Andrew Gordon Wilson

I aim to develop a prescriptive approach to building autonomous intelligent systems. This effort involves a variety of different research initiatives, which cumulatively work together towards achieving this vision. A major theme that unifies many of these initiatives is a desire for an actionable understanding, so that we can select for particular properties aligned with our goals. These areas, and some example papers, include: • Understanding deep learning models, including LLMs and vision models, generalization theory, and reasoning [e.g., 1, 2, 3, 4, 29, 30, 31] • Uncertainty representation, Bayesian methods, online decision making [e.g., 1, 5, 6, 7] • Distribution shifts, spurious correlations [e.g., 8, 9, 10, 11] • Encoding and learning inductive biases (e.g., equivariances) [e.g., 12, 13, 14, 15] • Linear algebra as a foundation for ML [e.g., 16, 33, 34, 17, 18, 19, 20] • Machine learning for physics, and physics for ML [e.g., 21, 22, 13, 20, 15] • Simple practical methods [e.g., 23, 24, 25, 26, 4] • Scientific discovery (protein engineering, materials design) [e.g., 27, 28, 32]