Exploring Neural Correlates of Flexible Cognition During a Complex Decision Making Task

Abstract

Efficiently solving complex decision making tasks typically arises from the use of schemas, abstract structures that organize experience and guide integration of novel information. There is a wealth of knowledge related to how the hippocampus, the prefrontal cortex, and communication between the two areas relate to schema formation, decision making, memory encoding, reward representations and behavioral planning. Less focus, however, has been placed on tasks involving ambiguous credit assignment, where feedback cannot be uniquely attributed to a single feature. In this case, information associating rewards with multiple stimulus features needs to be integrated across choice history to identify which features give rise to specific rewards. How does neural circuitry support a learned schema in complex credit assignment problems? To understand the neural dynamics, I focus on data from non-human primates solving a multi-variate bandit task involving ambiguous feedback of stimulus features, recorded in the Buffalo lab at UW. Through behavioral modeling, I find that non-human primates solve the problem with a reinforcement learning framework, suggesting schema-like learning by biasing choices towards recently rewarded features. Behavioral modeling allows us to fit an estimate of the belief state of the non-human primate, which we leverage to seek encoding of abstract feature variables in the hippocampus, prefrontal cortex, and basal ganglia. Particularly, decoding of abstract feature variables during the feedback period, along with multiple timescales of feedback specific modes, suggest that the neural circuitry updates belief states by integrating feedback information with working memory. Further, bidirectional information flow between the hippocampus and prefrontal cortex provides evidence for the hypothesis that the hippocampus guides neocortical plasticity while the prefrontal cortex stores and retrieves memories. Altogether, these findings suggest possible dynamical representations for computations following a reinforcement learning framework, and these findings lay the groundwork for understanding the underlying circuitry that supports schema representations.