Is statistics PhD program good enough to learn ML?

[Taxxi]

I had a brief discussion with one of my professors, and he advised me to apply for CS or ML. His opinion is that if I go to the statistics department, it is often way too theoretic (only dealing with something like Lasso) and he was worried that I might have way too few chances for applications. What is your opinion about this? Personally, I heard that CS focuses on research experience so I'm done for, but I might as well apply for some ML departments if this is the case.

[bayessays]

The line between ML and statistics is pretty blurry, but if you want to do things like deep learning, computer vision, NLP, a CS department will probably be a better fit.  You'll get plenty of practice doing things like linear models in statistics, if you consider that ML.  There are obviously exceptions though.

[Stat Assistant Professor]

The top statistics departments should have at least a few faculty who publish consistently in  top ML conferences, including ICML and NeurIPS. You can check the faculty pages to see if there are any who publish in such venues. Agreed with bayessays that CS would be much better for things like NLP and computer vision though. 

[Taxxi]

Thank you for your advice. I still decided to go for statistics since I want to focus on mathematical training. I guess I will take some deep learning courses later, if I am admitted somewhere.

[StatsG0d]

IMO, you can do ML research at either statistics or CS departments, but there are some differences (others can weigh in if it's different in their experience).

In stats, you might take a current ML algorithm or tweak one to prove that, under certain conditions, e.g., the classifier you developed is Fisher consistent or has some sort of nice asymptotic properties.

In CS, the focus is more on developing new algorithms or finding new optimization techniques to make current algorithms faster, and they are much less concerned with any probabilistic properties of these algorithms. 

If you're only interested in ML, my advice would be to choose a CS department, or choose a stats department like CMU, which has ML training embedded in its curriculum. I have found that for the most part, stats programs tend to consider ML an elective, but the stuff you learn in stats departments is only tangentially related to the stuff you need to learn ML, so it'll probably seem like a huge waste of time to deeply learn (no pun intended) linear models, generalized linear models, and measure theory to simply end up doing ML.

[Stat Assistant Professor]

Another main difference, I would say, is that it is virtually impossible to get into a CS PhD program without any research experience, whereas Stat/Biostat departments tend to de-emphasize research experience in favor of mathematical maturity. In CS, a 3.2 GPA (either cumulative or major) could be mitigated by strong research experience and a first/second-author publication in a prestigious ML conference, whereas Stat/Biostat departments place greater emphasis on grades (especially in math classes). In fact, for CS departments, the research practically *is* your application (see http://www.pgbovine.net/PhD-application-tips.htm)

But yeah, I agree with @StatsG0d that statistics departments will often emphasize things like asymptotic properties, uncertainty quantification, and mathematical foundations, whereas (applied) machine learning research groups in CS departments do not seem to care as much about theoretical properties or statistical inference/uncertainty quantification.  Predictive accuracy and computational efficiency are generally what is emphasized in CS. The culture is slowly changing at a lot of Stat departments though, with more and more departments placing high value on ML conference papers in top conferences rather than only mathematical papers in traditional (bio)stat journals. And nowadays, a lot of papers in JASA, JRSS-B, and Biometrika feature novel contributions to statistical computing, not just fancy-schmancy math. In the Bayesian statistics community,  it is also now well-acknowledged that scalability and computational efficiency are essential concerns from practitioners that we have to address. So the lines between CS and Stat are becoming blurrier... but still, certain topics like NLP and computer vision are almost exclusively in CS.

Edited December 13, 2019 by Stat PhD Now Postdoc

[Taxxi]

It was a good thing that I didn't attempt to apply for CS since I do not have any formal research experience ?

At least it is a good thing that I have genuine interest in CS as well. I wish I can study all the fun subjects.

Edited December 13, 2019 by Taxxi

[Spaghettini Plot]

It is true that computer vision is still primarily found in CS departments, but there are a few stats profs who do study it. UCLA for instance has a number of researchers https://vcla.stat.ucla.edu/people.html. There are others elsewhere who frequently work in CV such as Yali Amit at U Chicago and Zaid Harchaoui at UW.