Bayesian inference optimally estimates probabilities from limited and noisy data by taking into account levels of uncertainty. I noticed that human probability estimates are accompanied by rational confidence levels denoting their precision; I thus propose here that the human sense of probability is Bayesian. This Bayesian nature constrains the estimation, neural representation and use of probabilities, which I aim to characterize by combining psychology, computational models and neuro-imaging.

I will characterize the Bayesian sense of probability computationally and psychologically. Human confidence as Bayesian precision will be my starting point, I will test other formalizations and look for the human algorithms that approximate Bayesian inference. I will test whether confidence depends on explicit reasoning (with implicit electrophysiological measures), develop ways of measuring its accuracy in a learning context, test whether it is trainable and domain-general.

I will then look for the neural codes of Bayesian probabilities, leveraging encoding models for functional magnetic resonance imaging (fMRI) and goal-driven artificial neural networks to propose new codes. I will ask whether the confidence information is embedded in the neural representation of the probability estimate itself, or separable.

Last, I will investigate a key function of confidence: the regulation of learning. I will test the implication of neuromodulators such as noradrenaline in this process, using both within and between-subject variability in the activity of key neuromodulatory nuclei (with advanced fMRI), the cortical release of noradrenaline during learning and its receptor density (with positron-emission tomography) and test for causality with pharmacological intervention.

Researcher: Florent Meyniel

Contact: florent.meyniel@cea.fr 

Grant agreement ID: 948105