A study on neural task dimensionality of prefrontal cortex for context-dependent reinforcement learning

Abstract

An increasing amount of evidence suggests that high-dimensional neural representations of task-relevant state information, such as stimuli, might play an important role in perceptual decision making. However, little is known about the neural representational geometries of goals and contexts during learning and their effects on choice behavior. In this work, we used human functional magnetic resonance imaging data collected with a sequential two-stage Markov decision task to investigate the neural embeddings of goals and contexts during reinforcement learning (RL). We focused on the inferior lateral prefrontal cortex (ilPFC), the brain area known to guide goal-directed behavior by choosing optimal RL strategies for a given context. Unlike other brain areas, the ilPFC represents high-dimensional multiplexed context representations. Notably, the ilPFC’s goal dimensionality is seemingly a unique predictor of task performance in terms of average reward and choice optimality. Our study elucidates the critical role of goal and context embeddings of the ilPFC in guiding goal-directed behavior.