Talk Title:

Latent States of Brains and Behavior

Abstract:

New recording technologies are transforming neuroscience, allowing us to measure the spiking activity of thousands of neurons in freely behaving animals. These technologies offer exciting opportunities to link brain activity to behavioral output, but they also pose serious statistical challenges. Neural and behavioral data are noisy, high-dimensional time-series with complex dynamics. I will present our work on state space models (SSMs) for neural and behavioral time-series. The key idea is that underlying these high-dimensional data are low-dimensional latent states, which shed light on how neural circuits compute and how natural behavior is organized. First, I will present new models that decompose natural behavior into sequences of discrete, stereotyped, and reusable actions called “syllables,” while simultaneously allowing for continuous variations in speed or vigor. By disentangling discrete and continuous variability, we obtain parsimonious representations of moment-to-moment behavior that correlate with fluctuations of dopamine in the striatum. Next, I will show how the same simple building blocks can be composed into flexible and expressive models for neural recordings. Our models are not only among the most accurate models for multi-neuronal spike trains, they also achieve state-of-the-art performance on a host of other sequence prediction benchmarks in machine learning. Together, these lines of work highlight how advances in machine learning and statistics offer powerful new tools for linking brain activity and behavior.