Abstract
Learned associations between sensorimotor features of available choices and the value of outcomes guide many real-world decisions. Reinforcement learning (RL) approaches attempt to account for such decisions, but are often applied in a manner that implicitly assumes only relevant, attended feature-value associations are tracked, updated via reward prediction errors (RPEs), and employed in decisions. We have previously found evidence that "task-irrelevant" location/motor-related value tracking signals appear in subcortical regions related to motor and reward, such as the ventral pallidum, as well as a cortical region of ventral temporal cortex. In this study, we examined whether task-irrelevant, purely visual features are tracked in a similar manner. Participants (N=20) were scanned using fMRI while they completed a "4-armed bandit" dynamic reward-learning task and were instructed to track shape-value while ignoring color and location. On each trial, four shapes (circle, square, pentagon, and octagon) were randomly positioned in four locations on the screen, and randomly bound to each of four different colors (red, yellow, green, and blue). Participants were correctly informed that value was associated with shape, with the value of each shape determined by a hidden probability of receiving a reward that varied stochastically. Behaviorally, we found weak evidence based upon RL modeling that, in addition to shape-value, participants also tracked color-value and location-value and allowed them to influence choice. Replicating earlier work, we found correlates of location-linked RPE signals in subcortical regions of the basal ganglia, as well as a cluster of ventral temporal cortex. Critically, we also saw color-linked RPE signals in subcortical regions of the basal ganglia. These results suggest that both motor and non-motor "irrelevant" attributes are latently tracked with respect to their association with reward. Such latent signals may serve to guide exploratory actions, or actions taken under high uncertainty.
Meeting abstract presented at VSS 2017