Yifan Sun

Primary research interests

My research centers on the design and analysis of optimization algorithms, particularly those arising in machine learning and scientific computing, along with the learning-theoretic questions that accompany them. I'm interested in first-order methods and their convergence properties, including convergence rates under realistic, non-idealized assumptions; minimizing functions with extremely ill-conditioning (most good ML models); predicting the implicit bias induced by different optimization schemes; sparsity; handling and predicting stochasticity and stability in training dynamics; and characterizing the geometry of loss landscapes more broadly. (See publications / blog)

Student projects 

As machine learning, AI, and data science grow more complex and increasingly black-box, it becomes harder (for students, researchers, and practitioners alike) to keep pace and build the kind of foundational intuition needed to design good solutions to hard problems. With this in mind, students who work with me take on complex concepts: graph-based data structures, evolution of model embeddings, knowledge flow in large language models, loss landscapes of simple models; and build meaningful, web-based visualizations that make these ideas tangible and intuitive. The guiding belief is that making this technology simpler makes us stronger, more capable users of these tools. (See student projects)

If you are a student (undergrad or grad) interested in such a visualization project, contact me.  Strong coder preferred.