Benefits *and costs* of selective attention for human decision-making under uncertainty

Making decisions under uncertainty requires inferring the state of one’s environment from imperfect data, whether it is the cause of ambiguous sensory signals or the value of a possible course of action. Beside sensory errors and exploratory choices, recent research has identified the limited computational precision of hidden-state inference as a large contributor to the variability and suboptimality of human decisions made under uncertainty. In this talk, I will present recent work from my group which studies how selective attention provides benefits but also incurs significant costs for sensory- and reward-guided decision-making in the presence of ‘computation noise’. During visual categorization, humans use selective attention to filter and compress sensory signals to mitigate the negative effects of computation noise on inference. And during reinforcement learning, humans use selective attention to facilitate credit assignment in a ‘closed-loop’ regime, at the risk of reinforcing spurious associations between stimuli and outcomes due to computation noise. Together, these findings offer a nuanced view of the role of selective attention for decision-making based on noisy computations.