Presynopsis seminar
Speaker: Unnati Nigam, IIT Bombay
Host: Radhendushka Srivastava
Title: Quasi-Periodic Gaussian Processes: Statistical Foundations, Predictive Control, and Deep Generative Models
Time, day and date: 11:30:00 AM – 12:30:00 PM, Wednesday, July 29
Venue: Ramanujan Hall
Abstract: Quasi-periodic signals arise naturally in a wide range of applications, including engineering systems, robotics, and biomedical data, yet existing modeling approaches often lack either computational scalability or the flexibility to capture complex temporal structures. This talk presents a unified framework based on Quasi-Periodic Gaussian Processes (QPGPs) that addresses these challenges through a dynamic equation formulation. The framework enables efficient modeling and prediction of quasi-periodic processes and is subsequently applied to predictive Iterative Learning Control, where it is shown to improve trajectory tracking in both simulation and real robotic experiments. The talk further introduces an extension to heterogeneous quasi-periodic processes for modeling non-stationary disturbances with theoretical performance guarantees and concludes by demonstrating how the same mathematical framework can be incorporated into deep generative models through a QPGP-constrained Variational Autoencoder for structured latent representation learning. The presentation will highlight the key mathematical developments, theoretical results, and experimental validations that form the main contributions of my thesis.