A central goal of systems neuroscience is to determine the functional-anatomical basis of brain-wide activity dynamics. While brain activity patterns appear to be low-dimensional and guided by spatial gradients, the set of gradients remains provisional and their mode of interaction is unclear. Here we applied deep learning-based dimensionality reduction to task-free fMRI images to derive an intrinsic latent space of human brain activity. Each dimension represented a discrete, dynamically fluctuating spatial activity gradient. The principal dimension was a novel unipolar sensory-association gradient underlying the global signal. A small set of gradients appeared to underlie key functional connectomics phenomena. Different task activation patterns were generated by gradients adopting task-specific configurations. Dynamical systems modelling revealed that gradients interact via state-specific coupling parameters, allowing accurate forecasts and simulations of task-specific brain activity. Together, these findings indicate that a small set of dynamic, interacting gradients create the repertoire of possible brain activity states.

A dynamic gradient architecture generates brain activity states