The Compression of Mutually Predictive Task Sets During Learning

Abstract

During learning, humans form basic stimulus-response (S-R) associations in prefrontal brain regions. Higher-order structures in working memory (WM), know as task set representations, group together S-R associations necessary to perform a specific task. Task sets representations may form when S-R associations covary in a novel task, resulting in mutual information between task sets. We predict that compression of redundant information will occur in WM as human subjects learn the statistical structure of a task, leading to more efficient storage and processing in prefrontal brain regions. In a task containing mutually predictive task sets, we found evidence that participants consolidated task sets containing redundant information by examining response time (RT) differences when subjects switched from independent task sets to symmetrical task sets. In a second experiment, we ruled out an alternative episodic memory explanation for the RT effects. Taken together, these results suggest that prefrontal regions capitalize on statistical regularities in order to efficiently encode novel representations.