Abstract:
Uncovering the algorithms through which neural populations encode and produce rich behavior is a central goal in systems neuroscience. Existing methods to link neural activity and behavior, however, struggle to scale to these high-dimensional spaces involving thousands of neurons and complex behavior evoked in diverse environments. Modern machine learning techniques offer a powerful solution for discovering latent (lower-dimensional or unobserved) features in such high-dimensional datasets. Combined with precise stimulation technologies, we can begin to dissect large-scale circuits in vivo, constructing models that causally relate neural activity to behavior. Perturbative testing of hypothesized brain-behavior links, however, requires statistically efficient methods for both estimating behaviorally-relevant features and intervening on neural activity in real time. Here I will discuss a number of ways in which we can construct and refine machine learning models built in real-time, as neural and behavioral data are simultaneously acquired, and use them to identify and then causally test which behavior features are evoked and modified by ongoing neural latent dynamics.