Learning in changing environments
People often take advantage of trends and regularities that they detect in their environment to adapt their behavior. It is true even when such trends actually change across time, as it is often the case in real-world, dynamic environments. In this talk, I will characterize some properties of the powerful machinery that the human brain uses to perform such a statistical inference, based on behavioral data, functional magnetic resonance imaging (fMRI), magneto- and electroencephalography (M-EEG) and computational modeling. Notably, this machinery is:
(1) Based on an estimation of transition probabilities between successive observations.
(2) Bayesian: people use and estimate uncertainty in a rational manner.
(3) Accessible to introspection: people can report both the estimated probabilities of future observations and the confidence that accompanies their inference.
(4) Flexible, because it gives more weight to recent observations, which is indeed optimal in dynamic environments.
(5) Hierarchical: it disentangles distinct but inter-related sources of uncertainties.