Stats or CS PhD for ML Theory

[vraman]

Hi all, 

I'm interested in Machine Learning Theory and CS Theory and have been admitted to both CS (NEU, USC, BU) and Stats (UM) PhD programs. Which program is generally "better" for ML theory work? I am sort of leaning towards Stats because I feel the coursework would be more relevant (compared to like systems classes in CS). However, right now, I plan to go to industry after a PhD and can't help but feel that a CS PhD would make me more marketable. In the case I do decide to go the academia route, would a Stats PhD lower my chances of getting a TT position for ML theory work? My reasoning here is that I see a lot of ML theory peeps fall into the CS department

Edited March 4, 2021 by vraman
clarification

[Stat Phd]

Even though I want to do a stat PhD, I must admit that CS and EE is way ahead in ML. So if you really want that, go to CS.

[bayessays]

My suggestion would be to skim papers from professors at these departments to see what you're really interested in -- you can do work that falls in the "ML" category at each of them, but they'll likely be very different.  If the statistics papers are interesting to you, then you can do that type of theory work and you'd be set up well for a good stats job where you can do work that is "machine learning" but is close enough to statistics that it's fine -- many statistics professors now publish in conferences like NeurIPS.  Michigan is also a prestigious program and would be a very good place to launch a career from -- I'm not familiar with the CS academic job market, but I suspect it will be somewhat harder to get tenure-track jobs coming out of lower-ranked departments like BU/NEU but don't quote me on that.  Look up Tewari, Regier, Nguyen, Scott, Zhu, and Tan at Michigan, among others -- they have a lot of good ML people.

[Ryuk]

I can only provide advice as a fellow applicant interested in similar areas. I was also admitted to both CS and stat PhD programs. 

Because I am interested in both, the relationship between the stats and CS programs has been extremely important to my ultimate decision. For each of these schools, maybe see how many professors have dual appointments and publications across departments.

[vraman]

Thanks all for your insight! One thing to note is that after speaking with PhD Students in both programs, it seems like the coursework is more extensive for a Stat PhD. This may in turn slow research progress in the first couple years compared to a CS PhD. @Ryuk have you thought about this aspect of it? If so, how have you reconciled this "issue"?

 

[Ryuk]

@vraman To be honest, I didn't spend as much time talking to students in CS departments because I got into much better stats programs. I'm not sure what you learned from the CS students, but it seems that stats courses will definitely take up most of my first year and a decent chunk of my second. However, I'm certain that I'll get started on research next summer. I think that's expected at most stats programs in order to start finding an advisor. Do CS students begin research even sooner than that?

[vraman]

@RyukYup its not uncommon for CS students to publish in their first year. This is because: 1) coursework is very flexible 2) you get admitted basically with an advisor since you are typically not funded by the department (but rather by an advisor). Btw, I did a bit of stalking on your posts on grad cafe, and you got into some pretty amazing schools! Congrats and well deserved! I would love to collab

Edited March 6, 2021 by vraman

[bayessays]

CS PhD students that I know had lab groups from the time they started and high research output was pretty much expected from the start.  I agree that even though Stats might have some heavier coursework, you are also taking less completely irrelevant classes like operating systems and programming languages.

[vraman]

Right, I plan on pretty much doing the same stuff whether or not I go CS or Stats, so I guess I need to figure out whether the extra relevant coursework is worth it