(667g) Discovering Latent Dynamics Embedded in Large-Scale Neural Spiking Activity

Over the last several decades, there has been increased interest in developing computationally affordable models of large-scale neuronal activity and brain dynamics to better understand complex behaviors underlying cognitive functions. There is growing evidence in the brain-machine interface literature that there exist low-dimensional neural manifolds embedded in the high-dimensional spatiotemporal neuronal spiking patterns underlying reaching tasks. These low-dimensional neural dynamical structures have not only been found in the motor cortex, but also the somatosensory and visual cortices, as well as the hippocampus. Recent developments in machine learning techniques have enabled reduced-dimensional analysis of high-dimensional dynamical systems in data-driven frameworks with a low computational cost. Here, we develop a scalable data-driven computational framework to discover and systematically analyze the spatiotemporal dynamics of these low-dimensional neural manifolds embedded in the high-dimensional neuronal spiking activity underlying cognitive/non-cognitive tasks. We develop a bidirectional recurrent artificial neural network (RNN) framework to infer the firing rates of a neuronal population directly from the binned spiking data. Using a data-driven sparse identification of nonlinear dynamical systems approach along with autoencoders, we extract a reduced-order model of the low-dimensional neural manifold dynamics and identify the governing equations describing the time-evolution of the low-dimensional neural manifolds. We first validate this framework on synthetic datasets of large-scale network spiking activity with known governing firing rate dynamics. We then apply this framework on the datasets consisting of large-scale spiking activity recordings from the rat hippocampus during various spatial navigation tasks to understand the underlying neural dynamics.