This article is part of a series of articles on different European master’s programs related to artificial intelligence and machine learning.
Summary
This programme runs for 1 year, composed of 2 terms of 8 weeks, followed by a 3-month dissertation over the summer. You typically take 4 courses per term and are able to choose 1 course of the 4. There are 3 week-long practicals throughout the year, and two exams at the end of May. The courses are pretty theoretical, sometimes dry, but mostly good. The course as a whole is pretty good preparation for a PhD in stat-heavy ML, particularly if supplemented with seminars, keeping up-to-date on ML research trends, as well as extra-curricular coding.
The biggest downside of the course is the coding—it’s in R :( However, you don’t have to do much of this and can choose whatever language you want for your dissertation. The course itself doesn’t take too much time, so you should be able to spend a fair bit of time doing the supplementary activities I mention above, as well as maybe doing some research throughout the year with a prof, who are mostly keen to take students.
The biggest upside of the course is the people—academic and otherwise. Being in an Effective Altruism (EA) and AI-safety (AIS) hub has strong benefits, as does being able to attend lots of seminars and talks across departments.
Getting In
The acceptance rate is around 1⁄15. You’ll need strong grades from a good university. A decent majority of the cohort had research experience before they came. A minority had published. As a UK master’s, I expect that they expect less research experience than similar-quality non-UK programmes.
The Course
Relevance of courses
As you might expect, a statistics master’s will approach areas from a theoretical angle. While you gain a deep understanding of the underpinnings of ML, a fair bit of the material is dry (maybe ⅓?), and you won’t be learning the hottest ML—i.e. little discussion of the latest breakthroughs in DL or RL.
That being said, doing the master’s will make it easy to read any new ML papers, and the coverage of important probabilistic inference topics is good.
Quality of courses
¾ of the courses are of very high quality—lectures and lecture notes and problem sheets are typically very good and inspire insight.
Difficulty of courses
⅓ are easy, ⅓ are moderately difficult, ⅓ are pretty difficult.
There isn’t actually that much work—I actually ended up doing the vast majority of the course 2-3 months before exams (but I had lots of prior ML knowledge).
Practicals are easy.
Exams are, according to the external examiner, ‘significantly harder than any other master’s programme that she has experience with’ - for reference, she’s a Warwick stats prof.
Yet, 50% of people get a distinction.
Academics, Research Opportunities, Seminars
You can work with anyone for your dissertation, and people from all departments seem happy to take on stats students throughout the year, particularly if your coding skills are hot.
Yarin Gal—has weekly group meetings that you probably can and probably should attend.
Bayesian DL, but keen on safety topics as well: e.g. safe exploration, distributional shift.
His lab has many EAs in it, and some of the PhD students might be keen to have an RA.
Shimon Whiteson—has weekly group meetings that you probably can and probably should attend.
Inverse RL, imitation learning.
Yee Whye Teh—has weekly meetings in stats, which you can and probably should attend
He’s a big boi in stats who works on DL and probabilistic inference.
Less of a safety focus, but if you can work for him it will be good for career capital.
Various people at the future of humanity institute (FHI) have weekly group meetings that you can and probably should attend
Mainly Gaussian processes, active learning, Bayesian things. Could be interested in AGI. I don’t know much about him.
Oxford and Other People
The people in your course will be smart, as will be most of the people around you.
There is a very strong EA community, being around which I have found incredibly useful for forming better beliefs, being happier, and being a better human. If you can live in an EA house, then that helps lots.
There are lots and lots of people around that care about AIS, and whose models you can interrogate/steal.
The town is beautiful, and there are more cool things to do than you could possibly have time for.
Miscellaneous Considerations
The stats department is only for graduates, and you have lots of opportunities to interact with PhD students. The vibe of the department is really nice.
Increased likelihood of getting into a DPhil at Oxford.
If you choose to move careers outside of technical ML, then I think that having a master’s from Oxford looks comparatively stronger than most other programmes (provided you don’t have even cooler further credentials).
This course is a particularly good idea if you are considering econ (e.g. for GPR) alongside AIS.
University of Oxford, Master’s Statistical Science
This article is part of a series of articles on different European master’s programs related to artificial intelligence and machine learning.
Summary
This programme runs for 1 year, composed of 2 terms of 8 weeks, followed by a 3-month dissertation over the summer. You typically take 4 courses per term and are able to choose 1 course of the 4. There are 3 week-long practicals throughout the year, and two exams at the end of May. The courses are pretty theoretical, sometimes dry, but mostly good. The course as a whole is pretty good preparation for a PhD in stat-heavy ML, particularly if supplemented with seminars, keeping up-to-date on ML research trends, as well as extra-curricular coding.
The biggest downside of the course is the coding—it’s in R :( However, you don’t have to do much of this and can choose whatever language you want for your dissertation. The course itself doesn’t take too much time, so you should be able to spend a fair bit of time doing the supplementary activities I mention above, as well as maybe doing some research throughout the year with a prof, who are mostly keen to take students.
The biggest upside of the course is the people—academic and otherwise. Being in an Effective Altruism (EA) and AI-safety (AIS) hub has strong benefits, as does being able to attend lots of seminars and talks across departments.
Getting In
The acceptance rate is around 1⁄15. You’ll need strong grades from a good university. A decent majority of the cohort had research experience before they came. A minority had published. As a UK master’s, I expect that they expect less research experience than similar-quality non-UK programmes.
The Course
Relevance of courses
As you might expect, a statistics master’s will approach areas from a theoretical angle. While you gain a deep understanding of the underpinnings of ML, a fair bit of the material is dry (maybe ⅓?), and you won’t be learning the hottest ML—i.e. little discussion of the latest breakthroughs in DL or RL.
That being said, doing the master’s will make it easy to read any new ML papers, and the coverage of important probabilistic inference topics is good.
Quality of courses
¾ of the courses are of very high quality—lectures and lecture notes and problem sheets are typically very good and inspire insight.
Difficulty of courses
⅓ are easy, ⅓ are moderately difficult, ⅓ are pretty difficult.
There isn’t actually that much work—I actually ended up doing the vast majority of the course 2-3 months before exams (but I had lots of prior ML knowledge).
Practicals are easy.
Exams are, according to the external examiner, ‘significantly harder than any other master’s programme that she has experience with’ - for reference, she’s a Warwick stats prof.
Yet, 50% of people get a distinction.
Academics, Research Opportunities, Seminars
You can work with anyone for your dissertation, and people from all departments seem happy to take on stats students throughout the year, particularly if your coding skills are hot.
Yarin Gal—has weekly group meetings that you probably can and probably should attend.
Bayesian DL, but keen on safety topics as well: e.g. safe exploration, distributional shift.
His lab has many EAs in it, and some of the PhD students might be keen to have an RA.
Shimon Whiteson—has weekly group meetings that you probably can and probably should attend.
Inverse RL, imitation learning.
Yee Whye Teh—has weekly meetings in stats, which you can and probably should attend
He’s a big boi in stats who works on DL and probabilistic inference.
Less of a safety focus, but if you can work for him it will be good for career capital.
Various people at the future of humanity institute (FHI) have weekly group meetings that you can and probably should attend
FHI also has a summer research fellowship, which you might be able to tie your dissertation.
There is a strong AIS reading group.
Michael Osborne
Mainly Gaussian processes, active learning, Bayesian things. Could be interested in AGI. I don’t know much about him.
Oxford and Other People
The people in your course will be smart, as will be most of the people around you.
There is a very strong EA community, being around which I have found incredibly useful for forming better beliefs, being happier, and being a better human. If you can live in an EA house, then that helps lots.
There are lots and lots of people around that care about AIS, and whose models you can interrogate/steal.
The town is beautiful, and there are more cool things to do than you could possibly have time for.
Miscellaneous Considerations
The stats department is only for graduates, and you have lots of opportunities to interact with PhD students. The vibe of the department is really nice.
Increased likelihood of getting into a DPhil at Oxford.
If you choose to move careers outside of technical ML, then I think that having a master’s from Oxford looks comparatively stronger than most other programmes (provided you don’t have even cooler further credentials).
This course is a particularly good idea if you are considering econ (e.g. for GPR) alongside AIS.
Quick Links
Course website.
Handbook (descriptions of courses).
Feel free to email me at charlierogerssmith@gmail.com.