Abstract
In the study of connectivity in large-scale networks of brain regions, a standard assumption is made that the statistical dependence between regions is univariate and linear. However, brain regions encode information in multivariate responses, and neural computations are nonlinear. Multivariate and nonlinear statistical dependence between regions is likely ubiquitous, but it is not captured by current methods. To fill this gap, we introduce a novel analysis framework: fMRI responses are characterized as points in multidimensional spaces, and nonlinear dependence is modeled using artificial neural networks. Converging evidence from multiple experiments shows that nonlinear dependence 1) models mappings between brain regions more accurately than linear dependence, explaining more variance in left-out data; 2) reveals functional subdivisions within cortical networks, and 3) is modulated by the task participants are performing.