Abstract
The visual world is always changing. While we learn from regular patterns in our environment, we also learn from surprising changes that violate our beliefs, such as when objects move abruptly or people behave unpredictably. What are the brain networks predictive of surprise in the visual environment? To address this question, we reanalyzed openly available fMRI data collected as participants performed a task in which they learned to predict the location of an upcoming object (N=32; McGuire et al., 2012; Kao et al., 2020). McGuire et al. (2012) developed a normative model tracking surprise (sudden changes in the mean of an occluded generative distribution of the object’s location) and uncertainty (about the generative mean) in this task. To identify brain networks whose strength predicted these measures, we calculated the co-fluctuation time course of all pairs of 268 brain regions in a functionally defined atlas as the product of their z-scored BOLD-signal time series. Using leave-one-subject-out cross-validation, we identified region pairs (“edges”) whose co-fluctuation varied across trials with normative surprise and uncertainty. These edges predicted surprise and uncertainty in held-out individuals: edges positively correlated with surprise were stronger on trials with more unexpected outcomes (mean within-subject rho=0.08; t(31)=7.87, p<0.001) whereas edges negatively correlated with surprise showed the opposite pattern (mean rho=-0.07; t(31)=-6.61, p<0.001). We next asked whether the same edges predicted surprise in a naturalistic context. We measured dynamic changes in these edges in openly available fMRI data collected as novel participants watched NCAA basketball games (N=20; Antony et al., 2021). Edge strength tracked a validated measure of surprise (change in a team’s win probability) from advanced basketball analytics. These results suggest that brain dynamics in a common functional network predict surprise under very different visual contexts.