Robust modeling of within- and across-area population dynamics using recurrent neural networks

点击翻译按钮获取中文摘要

Over the past several decades, the ability to record from large populations of neurons (e.g., multi-electrode arrays, neuropixels, calcium imaging) has increased exponentially, promising new avenues for understanding the brain. These data have the promise to provide a qualitatively different view of activity within and across brain areas than was previously possible, but the effort will require the development of advanced analytical tools. One natural framework is provided by the tools of dynamical systems, which offer the means to uncover coordinated time-varying activation patterns expressed across an interconnected network of recorded neurons, and to characterize how these patterns relate to behavior. This framework has provided fundamental new insights into information processing in these cortical circuits, including those underlying motor, sensory, and cognitive processes. However, previous analytical approaches to uncovering dynamics have typically been developed and tested in specific brain areas, for limited behaviors, in restricted behavioral settings. Ironically, it is not unusual for these methods to have 10^5 parameters that need to be set or learned, and require careful tuning to properly function. Yet the brain is not homogenous, and it is unclear how well these approaches can be made to generalize to a variety of brain areas and behaviors, let alone by researchers who are not intimately familiar with the methods. Further, assuming that the brain's dynamics stem from independent, isolated areas is a vast oversimplification. Clearly, perceptual, cognitive, and motor functions all rely on activity distributed across multiple, interacting brain areas, each of which likely has distinct dynamics. Communication between areas is a dynamic process that underlies flexible function. There is growing recognition that population dynamics are specifically structured to support inter-area interaction, and an immediate need for methods to accurately uncover dynamics between interacting areas. We will address the challenge of generalized applicability to diverse brain areas by developing a powerful new open-source toolkit for automated discovery of neural population dynamics, within highly divergent brain areas. Further, we will extend this toolkit with new neural network architectures to model the dynamics between interacting areas. Our approach, the Dynamical Systems ID toolkit (DSID), will support accurate and straightforward application to data from different brain areas and behaviors without requiring great expertise or infrastructure setup. DSID will leverage sequential autoencoders (SAEs), powerful and flexible deep learning architectures that use recurrent neural networks to characterize nonlinear dynamical systems. We will validate the generalizability of DSID using a combination of previously-collected and new multi-electrode recording data from monkeys, including motor, sensory, and cognitive areas of cortex. Following their development and validation in our labs, we will work to disseminate them throughout the appropriate research communities where we expect they will be further developed with application to an even broader range of brain areas and behaviors.