Dynamic interaction between reinforcement learning and attention in multidimensional environments

When you are learning a new task with many possible cues, what you look at and what you learn from can interact with each other; but how that loop is implemented in the brain has not been studied. We had people learn by trial and error in a setting where only one of three kinds of information actually predicted rewards, while tracking their eye movements and using fMRI to decode which dimension they were paying attention to. Using a computational model of choice, we found that attention biased both choices and updating after outcomes. Reward-related brain signals scaled with that focus, and where people directed attention moved as learning progressed – supporting a two-way interaction between selective attention and reinforcement learning in high-dimensional environments.