Grad student recruiting for 2024 has recently wrapped up. This round, I sent my admitted students a sort of brain dump about my advising style and other lab logistics. I've realized this may be useful as a point of reference for future prospective students (or just prospective grad students in general, to get at least one data point of what they might expect during a PhD). So, I've decided to post it publicly here (after editing a bit to make it less 2024-specific and detail-oriented). Note that this is all subject to change! I am always trying to tweak my advising style to maximize my student's happiness and success. On that note, if, after reading, you have any ideas as to how I might improve, please let me know!
I generally aim for my lab to have ~12 people in it across PhDs, undergrads, master's students, and postdocs. I find 12 is big enough that you can easily find collaborators and people to ask questions but small enough that you still know what everyone is working on and I have enough time to give everyone a good amount of attention. Regarding undergrads and master's students, I'm very happy to have PhD students mentor them, and there generally is a lot of undergraduate student interest at U of T.
I generally meet with PhD students weekly either 1:1 or as part of a project-specific meeting with multiple attendees. If I haven't met 1:1 with a PhD student for a while because we've only been meeting in project meetings, I make sure to meet with them 1:1 from time to time to see how things are going at a broader level. I also try to have longer yearly 1:1 meetings with my students to check in on progress and give feedback. We have a 2-hour weekly lab meeting; the first hour is generally devoted to something specific for that week and the second hour is for all of us to briefly and informally talk about what we've been up to over the past week. Finally, we also have weekly “office hours”/co-working time slots where we all aim to be in our lab space at the same time to make sure we have additional time together.
In general, my goal is to help each of my PhD students become totally self-sufficient expert researchers by the time they graduate. This involves competency in a range of skills - coming up with project ideas, deciding what experiments to run, running them, finding relevant literature, writing papers, presenting them, etc. etc. I prioritize helping each student develop the skills that they need the most work on. For example, if a student has limited experience writing papers, I might do most of the writing for their first paper, do a lot of co-writing/editing for the second and third papers, and then try to have them write full first drafts on their own after that. Part of the motivation for prioritizing broad skill development is that all of these skills are necessary for a professor to have, and I assume that all of my students might try to go on the academic job market (mainly because I think the preparations needed for academic jobs are a superset of those needed for industrial jobs). To give a rough outline of how I'd expect a typical PhD to go, in a student's first year, my highest priority is just that they publish a first-author paper to help get their feet on the ground and build confidence. For the first paper, I'm less concerned with skill development or a long-term research theme. During the second year, though, I'd try to help the student identify a longer-term research theme to work on and maybe publish a first paper on that theme. Assuming that happens, the third year should be about executing on that theme, possibly by starting to build some collaborations in order to start to complete a larger body of work. In the fourth year, my hope is that the student can start to shift from “first author” to “middle author” on many papers, and perhaps start to mentor some junior students if they haven't already. The fifth year is then dedicated to thesis writing and the job search.
I like to have the members of my lab work on one or a few unified high-level themes (for example, “decentralized ML”). This helps foster collaboration and amplify the impact of each individual project. However, I don't ever force my students to work on specific projects (and fortunately I virtually never have funding that would require me to do so). Instead, I try to brainstorm projects that are interesting to them and leverage their strengths while fitting into the broader goals of the lab. I also don't force all of the work my students do to be relevant to our high-level themes, and I value some exploration. Generally I do think it's best if a student is not spending the majority of their time working projects that are outside my interests and area of expertise, simply because I can't be as effective as an advisor on such projects. On the whole, I am very collaboration-forward, both within and outside of the lab — I am thrilled to add collaborators to a project if it will speed up progress and improve results. I personally like to be relatively closely involved with the projects my students are working on. One way to characterize my minimum level of involvement would be to say that I want to at the very least able to write the introduction of a paper on the project. In some cases, I like to actually contribute (e.g. write code) but this is somewhat rare (though I've been trying to put things in place to make this more common in the future).
As my PhD student, you will be affiliated with two institutions: The University of Toronto and the Vector Institute. Among those two institutions, Vector is the more atypical one, so let me describe it in some detail. The Vector Institute is not a “part” of U of T; it's a separate institute that faculty at various universities can be affiliated with. Vector currently provides three major benefits to my students: they receive an additional stipend, they have access to more compute (more on this later), and they have access to the Vector space. Vector recently moved into a nice new building on U of T's campus. The space serves as a centralized location for all Vector members to work. The benefits of this centralized space are huge - it means you're part of a larger community of experts working in the field. So, for example, if you want some advice on some specific topic or want to find a collaborator, you can often walk a few desks away and find one. The size and benefits of this community I think are probably unmatched in North America.
We have roughly three sources of compute: The U of T cluster, the Vector cluster, and cloud compute. Faculty in the Department of Computer Science (like myself) can add GPUs to the U of T cluster; I haven't done so yet but plan to soon. In the meantime, we can use the other faculty's compute on a preemptible “scavenger” tier. Vector has its own cluster which currently has ~1000 GPUs of varying power (from T4 to A100), but Vector is currently in the process of spending a huge amount of money on additional GPUs. I won't get into specifics here except to say that I think in a few year's time, the compute access at Vector will be world-class, if not the world's best, among academic institutes. Finally, I occasionally acquire cloud compute to try to prevent any of our projects from being bottlenecked by compute. This sometimes is free (e.g. TPUs from the TRC program) and is sometimes on random cheaper clouds (like vast.ai) but I think serves an important need. We also sometimes get access to compute via collaboration, which I'm always supportive of.
I am totally happy with my students doing internships and encourage it. It's a good way to make more connections, work on more resource-heavy projects, and make some extra money. I'm happy with students doing internships at any time of the year, and am happy to be involved in the projects but only really want to be involved if I would be concretely helpful (i.e. I don't think there's much benefit in adding me if you're already getting all the “advising” you need). Since I've worked in industry a good amount, I have a good network of connections to help students get internships, and I think my students have been pretty successful at getting good offers (at Google, Microsoft Research, Meta, etc. etc.).