Conference on Cognitive Computational Neuroscience

Accurate representations of uncertainty facilitate optimal behavior during learning and decision-making. Most previous human neuroimaging studies have attempted to link uncertainty to (average) BOLD signal magnitude. Although untested experimentally, it has also been postulated that a higher, more flexible neural dynamic range may be required whenever the “correct” state is unclear, potentially promoting learning. Here, we probed whether state uncertainty may indeed be best reflected in the moment-to-moment instability (variability) of the BOLD response. Participants performed a Bayesian inference task where they estimated the ratio of marble types in an unseen jar following exposure to five sample draws. We found that BOLD signal variability strongly decreased with successive sample exposure. Crucially, more accurate subjects expressed a greater exposure-related collapse of BOLD variability. Our computational model of behavior suggested that the relationship between BOLD variability collapse and accuracy could best be explained by individual differences in their prior belief before sample exposure; participants with a wider prior were more likely to (a) collapse SDBOLD with successive exposure and (b) perform more accurately (i.e., learn better) overall. We provide first evidence that BOLD variability may track the evolution of state uncertainty representations during Bayesian inference.