Conference on Cognitive Computational Neuroscience

Reinforcement learning uses prediction errors to update expectations about future events. Model-free reward prediction errors (RPE) have been localized in the ventral Striatum, whereas model-based state prediction errors (SPE) can be found in the intraparietal sulcus. Here we use functional magnetic resonance imaging in humans solving a two-step Markov decision task to investigate changes in neuronal activation patterns following a prediction error (PE). First, we locate reward and state prediction errors to extract trial-wise BOLD signal estimates of both error types. Then, by applying a novel searchlight-algorithm, we use the PE-induced BOLD responses to predict pattern changes between corresponding events in current and upcoming trials. We found pattern changes distinctly correlated to model-free RPEs in the orbitofrontal cortex and to model-based SPEs in the superior parietal lobule. Additionally, both types of PEs were able to explain pattern changes in cognitive systems associated with attention, working memory and motor planning. These results reveal the immediate reconfiguration of neuronal representations following two different types of PEs during learning.