Robust and Conjugate Gaussian Processes Regression

This 44-minute lecture by François-Xavier Briol from the Finnish Center for Artificial Intelligence (FCAI) explores robust and conjugate Gaussian processes regression. Learn how to perform provably robust Gaussian process regression without sacrificing the computational advantages of closed-form conditioning. The presentation addresses a common limitation in standard GP regression—the assumption of independent and identically distributed Gaussian observation noise—which often leads to unreliable inferences when violated in real-world applications. Discover the Robust and Conjugate Gaussian Process (RCGP) approach that enables exact conjugate closed-form updates in all settings where standard GPs work, making it particularly valuable for applications ranging from Bayesian optimization to sparse variational Gaussian processes. Briol, an Associate Professor in Statistical Science at University College London who leads the Fundamentals of Statistical Machine Learning research group, shares insights from his research on developing statistical and machine learning methods that effectively merge large-scale scientific models with data while maintaining robustness to model misspecification.