Learning and Representation of Recent Structure in the Environment: Behavioral, Neuroimaging, and Computational Investigations

Abstract

Environmental statistics gradually come to be represented in cortical areas of the brain after extensive experience and long periods of time. In many contexts, however, we are exposed to new environmental regularities that influence our behavior very rapidly. What kinds of neural processes and representations support such rapid statistical learning? The medial temporal lobe (MTL) can learn new information rapidly, but it is traditionally thought to specialize in learning new arbitrary - not structured - information. Much of this dissertation work investigates whether this rapid learning ability may in fact extend to learning new structured information. In support of this idea, we found that representations of objects that appear nearby in time become more similar to each other throughout the MTL. Beyond indicating that the MTL is involved, these findings begin to suggest what kinds of representations it may construct to support statistical learning. We found the same kind of representational similarity in the hippocampus in a paradigm with more complex structure. In this paradigm, stimulus sequences were generated by a graph with community structure, where the strength of transition probabilities - a cue commonly considered to be critical for event parsing - was uniform, and therefore uninformative for parsing. We found that participants learned the structure nonetheless, as evidenced by event parsing behavior, and that representations of items from the same community came to be represented more similarly than items from different communities in the hippocampus, as well as in the inferior frontal gyrus, anterior temporal lobe, and superior temporal gyrus. Connectivity analyses suggest that the hippocampus may be a central hub in the network of regions involved in learning new events. We additionally found that a patient with MTL damage failed to learn new temporal regularities, providing evidence that the area is necessary for this form of learning. Finally, we ran experiments and developed a computational model suggesting that sleep may help consolidate recently learned structured information. This work begins to characterize the neural mechanisms underlying our ability to rapidly extract and consolidate regularities in a new environment.