Gaussian Variational Inference

Gaussian Variational Inference (GVI) is an optimization-based approach to approximate complex probability distributions with simpler Gaussian distributions, primarily aiming to efficiently estimate posterior distributions in challenging scenarios. Current research focuses on developing and refining GVI algorithms for diverse applications, including robotics (sequential GVI), full-waveform inversion, motion planning (GVI-MP, PGCS-MP), and contextual bandits (VITS), often leveraging natural gradient methods and manifold optimization techniques to improve efficiency and convergence. These advancements are significantly impacting fields like robotics, machine learning, and inverse problems by providing robust and computationally tractable solutions for uncertainty quantification and probabilistic inference.