Wu Lin

Wu Lin is a Research Fellow specializing in computational aspects of differential, geometric, and algebraic structures in machine learning. Their research focuses on geometric methods for numerical optimization and approximate inference, particularly natural-gradient (NG) descent techniques. They investigate adaptive optimization methods, structured optimization, and Bayesian approaches in deep learning. Their work emphasizes practical and numerically-stable algorithms, including memory-efficient implementations and second-order methods.

Research interests include exploiting hidden structures and symmetries in ML models, with emphasis on optimization challenges and statistical inference. Key contributions include foundational papers on structured NG descent, Riemannian gradient methods, and variational inference techniques. Their blog series provides accessible introductions to natural-gradient concepts.

Publications span top venues like ICML and AI&Stats, addressing topics from curvature learning in neural networks to scalable Bayesian methods. While no institutional affiliation is explicitly stated in the provided text, their work is associated with Toronto, Canada. No awards or grant information is provided here.