Conference on Cognitive Computational Neuroscience

Predicting what will come next in a changing and stochastic world is difficult. However, humans are able to do so based on past observations in a reasonably accurate manner by following the principles of Bayesian inference. In particular, they have a sense of confidence about their predictions and use it to regulate the updates of those predictions. Here, we probed the neural representation of this confidence in human adults during a probability-learning task with 7T fMRI. Participants were shown binary sequences of visual stimuli generated from an abruptly changing Bernoulli process. They covertly estimated the latent item probability and occasionally reported it along with their estimation confidence. We modelled learning with an (Bayesian) ideal model and distinguished prediction (the probability estimate), confidence (the estimate's precision), predictability, and surprise. Participants' reports correlated with those of the ideal observer. Our results unveiled a neural representation of confidence in a fronto-parietal network where the fMRI activity was 1) sensitive to confidence, 2) specifically so with respect to confounds (surprise and predictability), 3) invariant to which item is predicted in the sequence, and 4) functional inasmuch as it overlapped with a representation of surprise (relevant for confidence-weighted updates), and predicted the subjective confidence reports.