Context-invariant neural dynamics underlying the encoding of Bayesian uncertainty, but not confidence

Abstract

Classic decision theories, such as reinforcement learning, typically require the presence of explicit outcomes for learning and belief updating. However, ecological environments are often opaque and explicit feedback, such as those based on values, might not be immediately accessible. It remains unclear whether the neural dynamics underlying belief updating in absence of outcomes differ from those responsible for decision-making based on accessible outcomes. Here, we investigated this question in healthy humans (n=28) using Bayesian modeling and two multi-option fMRI tasks, one with and one without immediate outcome. Model-based fMRI analysis revealed two opposing networks encoding belief updating regardless of the presence of immediate outcomes. A "confidence-building" network including the hippocampus, amygdala, and medial prefrontal cortex (mPFC) became more active as beliefs about action-outcome probabilities were confirmed by newly acquired information. Meanwhile, an "uncertainty-building" network including the anterior insular (AIC), dorsal anterior cingulate (dACC), and dorsolateral prefrontal (dlPFC) cortices became more active as new evidence conflicted with action-outcome estimates. Interestingly, dynamic causal modeling revealed that the confidence network was driven either by the hippocampus when outcomes were not available, or by the mPFC and amygdala when value-based outcomes were immediately accessible. Convsersely, the AIC always drove the activities of dACC and dlPFC, under the modulation of increasing uncertainty, independent of outcome availability. These findings reveal similar network compositions but distinct neural dynamics underlying belief updating in changing environments with and without explicit outcomes, highlighting an asymmetric relationship between decision confidence and uncertainty computation across levels of analysis. HighlightsO_LIWe investigated belief updating in two tasks, with and without explicit feedback. C_LIO_LIModel-based fMRI analysis revealed similar neural responses across tasks. C_LIO_LIThe anterior insula drove an uncertainty-encoding network, across tasks. C_LIO_LIThe anterior hippocampus drove a confidence-encoding network, w/o feedbacks. C_LIO_LIThe medial PFC and amygdala drove a confidence-encoding network, with feedbacks. C_LI