Neurocomputational Dynamics of Sequence Learning

The brain is often able to learn complex structures of the environment using a very limited amount of evidence, which is crucial for model-based planning and sequential prediction. However, little is known about the neurocomputational mechanisms of deterministic sequential prediction, as prior work has primarily focused on stochastic transition structures. Here we find that human subjects' beliefs about a sequence of states, captured by reaction times, are well explained by a Bayesian pattern-learning model that tracks beliefs about both the current state and the underlying structure of the environment, taking into account prior beliefs about possible patterns in the sequence. Using functional magnetic resonance imaging, we find distinct neural signatures of uncertainty computations on both levels. These results support the hypothesis that structure learning in the brain employs Bayesian inference.