Recurrent interactions between neurons generate dynamic patterns of activity that serve as a substrate for behaviorally relevant computations. However, we do not yet have a principled framework for relating neural dynamics to neural computations. We have recently synthesized a theory that explains how low-rank recurrent neural networks may serve as a building block for computations. Our overarching goal is to integrate insights from this theory with behavior and electrophysiology in awake, behaving monkeys to establish a principled framework relating neural dynamics to neural computations. The project will start with reverse engineering low-rank network models that capture cortical dynamics in simple timing tasks. We then move systematically toward progressively higher rank network models that can perform timing tasks with progressively more sophisticated computational demands such as probabilistic inference of time intervals.
We aim to create models that simultaneously succeed in performing task-relevant computations (i.e., behavior) and emulate cortical dynamics recorded in monkeys performing those tasks. We will use this iterative process to establish a principled framework relating neural dynamics to neural computations underlying inference. Finally, we will put this framework to test using a novel task that demands an unprecedented level of computational flexibility.
It has become increasingly apparent that the neurobiology of behavior in health and disease has to be probed at the level of populations of neurons. However, we do not yet have a rigorous and quantitative language for linking population neural activity to behavior. Our work combines primate electrophysiology with neural network modeling and aims to develop such a language through the mathematics of dynamical systems. The results hold promise for future translational research to diagnose behavioral symptoms of brain dysfunction in terms of their computational modules and the dynamic patterns of activity that support those modules.
