Pathwise Conditioning and Non-Euclidean Gaussian Processes

Explore the intricacies of Gaussian processes and their applications in Bayesian optimization through this Google TechTalk presented by Alexander Terenin. Delve into the concept of pathwise conditioning as an alternative approach to traditional distributional methods for conditioning and computing posterior distributions in Gaussian processes. Discover how this perspective enhances the efficiency of acquisition function computations in decision-theoretic settings. Examine recent advances in this field and their broader implications for Gaussian process models. Learn about a novel class of Gaussian process models designed for graphs and manifolds, enabling Bayesian optimization that intrinsically accounts for symmetries and constraints. Gain insights from Terenin's expertise in statistical machine learning, particularly in interactive data gathering scenarios, and explore the connections to multi-armed bandits, reinforcement learning, and techniques for incorporating inductive biases and prior information into machine learning models.

Pathwise Conditioning and Non-Euclidean Gaussian Processes