Graduate Students Seminar

Speaker 1: Pratyusha Sarkar

Title

Variational Inference: A Scalable Alternative to MCMC

Abstract

Variational inference (VI) is a scalable alternative to Markov chain Monte Carlo (MCMC) for approximate Bayesian inference. It frames inference as an optimization problem by selecting a distribution from a parametrized family that is closest to the true posterior, typically via the Evidence Lower Bound (ELBO). Using simplifications like the mean-field approximation and coordinate ascent updates, VI achieves efficient computation, making it well-suited for large data and complex models such as Gaussian mixture models.

Speaker 2: Maliha Noushin

Title

COVID-19 Transmission Modeling and Sensitivity-Based Prevention Strategies

Abstract

Mathematical modeling is an important way to understand how diseases work and to identify effective strategies to stop them from spreading quickly, especially when there aren't any vaccines or specific antiviral treatments available. Accurate forecasting is crucial for planning and responding to events like the COVID-19 pandemic. In this talk, we discuss three SEIR-based models: a reaction-rate formulation with multiple quarantine and hospitalization compartments, an extension incorporating contact tracing and hospitalization strategies, and a protection-rate model that adds quarantined, recovered, deceased, and protected classes. We carry out time-series simulations to visualize outbreak dynamics, then apply partial rank correlation coefficient (PRCC) analysis and variance-based Sobol sensitivity analysis to identify critical parameters of the models. Our results show that at the start of an outbreak, how quickly the virus spreads matters most; in the middle phase, measures that protect people are most important; and the length of quarantine has the greatest effect toward the end. We also found that protection efforts work best when combined with well-timed quarantine. These insights enhance the predictive power of epidemiological models and offer concrete guidance on when and how to deploy interventions for maximal impact.