Learning How to Learn: The Interaction Between Attention and Learning as a Mechanism for Dimensionality Reduction in the Brain

Abstract

Evidence suggests that attention and learning interact to help extract task-relevant dimensions in a complex environment. How exactly these two mechanisms interplay and craft an internal representation of our environment is still unclear. We tested human participants on a multidimensional task, the Dimensions Task, in which they had to learn through trial and error to maximize reward. We found behavioral evidence suggesting that the learned reward value of a feature influences how that feature is encoded in the brain. Analyses also suggest that participants attend to the whole dimension of higher rewarding features. By drawing on theoretical and other behavioral findings of attention, we present an improved model of human decision-making in the Dimensions Task, which mimics learning in naturalistic task settings. Model comparison provides empirical evidence of informational benefits of selective attention. We hypothesize these benefits, implemented through competitive statistics and gated learning according to the confidence in state representation, aid the brain in simplifying and learning in a high-dimensional world.