Liu Hanqing

I am Liu Hanqing (刘汉青), a master’s student in the Koyama Lab at the University of Tokyo. Previously, I received a bachelor’s degree in physics and a master’s degree in history from Tsinghua University.

I study the theoretical foundations of deep learning, with a particular interest in how simple learning rules give rise to complex, sometimes surprising behavior.

I am currently studying the properties of deep generative models, focusing on generalization and memorization in diffusion models. Previously, I investigated loss spikes during neural network training and provided an explanation for the Slingshot mechanism.

My interests span learning dynamics and generalization, generative models, and statistical physics. I like connecting mathematical ideas with what actually happens when we train neural networks. See details on my research page.

About my name: Liu is my surname, and Hanqing is my given name. I prefer the Chinese name order, Liu Hanqing (roughly pronounced “Lyoh Hahn-ching”). However, I consistently publish as Hanqing Liu to avoid duplicate Google Scholar entries (a lesson I’ve already learned the hard way).

In my free time, I enjoy traveling, cycling, reading, live shows, anime, and video games.

I am seeking summer research opportunities for 2027 and plan to apply for Fall 2028 PhD programs. I also welcome research collaborations — please feel free to reach out.