Uncertainty-robust goal embedding in the prefrontal cortex for flexibly stable learning

Mounting evidence suggests that the prefrontal cortex encodes environmental uncertainty for value-based decision-making. However, how uncertainty influences goal-seeking at the neural level remains elusive. We used human fMRI data collected with a two-stage Markov decision task to investigate the neural embeddings of goal and uncertainty during reinforcement learning (RL). We found that the neural activity patterns of the lateral prefrontal cortex (LPFC) and orbitofrontal cortex (OFC) encode the immediate goals and the environmental uncertainty levels, compared with other brain areas involved in decision-making, including the hippocampus, primary visual cortex, and ventral striatum. We also found that the LPFC and OFC selectively represent uncertainty in specific goal-seeking conditions, suggesting that these brain regions use uncertainty information to guide goal-directed behavior. These results motivated us to examine whether and how the goal and uncertainty embedding in the PFC guides goal-directed learning. Our key findings are as follows: 1) LPFC shows mixed representations of specific goal x uncertainty (shattering dimensionality analysis), which explains behavioral flexibility. 2) Those neural representations of goal information are robust against uncertainty change (cross-condition generalization performance analysis), which explains behavioral stability. 3) Notably, both separability and robustness of goal representation predict optimal goal-seeking behavior. In summary, our study provides a detailed account of how goal and uncertainty embeddings in the LPFC guide flexible and stable RL.