The effectiveness of our collaboration largely depends on whether there’s a good fit. Are we intrigued by similar questions? Can we meet each other’s expectations? I’ll try to summarize my working style below and cover typical questions I get from prospective students.
What do you work on? Currently, my group is working on model robustness (overcoming spurious correlations, model calibration, OOD detection) and human-machine collaboration (collaborative writing, learning from human feedback). We are also trying to gain a better understanding of large language models (in-context learning, reasoning, safety). The best way to get an overview of the type of research we do is to skim our papers on these topics, or talk to students and postdocs in the group. One’s research interests may change over time, their style/taste less so. I’m generally excited about new ways of thinking about a problem or solving a problem. In other words, I’d be more interested in adding a dimension to a problem space than improving along an existing dimension. I use a simple rule to judge the value of a paper: how many people would find it useful? There are many ways to make contributions. It may directly benefit a downstream user or provide insights to people working on similar problems. But the bottom line is that it has to be useful to someone.
How do your students choose research topics? This varies from person to person. Some students come in with a broad interest in an area (e.g., generation), in which case I would suggest concrete directions and then we scope the project together. Others have more specific proposals on what they want to work on and I give suggestions on how to make the project more interesting or tractable. I also encourage students to look beyond a single project and think about long-term questions they would like to pursue. In general, both of us contribute ideas and shape the project. The goal is to reach something that we both are excited about. I’m open to work on new directions, but if you feel strong about working on something that is not immediately within my research interests, then you’ll need to convince me that it is exciting to pursue :). The process is more enjoyable if our goals are aligned, which is why research fit is critical.
How do you work with students? I like to play an active role in all projects of my students. We have 1-1 or project meetings at least once a week for 30 to 60 minutes. The meeting can be about any aspects of the project, such as brainstorming ideas, going over results, discussing technical details, planning projects, etc. Sometimes we also chat about high level stuff like career planning or retrospect of a project. Outside of the meetings, we communicate on Slack frequently to discuss ideas, quick results, and related work.
What is it like being in the group? We have weekly group meetings where students present their research update and receive feedback from each other. New ideas or collaboration often come out from such discussions. Every month we hold an open-ended discussion on broad topics in AI (e.g., poorly understood phenomena in deep learning, AI alignment). There is also a low-committement social hour every week where we have snacks and chat about anything that interests us. Most students also interact with the larger ML^2 and CILVR labs. The ML^2 lab has weekly lunch meetings where NLP students present their work or recent papers. The CILVR lab has a weekly seminar that features in-person or online talks from external speakers with a broad range of topics (vision, robotics, optimization, machine learning etc.). Outside work, we have regular social events (board game, hiking, picnic etc.).