Physics Colloquium with Anqi Wu on Addressing Challenges in Modeling and Understanding Neural Connectivity with Generalized Linear Models

Exposing meaningful and interpretable neural connectivity is critical to understanding neural circuits. In this talk, I will discuss two key challenges in modeling and understanding neural connectivity: dynamic changes in functional connectivity with varying behavioral states and dealing with missing neuron data. To address the dynamic nature of functional connectivity, we propose a novel prior-informed state-switching generalized linear model (GLM) that incorporates learnable Gaussian or one-hot priors. I will show that the learned priors should capture the state-constant connectivity, shedding light on the underlying anatomical connectome and revealing more likely physical neuron interactions. In the second part, I will introduce the missing-neuron scenario as a partially observable GLM (POGLM). To tackle the challenging inference problem of POGLM, I will present "variational importance sampling" (VIS), a novel approximate inference approach combining the principles of importance sampling and variational inference. It yields a substantially tighter lower-bound to the log marginal compared to the widely used Evidence Lower Bound (ELBO). I will demonstrate the effectiveness of POGLM with VIS through synthetic and real-world neural datasets, highlighting its practical utility in deciphering neural connectivity from spike train recordings.

Bio: Anqi Wu is an Assistant Professor at the School of Computational Science and Engineering (CSE), Georgia Institute of Technology. She was a Postdoctoral Research Fellow at the Center for Theoretical Neuroscience, the Zuckerman Mind Brain Behavior Institute, Columbia University. She received her Ph.D. degree in Computational and Quantitative Neuroscience and a graduate certificate in Statistics and Machine Learning from Princeton University. Anqi was selected for the MIT Rising Star in EECS, DARPA Riser, and Alfred P. Sloan Fellow. Her research interest is to develop scientifically-motivated Bayesian statistical models to characterize structure in neural data and behavior data in the interdisciplinary field of machine learning and computational neuroscience. She has a general interest in building data-driven models to promote both animal and human studies in the system and cognitive neuroscience. 

This lecture was made possible by the William C. Ferguson fund.