First Year Ph.D. Life at Penn

I came to Philly at Aug. 18, 2023 to start my Ph.D. study with a Hong Kong - Dubai - New York flight. This is the first time that I have took a A380, which is a double-deck, wide-body, four-engine jet airliner looking from outside, but the seat is so small. It was amazing to see the long island from the plane.

We rode a 3-hour ride taxi from JFK to Penn with traffic jam, that's why I choose to go to New York by train for all the later trips.

The biggest challenge of my life here is deciding what to eat. The only thing the food in US can do is to make me not starve. I rarely find myself enjoying the food, except for the few occasions when I visit Chinese restaurants. However, I am always happy when free food is available, as it means I don't have to worry about what to eat for that meal.

One of the positive aspects of the new Ph.D. requirements is that I only need to take four courses. In my first year, I took two classes, one each semester. For Fall 2023, I enrolled in CIS 5200. Although much of the content was familiar, I greatly enjoyed Jacob Gardner's teaching. He had a profound understanding of machine learning concepts and explained them clearly and thoroughly. This was also my first experience in an American-style class, which is very different from classes in China. Most students listened attentively and actively asked questions. In the Spring, I took Equivariant Deep Learning with Kostas and Jean. I must admit that I struggled with the course because the mathematics were too difficult for me to follow. It involved complex topics like groups and the Fourier transform. Interestingly, at the end of the class, I learnt the latest AlphaFold3 model even abandoned equivariance during training.

For Ph.D., of course the most important thing is research. My goal is to explore feedforward 3D generation. I began by investigating distillation from multistep diffusion models. I encountered mode collapse for a long time and then attempted to address it on ImageNet by referring to a previous paper. However, transitioning to StableDiffusion still didn’t work, and I eventually discovered that it was because I trained on only one prompt, which had little variance. It was quite frustrating that just one or two days after I achieved some good results, a paper was published on arXiv that did exactly the same thing with the same method. At first, I felt like all my efforts were in vain and didn't know what to do, but finally I tried to calm myself down. Although these efforts didn't result in a paper, at least I made it work and gained some valuable experience.

The first year of my Ph.D. went by much faster than I could have imagined. This week, I had a long meeting with my advisor, and she shared her own stories from her Ph.D. journey. I realized that one point that I am definitely lacking is having a high-level understanding of the papers I have read. I need to have a clear vision of what my future research should aim to achieve and dedicate myself to it. Previously, I was too focused on my small project and didn't think much beyond it. So, I am determined to establish this habit from now on.