We constantly generate predictions about the world, from tracking statistical regularities in our environment to anticipating the flow of a conversation. When these expectations are violated, we experience surprise—a fundamental signal that shapes attention, learning, and memory. My dissertation investigates both what impacts the formation of predictions and the consequences of their violation: how do our internal attentional states shape prediction, and how does the mind and brain compute surprise when predictions are violated? To address these questions, I combine behavioral measures, neuroimaging, and computational modeling. In Chapter 1, I show that sustained attentional states modulate how effectively individuals learn to predict visual regularities during ongoing experience. In Chapter 2, I demonstrate that the dynamics of a common fMRI network—edge-fluctuation-based predictive models (EFPM)—track belief-inconsistent surprise across diverse contexts, including associative learning, naturalistic sports viewing, and observing unexpected actions by agents. In Chapter 3, I investigate the computational basis of subjective surprise during narrative comprehension. Using large language model–derived metrics, I show that surprise is better captured by deviations in high-dimensional internal representations than by scalar measures of unpredictability. Specifically, representational prediction error predicts both self-reported surprise and neural dynamics in the EFPM network. Together, this work advances a unified account of predictive processing in which predictions are shaped by internal states and structured representations of the world, and surprise reflects representational prediction error shared across mind and brain. Looking ahead, this framework provides a foundation for investigating how predictive processes vary across contexts, development, and populations, and how they support learning and memory.
