Hi
Firstly, thanks a lot for such a detailed overview.
A prospective student here. I have an offer from ETH Zurich (not funded) and UvA (funded). I really like the weightage on research at UvA (48 CPs for thesis + potentially 2 AI Projects) which is conducive for me since I want to explore a couple of areas before I apply for my Ph.D. after masters. However, ETH is a better-ranked institute and also has a lot of wonderful labs, and is more flexible in terms of course choices.
Can you provide an update on what you decided in the end (and on what basis)? Are you happy with the choice you made?
(I'm also applying to these schools and would benefit from your take on this.)
Hi Piyush,
I'm sorry, but I do not have good answers to your questions.
Note that many students in Amsterdam are not at all interested in doing a PhD, and so the number of such students doing a PhD at UvA or a better institute may also not be that informative.
But I do know one UvA master's graduate who recently was admitted for a PhD in Cambridge (after he was rejected by Amsterdam itself - there's probably lots of noise in these decisions).
Other possibilities for earning money is to try to find a thesis at other institutions that pay their students like CWI, Mila,,that or CHAI in Berkeley (for all these, I know students who did their thesis there).
Prospective student here; what finding a thesis at another institution entail? Would a student who wanted to do this begin by emailing professors at these institutions about their research, applying to their "visiting researcher" programs, or something else?
Hi Aidan_Kierans,
In the three cases I know, it went like this:
CWI: the student got in contact with a researcher from CWI who was willing to supervise a thesis.
CHAI: Same.
Mila: Same.
Note that for CHAI and Mila, you may need recommendation letters if you go over their usual routes of research internships. I myself also applied to Mila and FHI and got recommendation letters for this, though in the end I was not able to get accepted there.
This article gives an overview of the master's program in Artificial Intelligence at the University of Amsterdam (UvA). It is meant to be both useful for people who want to decide where to study, as well as for UvA students who want to get information about how to make the most of their experience – including which courses are good and which research opportunities exist.
This article is part of a series of articles on different European master's programs related to artificial intelligence and machine learning.
Earlier versions of the post received feedback from Henning Bartsch, Remmelt Ellen, Micha de Groot, Benjamin Kolb, Evangelos Kanoulas, Dmitrii Krasheninnikov, Linda Petrini, Attila Szabó, and Oskar van der Wal. All mistakes and opinions are mine.
General note: The master’s program changed significantly in the last 3 years. I’m putting in the information of the newest edition, but cannot guarantee that the information is still correct when you apply.
I have a background in mostly pure mathematics, where I did a master’s degree at the University of Bonn before. After getting interested in the social movement of Effective Altruism, I tried out Software Engineering and Data Analytics and then began studying AI at the University of Amsterdam (not the same as the Vrije Universiteit Amsterdam). That was since I felt I want to go closer to research again, but in a field that’s more relevant to the planet than pure maths. I did my master’s in AI at the UvA from September 2018 to July 2020. In September 2020 I started a PhD in the intersection of Information Theory, Causality, and Deep Learning at the UvA in collaboration with the Computational Science Lab (CSL) and Amsterdam Machine Learning Lab (AMLab). If you have any questions, please contact me by email.
The master's program in AI at the University of Amsterdam is a research-oriented 2 years master program. “Research-oriented” means that a significant proportion of the degree consists of the students doing research, but usually still less than 50%, with the remainder being coursework. Graduation requires the students to collect 120 Credit Points (CP). Most courses (lectures/projects/seminars) are awarded 6CP upon completion, and a research thesis in the end is rewarded with 48CP. In continental Europe, master’s programs seem to be the requirement to start a PhD at most Universities, and this program tries to be a useful preparation both for a scientific and industry career, depending on the interests of the students.
Here you can find the admission criteria. The basic requirements are relatively low: You need some, but not much, prior knowledge in computer science, programming, calculus, linear algebra, probability theory and statistics, and meaningful motivation to do a master’s in AI. The more you know, the easier it is to do the master’s obviously, and a lack of mathematical ability seems to be the most common bottleneck.
On the website, you can find more ranking criteria that finally determine who gets accepted. The ranking criteria are roughly what you would expect.
The program accepts a maximum number of 150 students to the program. Prior to the last year, all students who fulfilled the mandatory criteria were also accepted, but this changed significantly due to the exponential increase in interest in AI. In the 2019/2020 academic year, only 30% of students who fulfilled the mandatory criteria were accepted. I do not know about the situation in the 2020/21 academic year.
The program starts in September. There are overall 40 working weeks per year, and 12 weeks of holidays: 2 around Christmas and 10 in the summer.
The 40 weeks are distributed across 6 so-called "periods", of which four are eight weeks long and two are four weeks long each. In an eight-week-period, a student usually has two lecture series which are taught in the first seven weeks, with an exam in the eighth week (with some exceptional lecture series extending over two periods). In a four-week-period, students have only one lecture series or project in a more intense format.
In the second year, students are encouraged to start their thesis work with an already chosen supervisor and a signed contract in November. The expectation is that the thesis can be finished within 8 months, i.e. until June of the second year. It is usually not a problem to take longer. However, this means the student has to pay the tuition fees for the additional months, which especially for overseas students is a significant amount.
Grading happens on a scale from 1 to 10, where a grade below 5.5 generally is a fail. The grade is often roughly equivalent to the percentage of points obtained in tests and homework. A Dutch 9.0 is in my experience roughly comparable to a German 1.3 or even 1.0. Generally, the dutch grading scale is viewed as demanding, and grades above 8.0 often considered to be good or very good grades.
The highest honor for a degree in the Netherlands is "Cum Laude" (not “Summa Cum Laude” as one might expect coming from Germany).
There are several rules for receiving Cum Laude, with the most important being that the grade average should be more than 8.0, no grade below 7.5, and the thesis at least 8.0. An 8.0 in courses requires work but mostly is doable, whereas for the thesis this is quite an achievement: I saw several thesis defenses with publishable (or already published!) work receiving an 8.0, and this grade is generally thought of as only attainable for conference-publishable work. Probably the most accurate information on what needs to be done to receive Cum Laude can be found somewhere here.
As said, there are mostly three types of courses: lectures, seminars, and projects. Attendance is usually not mandatory, apart from seminars.
First seven weeks:
Eighth week: Exam, usually 2 or 3 hours long.
See the section below on "Seminar combining symbolic and statistical methods in AI" for more details on a specific seminar (this was the only one I attended). In seminars, the students read contemporary work in a specific area, present on it, and work on reproducibility studies or on self-chosen research projects.
Furthermore, there is the course Project AI, in which students can either do an industry internship or a project together with a researcher at the UvA. There is no structured process for finding projects to the best of my knowledge. If you find a project, it may allow you to publish – we managed to put a workshop paper out (which usually is considerably less valuable than a conference paper, so choose your project wisely!).
A list and description of courses for the 2020/21 academic year can be found here. Note that there can be quite some changes to the courses and to what courses are obligatory, so when in doubt, consult the information on the website. A GitHub repository of a student with summaries and very detailed information about courses can be found here. Though note that there are some changes to the courses and not all this information may still be up-to-date.
7 courses are obligatory and cannot be chosen. Currently, the courses are:
Additionally, students need to choose 3 constrained-choice electives and 2 courses which are either constrained-choice or "free-electives". The free elective courses may or may not be related to AI, as long as the Examination Board approves. You can even choose courses from other universities, like the excellent Information-Theoretic Learning course at Leiden University which I enjoyed doing. In the appendix, you find more detailed information and my opinion on the courses.
The biggest part of the whole master's is writing a thesis in the end. The student receives 48 CP for it, which corresponds to roughly eight months of full-time work, i.e. 40% of the whole program.
When I started the program, the thesis was supposed to be only 36 CP, i.e. 30% of the program, but this has changed and I switched into the new regulations. The named reason for the change was that students took on average considerably more than the originally intended 6 months in order to finish their graduation project. Then, so the story, they wanted to adapt to the reality and give more room for the thesis.
I can think of several more reasons for this change and assume they are also part of the truth, although they were not officially mentioned. The following is pure speculation:
After a short, not necessarily representative, poll among students, it seems that roughly 55% of students do their theses at the university (including company-funded university labs) and roughly 45% at a company. Also, theses done at a company are judged based on scientific merit.
Roughly one month before the thesis starts, there was a big thesis fair with companies in which they presented their projects. The range of quality of the topics was really huge, with some being cutting-edge research problems that seemed more fundamental than many of the projects at the university itself, and others being more of the type "Hey AI student, I have a data set, please do something with it". Even though I did not end up at a company, my recommendation for AI students, even with the aim to do a PhD, is to at least consider the thesis fair and try to find the worthwhile opportunities among the projects.
Working at a company has the clear advantage that the student will usually also earn some money and can use the project as an opportunity to jump-start their working life. Note that some university labs also pay the students, although usually less than companies (I received 200€ from the QUVA Lab, and came with the cost of signing an NDA. This didn’t matter in the end, though). Other possibilities for earning money is to try to find a thesis at other institutions that pay their students like CWI, Mila,,that or CHAI in Berkeley (for all these, I know students who did their thesis there).
Projects at the UvA themselves are often not really advertised. There is a website with projects, and some researchers advertise their projects on their homepage. But overall my impression was that the students need to proactively reach out to researchers in order to find a project that fits.
My project is in the area of steerable and gauge equivariant CNNs. Gauge equivariant CNNs recently got some coverage. It's awesome to work in such a promising research area. My thesis is entirely theoretical (which was not originally planned as such, but turned out to be the best choice given how it developed).
It was recently accepted at ICLR.
I don’t feel qualified to comment on the high-level strengths and weaknesses of current research at the UvA. Some impressive research that originated in Amsterdam – highly biased by what interests me personally – are the variational autoencoder, graph convolutional networks, and the research surrounding gauge-equivariant convolutional networks. All of these were projects of PhD students supervised by Max Welling.
If you want to do research with one of the groups, the most natural courses for doing so are your 48CP master’s thesis or up to two 6CP Projects (or instead one 12CP project if it is long). A third possibility is to be an honours student. A fourth possibility is to be really ambitious in your project attached to a course. For example, projects in the new course in “Fairness, Accountability, Confidentiality and Transparency” or in the “Seminar Combining Symbolic and Statistical Methods in AI” have the chance of being (workshop)-publishable with extra work.
There are several research labs in Amsterdam that do research in machine learning and surrounding areas. Some have a large overlap – for example, the Amlab overlaps with both the Delta lab and the QUVA Lab, so don’t take them necessarily as independent evidence for excellence. Additionally, I don’t have a good sense of the hierarchical structures of “labs”, “groups”, and “professors”, and so some of the information below may be confusing. No guarantee for completeness!
Part of the Informatics Institute (IvI):
Institute for Logic, Language and Computation: As the name of the institute suggests, it conducts research on the intersections between Logic&Language, Language&Computation, and Logic&Computation. They have many groups which I do not all list and which may be interesting for AI students. An incomplete list of labs/groups:
Mathematical Institute: This is an institute on mathematics, and thus has only limited relation to artificial intelligence. However, in the track on stochastics, Joris Mooij leads the project on mathematical statistics. He conducts research in causal discovery and inference and they have a reading club.
This whole section is my own opinion and does not necessarily reflect the opinions of other students. A summary of my perspective: I’m a student mostly interested in research and very theoretical topics. I was certainly not the typical student of this program.
I have basically not much of an idea about how well the UvA does compared to other top Universities in AI and ML. It seems to me that it is the university in the Netherlands with the broadest AI abilities and options. As mentioned before, it is also an ELLIS unit and there seems to be a lot of new funding (as judged from the many new labs and PhD openings this year). Other strong options in continental Europe (non-exhaustive!) are ETH Zürich and Tübingen University – I don’t have an opinion about whether these Universities are “better” or “worse” on average than Amsterdam, since I don’t know enough about these programs. Also, global judgments of this sort may be the wrong way to look at things, since all Universities have their own strengths and weaknesses, and so the more you know about what is important to you, the more your decision may differ from what would usually be suggested.
Thus, what I’m doing now is listing some things that I think is done well and not so well in Amsterdam, in the hopes that this might inform your decision. Below I then give a tentative decision guide for deciding whether you want to study in Amsterdam.
You should consider the AI Master’s program of UvA if you
You should not choose UvA if you
Note that I mostly studied in Amsterdam. I did not really explore the city that much, and so for a broader overview, I suggest looking on the internet for more information.
Some features:
In general, I would say that Amsterdam is a nice, international, and modern city worth living for many people. For people who think that larger cities are two stressful, I suggest considering Tübingen, which is way smaller and thus also a bit cheaper to live in than Amsterdam (especially considering rental prices).
Those three courses were mandatory when I started, but are not anymore since they are offered from the Vrije Universiteit Amsterdam which does not collaborate as closely with the University of Amsterdam anymore. Their “Knowledge Representation” course is not the same as the new “Knowledge Representation and Reasoning” course from the University of Amsterdam.
Computer Vision is usually not well received from the students, but fortunately requires less work than other courses. It seems from my perspective a bit out of date and the lecturer not that motivated. In the past three years at least, students gave lots of feedback to improve this course. I know that this led to some changes in the course, but I do not know yet whether this led to a larger approval of the course by students.
Machine Learning 1 is basically a crash-course on maybe half of the Machine Learning book by Bishop. It’s a good course and alongside the course on Deep Learning probably the most important to get into modern ML research. Usually, the course is considered a lot of work.
This course from the second period covered many different topics in NLP. Overall, I liked this course considerably less than Machine Learning 1. The homework and lecture were overall a bit less structured, but part of it was also me being a bit less interested in the topic.
Overall, the course is amazing. Alongside with Machine Learning 1, it is probably the most important course for getting into ML research early. Here's the website of the course. And this is the homework. You learn to derive gradient descent in generality for yourself and how to code MLPs, including a manual backpropagation in pure NumPy routines. You learn to use PyTorch and to train big models on the Lisa cluster. Other topics include RNNs and LSTMs, graph-convolutional networks, and generative models, including GANs, VAEs, and Flow models, which the students investigate theoretically and in practice.
Students usually like the course but judge its workload as too high.
In my time, this was still a constrained-choice course, but they upgraded it to be in the mandatory courses, which is a useful change in my mind.
I was recently teaching assistant for the new FACT-AI course, which didn’t exist when I started studying. My impression is overall pretty positive. There’s no exam and only the homework and a final presentation are graded. In the homework, students could choose which paper in fairness or transparency of AI to reproduce/extend in teams. Here are two examples of papers that students could choose, the first about transparency and the second about fairness.
Overall, this course is very valuable after Machine Learning 1 and Deep Learning and brings students considerably closer to doing research.
Information Retrieval is about how modern search engines work. My personal opinion – this may not at all coincide with the opinions of other students – is that this course should not be in the mandatory courses: It’s structure was good, but I felt like the topics were not relevant enough as background for the rest of the master AI program in order to warrant a mandatory course.
This course is new. I assume it is meant to replace the course on Knowledge Representation of the Free University that dropped out of the main curriculum. I think learning about Knowledge Representation and Reasoning is actually a useful thing in order to understand both the limitations of these methods as well as the limitations of modern machine learning.
Between 3 and 5 constrained-choice courses have to be chosen by the students. Some words about the courses I took:
I took this course in my second period. It is basically about the modern formulation of Shannon's original work on the mathematical theory of communication. It differs considerably from other courses:
I really liked the course!
(The course does not appear in the course overview anymore. That may mean it does not exist anymore, but I’m not sure.)
This is the only seminar I went to in my master's, so not sure how representative this seminar is for other seminars. Unfortunately, this course will in the future only be “free elective” instead of constrained choice and so it is overall harder to find the time to do it.
Each week, students presented papers, each student presented 2 overall. Finally, in pairs of two students, everyone had to do a research project (with self-chosen ambition). Sometimes students produce something which is workshop-publishable.
Also, this excellent course now moved from being constrained-choice when I took it in period 6 to being a "free-elective" course. That means again that incentives to take this course, unfortunately, got smaller. I expressed to the program-committee my regret about this change and hope they will take it into account. Some projects lead to publishable results.
The course on Reinforcement Learning is basically a crash course on parts 1 and 2 of Sutton and Barto's excellent book on RL. Furthermore, also some Deep Reinforcement Learning not covered in the book is treated, including for example Trust Region Policy Optimization.
Next to theoretical and practical homework, the students also write a blogpost on a reproducibility study they do, where me and my colleagues investigated DQN.
Overall a great course!
This course basically teaches the other half of the Machine Learning book from Bishop?
Overall, this was in my mind a pretty good course. Many students found it to be too theoretical, but I personally found it pretty good. The teacher of the course was now promoted a professor for the maths department, but I heard it’s essentially the same under the new teacher. I'm not sure how the course will change with a new teacher.
In this category, students can collect at most 12 CP, which corresponds to two courses. I did two courses in this category:
This course is part of Mastermath, a joint effort of Dutch universities to deliver mathematics courses at the master level for everyone in the Netherlands. It is a completely theoretical course with no coding involved. Different from the other courses, it is a 4 months course instead of only 2 months.
Topics of the course include PAC-learning, Rademacher Complexity, Online Convex Optimization, and AdaBoost. It’s a good course with quite some work. The theoretical exercises are more difficult than in any AI master’s course in Amsterdam, but not more difficult than exercises in typical maths lectures.
Information-Theoretic Learning is a course by Peter Grünwald from the University of Leiden, who also coordinates the course on Machine Learning Theory. It is basically a course on the Minimum-Description Length Principle on which Peter Grünwald has written a book.
Topics covered in the course include Kolmogorov Complexity, Shannon's Coding theorem, Markov Models, Jeffrey's prior, MDL Prediction/Model Selection/Estimation, ...
I would say the course is pretty good, albeit the connection to modern Machine Learning is a bit lacking.
Of course, I only took a handful of the courses available. Other courses worth considering (not exhaustive, heavily biased towards my interest in more principled theoretical courses):