Impacts of feature-based value generalization on learning

Abstract

2022R1A2C2009174 차채녕 Given the limitation of resources, such as cognitive capacity and time, humans use various short-cuts in learning and making judgements. We propose that such an ‘efficient’ strategy in learning underlies how values of the choices which one never experienced or learned get determined. To test our hypothesis, we constructed a probabilistic reward learning task where reward contingencies associated with each item may change during the task. Most critically, for some individuals, a perceptual feature embedded in choice options and its relative position on the feature continuum (e.g., orientation of a Gabor patch) were informative for the reward contingencies based on the pre-existing experiences, while for other individuals, pre-existing experiences were not informative for learning reward contingencies of new items. Computational modeling revealed that individuals generalize previously learned feature-outcome association to subsequent learning processes and that this generalization is arbitrated against usage of belief about the task structure. These data provide a computational account of how individuals efficiently learn from a volatile and uncertain environment, and highlight the possibility where overgeneralization backfires and harms individuals’ performance.