Is human statistical learning Bayesian?
Statistical learning is the capability to estimate the latent statistics of events in the world based on (sequences of) observations. In this talk, I ask whether human statistical learning is “Bayesian”, distinguishing different aspects of Bayesian inference. Bayesian inference is often used to define the optimal performance in a task; response accuracy is thus taken as evidence (although not sufficient) of a Bayesian process. Bayesian inference is also notorious for the use of priors, the estimation and use of uncertainty, and the use of conditional probabilities, making computations over hierarchical models possible. I will provide evidence based on behavioral experiments and neuroimaging (functional magnetic resonance imaging, magneto-encephalography) that human statistical learning exhibits, at least qualitatively, those different properties of Bayesian inference.