Real-time modeling and perturbation of stimulus-shaped neural manifolds

About This Grant

ABSTRACT Systems neuroscience is acquiring exponentially more neural activity data in vivo and is employing richer stimuli to study ethologically relevant behaviors. Models of neural dynamics in low-dimensional spaces (‘neural manifolds’) are increasingly state-of-the-art for describing the underlying neurobiological mechanisms to encode rich stimuli and evoke behavior. However, causally testing these theories remains challenging, as it requires dynamic manipulation of activity across neurons and time, conditioned on the individual brain, task, or environment. Here, we propose to build new machine learning methods to construct low-dimensional neural manifolds shaped by external stimuli in real time, and design perturbations of neural dynamics on these manifolds to directly test hypotheses of neural circuit function. We will broadly learn how external stimuli drive ongoing neural dynamics and which neuronal stimulations optimally align with these latent vectors. Our real-time approach will also enable us to consider multiple competing models in parallel and choose stimulations to differentiate between them, causally testing competing hypotheses of neural manifold landscapes. We will validate our models in vivo in the larval zebrafish using high-dimensional visual stimuli concurrently with holographic optogenetic photostimulation. In Aim 1, we will develop real-time dimensionality reduction algorithms to construct neural manifolds shaped by external stimulus information. In Aim 2, we will design probabilistic models for predicting the effects of neural stimulations on latent neural dynamics. In Aim 3, we will develop a method for jointly optimizing external stimuli and direct neural stimulation patterns to shape ongoing neural dynamics in real time. Successful completion of this project will result in generalizable machine learning methods that can automatically learn stimulus-shaped neural dynamics and optimize neural stimulations to drive dynamics on the manifold. These tools will be widely available and broadly useful to many neuroscientists.