Cognitive Science Lunchtime Talk - Angela Radulescu
Dynamic interaction between reinforcement learning and attention
Abstract: Decades of animal learning research suggest that humans and animals alike learn to predict future outcomes. Based on such predictions, they are able to make decisions that maximize the amount of reward they expect to receive in the future. However, in most real-world tasks, there is not one stimulus that does or does not predict reward, but rather there is a plethora of stimuli, most of them irrelevant to prediction and decision. A fundamental and still mysterious aspect of learning is what features of our complex environment should the mind selectively focus on: how do we know what to learn about in the first place? In this talk, I will present data suggesting that reinforcement learning in multidimensional environments relies on selective attention to uncover those aspects of the environment that are predictive of reward; in turn, variables underlying the learning process provide a kind of “saliency map” that dynamically guides attention. Using a direct measure of attention obtained by tracking trial-by-trial changes in eye position, I will also show how two separate attentional processes contribute to decision-making: attention for choice determines which features are used for assigning value to each option and attention for learning determines which features to learn about in light of feedback.