Dr Jihao Andreas Lin

PhD thesis title

Scalable Gaussian Processes: Advances in Iterative Methods and Pathwise Conditioning

My PhD research introduces novel computational methods which make Gaussian processes more scalable and efficient. In particular, my work combines iterative methods for solving systems of linear equations with a technique called pathwise conditioning to allow Gaussian processes to run efficiently on modern hardware and handle millions of data points. Gaussian processes are a probabilistic machine learning method and a cornerstone of Bayesian optimisation. The latter is a powerful framework for global optimisation of complex black-box functions, which is commonly used in impactful applications such as automated machine learning, drug discovery, materials science, robotics, and autonomous systems.

After the PhD

After completing my PhD, I am joining the Adaptive Experimentation team in the Central Applied Science organisation at Meta as a Research Scientist. The team develops and applies optimisation techniques to various tasks, including A/B testing and experimentation, automated machine learning and hyperparameter tuning, infrastructure optimisation, and hardware design. It also creates and maintains the open-source software libraries Ax and BoTorch. Ax is a domain-agnostic platform for adaptive experimentation, which provides a user-friendly interface and automates the process of running sequential experiments. BoTorch is a library for Bayesian optimisation which is built on PyTorch. It provides a modular and powerful framework for researchers and developers to create advanced optimisation algorithms.