Bayesian model predicts confidence in perceptual decision

Bayesian models have been successful at describing perceptual decisions and behavior across tasks and stimulus modalities.To what degree they can also accurately describe meta-cognitive decisions, like confidence judgements, has been the subject of many studies, and is debated.One focus of those studies has been to understand whether confidence judgements reflect posteriors about (typically binary) decisions, posteriors over latent stimulus variables like orientation, or heuristic computations.In this work we point out that the predictions also depend on the details of the assumed generative model underlying perceptual inference in the brain.We demonstrate this point by comparing the predictions for an increasingly realistic set of generative models for one of the experiments presented in Adler & Ma (2018) in which human observers were asked to perform an orientation discrimination task, and report their confidence, using Gabor patches of a range of orientations and contrasts.We compare the generative model assumed in Adler & Ma (2018) to two alternative models.In the first, we explicitly include inference over contrast which was varied in the experiment.In the second model, we also explicitly model the fact that the brain's inference is not based on a single scalar observation but multiple (n) observations (e.g.diUerent spatial "channels") -similar to a generative model previously proposed to model sequential decision-making (Deneve 2012).We found that each of these Bayesian models made diUerent predictions for the previously described choice-consistency -confidence relationship (Caziot & Mamassian 2021).