Dynamic computation of hierarchical prediction errors during sequence learning

点击翻译按钮获取中文摘要

Neuroimaging research on sequence learning has proceeded largely independently of the research on computational reinforcement learning, because the focus is primarily on learning statistical dependencies of stimulus sequences rather than on optimizing performance through learning. Here we want to investigate whether and to what extend mechanisms of model-free reinforcement learning are also be employed in sequential learning. Using an experimental paradigm where we precisely control temporal dependencies by first and second-order transition probabilities, we propose a combined fMRI-computational modeling study in order to test the following hypotheses: (1) The human brain implements common computational reinforcement-learning mechanism for estimating higher-order sequential dependencies in a similar way as it estimates average rewards. (2) Distinguishable higher- and lower-order prediction errors are encoded in different temporal orders and in dissociable brain regions. (3) The connectivity among brain regions that encode different orders of conditional probabilities is modulated by average rewards.