Subjective confidence reflects representation of Bayesian probability in cortex

Abstract

What gives rise to the human sense of confidence? Here, we tested the Bayesian hypothesis that confidence is based on a probability distribution represented in neural population activity. We implemented several computational models of confidence, and tested their predictions using psychophysics and fMRI. Using a generative model-based fMRI decoding approach, we extracted probability distributions from neural population activity in human visual cortex. We found that subjective confidence tracks the shape of the decoded distribution. That is, when sensory evidence was more precise, as indicated by the decoded distribution, observers reported higher levels of confidence. We furthermore found that neural activity in the insula, anterior cingulate, and prefrontal cortex was linked to both the shape of the decoded distribution and reported confidence, in ways consistent with the Bayesian model. Altogether, our findings support recent statistical theories of confidence and suggest that probabilistic information guides the computation of one’s sense of confidence.

Awareness of the Relative Quality of Spatial Working Memory Representations

“…Similar results are observed for a single remembered location (Li et al, 2021), a single cued location among multiple items (Yoo et al, 2018), and a single cued color among multiple items (Honig et al, 2020). Moreover, a pair of recent fMRI studies have established a link between the quality of decoded neural representations from areas of visual and parietal cortex and reports of memory uncertainty for spatial location (Li et al, 2021) and grating orientation (Geurts et al, 2022). Both of these studies observed a trial-by-trial correlation between a model-based measure of uncertainty of the decoded neural representation and a behavioral report of memory uncertainty, supporting a model whereby observers are directly reading out the quality of their neural representations when making an evaluative judgment of a single WM representation.…”

“…For example, the activity pattern across a population of neurons in primary visual cortex accurately encodes the orientation of an oriented grating which can be decoded using a variety of machine learning methods (Berens et al, 2012;Stringer et al, 2021;Walker et al, 2020). Theoretical (Ma et al, 2006) and empirical (Geurts et al, 2022;Li et al, 2021;van Bergen et al, 2015;Walker et al, 2020) studies have supported the notion that neural population codes additionally (and implicitly) encode the 'quality' or 'uncertainty' of a representation. Rather than separately keeping track of the feature value represented by a neural population and its uncertainty, both of these types of information are simultaneously encoded in a probability distribution carried by the neural activity pattern.…”

“…Indeed, models involving random fluctuations in the representation of each remembered item outperform those lacking such variability (van den Berg et al, 2014Berg et al, , 2012. Moreover, participants appear to be sensitive to these random fluctuations: behavioral measures of recall precision track behavioral measures of uncertainty and/or confidence (Geurts et al, 2022;Honig et al, 2020;Li et al, 2021;Rademaker et al, 2012;van den Berg et al, 2017;Vandenbroucke et al, 2014;Yoo et al, 2018), and the quality of information represented by neural activation patterns in visual and parietal cortex correlates with behavioral recall performance and confidence reports on individual trials (Geurts et al, 2022;H.-H. Li et al, 2021). Altogether, these results point to a model whereby WM representations of individual objects are instantiated within noisy neural activity patterns, and participants 'read out' both the represented information and the uncertainty with which that information is represented from these neural activity patterns when making WM-guided decisions.…”

An insula hierarchical network architecture for active interoceptive inference

“…Finally, the aINS generates higher-order interoceptive predictions and interoceptive prediction errors by computing the difference between its predictions and forward signals arriving from the dINS. Evidence that the insula is involved in computation of predictions or more specifically in generation of interoceptive predictions and uses Bayesian belief updating is still scarce, but has started to emerge from human and animal experiments [134][135][136][137][138][139][140][141][142][143].…”

Perception is rich and probabilistic

“…The same also applies to moving stimuli in the MT cortex 3 and neural firing in response to visual stimuli in general resembles probability density functions 4 . Decoding of these neural activities, both in visual and prefrontal cortices, demonstrate correlations with reported confidence 5 , suggesting that people are aware of and can report their internal representations at least to some extent. Given that neural representations are probabilistic, how could perception be discrete?…”