Shaping dynamical neural computations using spatiotemporal constraints*

The human brain is a physically embedded complex system that exhibits exceptional computational capacity despite its limited resources. These limited resources, which include constraints on space, material, and metabolism, systematically shape inter-regional connectivity, which in turn circumscribes the brain's computational dynamics. A key question in neuroscience is how the brain navigates these resource constraints to enable diverse human behavior and cognition. Here, we address this question using recurrent neural networks (RNNs). RNNs are artificial neural networks that can be trained to perform a wide variety of behavioral and cognitive tasks with high performance, making them a natural candidate for studying brain computation. However, RNNs are typically trained in the absence of any resource constraints, which allows them to develop connectivity that---while tailored for high task performance---is often biological implausible. Thus, in this work, we examine how spatial patterns of cortical organization affect the computational dynamics of RNNs. Throughout human development, the formation and refinement of inter-regional connectivity is guided by a spatially-embedded axis of cortical organization known as the sensorimotor-association (S-A) axis. Critically, this spatially-constrained neurodevelopmental program is thought to give rise to the staged acquisition of increasingly complex human behavior and cognition. When constraining RNN training using the S-A axis, we observed multiple connectivity features that are hallmarks of brain organization---such as the presence of modular subnetworks and top-down modulatory connections---emerged naturally in RNNs. Crucially, the emergence of these connectivity features precipitated increases in task performance, illustrating their computational significance. Our work shows that neurodevelopmentally-informed spatial constraints enable the training of biologically plausible RNNs. These biologically-informed RNNs allow us to probe potential causal mechanisms for the brain's computational dynamics.