how probabilities are encoded in the human brain. We constantly evaluate probabilities to navigate the world, whether to determine if our train will be on time or to predict tomorrow's weather. Because our behavior is so clearly adapted to these probabilities, it is reasonable to assume that they are encoded somewhere in the brain. However, identifying this code has proven notoriously difficult. One exception concerns the probability of obtaining a reward, for which neural correlates are better characterized. However, these studies do not address the probability of neutral events. Previous work has also identified neural correlates of probability-related quantities such as surprise or uncertainty, but has often failed to find a direct correlation with probability itself.
Here, the authors propose a radical shift in perspective by testing the hypothesis that the brain may use a highly non-monotonic code. This hypothesis had never been tested in previous exploratory studies, which all assumed that neural activity simply increases proportionally with probability.

PROBABILITY, A NON-MONOTONIC ENTITY

To test their hypothesis, the researchers asked twenty-six participants to perform a probability learning task while simultaneously measuring their brain activity using ultra-high-field (7 Tesla) fMRI. Participants were presented with a rapid sequence of binary stimuli, A and B, and were required to estimate the probability of observing A, knowing that this probability could change abruptly and unpredictably over time. Thus, participants inside the MRI scanner had to continuously infer this hidden probability mentally, without pressing any button or receiving an immediate reward. This protocol ensured that the recorded brain activity was not confounded by motor responses or reward-related signals.
To approximate the optimal estimate of the hidden probability, the team used an ideal mathematical observer that computes the exact probability estimate for each trial. The neural code underlying these quantities was then characterized by combining two analytical approaches integrating encoding models with approximation theory, allowing the probability code to be modeled without strong assumptions about its form. This innovative approach enabled the identification of a previously unknown representation of probability within the human dorsolateral prefrontal cortex and intraparietal cortex. Both univariate and multivariate analyses revealed that this representation relies on a highly non-monotonic code, explaining why it had not been detected in previous standard analyses.

In this study, the authors provide evidence for the non-monotonic nature of the neural code for probability, and strengthen their conclusion by comparing it to another quantity, confidence, whose neural code is highly linear. Given the diversity observed in the response curves, future studies will need to move beyond the assumption of monotonic response functions, or simple canonical forms, and instead investigate richer and more complex representations.

Joliot contact : Florent Meyniel (florent.meyniel@cea.fr)​