Abstract:Neural manifolds are an attractive theoretical framework for characterizing the complex behaviors of neural populations. However, many of the tools for identifying these low-dimensional subspaces are correlational and provide limited insight into the underlying dynamics. The ability to precisely control this latent activity would allow researchers to investigate the structure and function of neural manifolds. Employing techniques from the field of optimal control, we simulate controlling the latent dynamics of a neural population using closed-loop, dynamically generated sensory inputs. Using a spiking neural network (SNN) as a model of a neural circuit, we find low-dimensional representations of both the network activity (the neural manifold) and a set of salient visual stimuli. With a data-driven latent dynamics model, we apply model predictive control (MPC) to provide anticipatory, optimal control over the trajectory of the circuit in a latent space. We are able to control the latent dynamics of the SNN to follow several reference trajectories despite observing only a subset of neurons and with a substantial amount of unknown noise injected into the network. These results provide a framework to experimentally test for causal relationships between manifold dynamics and other variables of interest such as organismal behavior and BCI performance.

Model Predictive Control of the Neural Manifold