Advising statement
The language and intelligence group ("Lingo") currently works on a relatively broad set of topics around (you guessed it) language and machine + human intelligence. Currently, most lab members are thinking about language models in way or another, with a focus on scientific understanding (how do they work? what tools do we need to understand them? what does understanding even mean?), interactive learning and teaching (how do we reliably get knowledge, skills and goals out of users' heads and into automated decision-making systems, or vice-versa?), and robust reasoning (how do we train models that know what they don't know? what inspiration can we draw from classical formal / probabilistic methods while maintaining the flexibility of the current ML paradigm?).
Here's a little about how we operate:
Goals PhD programs are designed both to generate knowledge and to train new researchers. These functions are equally important, but (especially in a research landscape that evolving as quickly as the current one) I believe it's more valuable for students to explore a broad set of problems, technical tools and collaboration styles than to produce a totally coherent body of work on a single topic. By the end of their time at MIT, students should have their own research agenda and the expertise to pursue it independently.
Research style We began life as an NLP group. (There have been some heated debates about whether we are still!) In any case, we value linguistic insight and other forms of domain expertise as much as machine learning expertise. And even though we now use the term ``language model'' to refer to just about any neural sequence predictor, we remain specifically interested in human language as a distinctive object of study. We try to avoid low-hanging fruit and tune our baseline models carefully. But at the end of the day, the primary function of papers is to generate insight, not improvements on evaluation metrics. We generally don't work on problems with established leaderboards or shared tasks. A successful project is one that shapes the way researchers think long after the specific model or task being described goes out of use.
Group structure Historically, about half of our lab members have come in with previous NLP experience and the other half had more general ML, math, or cognitive science backgrounds. Many people have spent some time in industry or post-bac programs (and several applied to MIT more than once!) The group is currently eight PhD students and three postdocs (some co-advised) along with a rotating group of visitors, undergrads and master's students. I expect the group to consist of six to eight PhD students and zero to two postdocs at steady state. I prioritize personal interaction and hope to play an active role in every project that lab members are working on. I also want to keep the group small enough that everyone knows (and has opinions about!) what everyone else is doing—peer mentoring is often more useful than top-down advising.
Meetings I meet with most lab members once or twice a week for at least half an hour. For students involved in bigger group projects, at least one of these meetings is one-on-one. (Some people also like to do short check-ins much more frequently.) These meetings can involve high-level project planning, low-level technical discussions, and general life / career chat. In addition to one-on-one meetings, we hold a lab meeting once a week which generally features research updates, paper discussions, outside speakers or practice talks; in some semesters we run additional focused on specific topics of interest.
Life outside the lab We like to spend time on things other than research! Work–life balance is important for intellectual development, health and personal relationships. We organize regular group outings (e.g. hikes, picnics, kayaking trips), and many people are also involved in athletic, artistic or cultural pursuits outside of MIT. The office is usually empty on weekends.