University of Tübingen, Master's Machine Learning

by Master Programs ML/AI6 min read14th Nov 2020No comments



The purpose of this post is to give an overview of the Machine Learning (ML) Master’s program at the University of Tübingen. It should function as guidance for people who are interested in studying ML and weigh the pros and cons.

This article is part of a series of articles on different European master's programs related to artificial intelligence and machine learning.

General Overview

The ML program in Tübingen has 120 credits, 30 of which are assigned for the thesis. The program has three mandatory courses, Deep Learning, Statistical Machine Learning, and Probabilistic Machine Learning. All other courses can be chosen more or less freely with some small restrictions, e.g. they have to be in the broad area of ML. The full range of lectures can be found in the module handbook, though not all of them exist yet. Since 2019 was the first year of the Master, I expect these gaps to be closed in the next two years. In the winter semester 20/21 there are already many new courses.

Regarding prerequisites: There are some specifications on the website but they can be a bit vague. According to the creator of the program, who also oversees admissions, absolutely necessary criteria are having a sufficient understanding of proof-based math (e.g. through a math or Computer Science (CS) Bachelor’s degree) and a basic understanding of algorithms and other CS concepts. To give you a prior probability for a successful application we can only look at the first iteration of the degree where aone-thirdround 150 people from all around the world applied, 60 were deemed to be sufficiently qualified and around 40-60 actually started the Master. The program probably could have handled more students but the creator decided that applicants need to pass a certain level of skill. This acceptance rate of one third does not seem very low but I expect it to get more competitive in the future. Last time the program was only announced three months before the application deadline and already 150 people applied. Since then the University of Tübingen/the MPI has had more exposure within the media and the official ML YouTube-channel hit 800 subscribers in its first week I would expect the program to become more competetive. An alternative in case of rejection is to apply for the CS Master’s and transfer to the ML Master’s later. However, if too many people use this loophole it might be closed or ML students might be prioritized in contested lectures.

I think the general self-understanding of the program is one of excellence, i.e. it wants to produce people who have a deep understanding of the current ML landscape. As far as I can tell Tübingen seems to put a lot of emphasis on the theoretical understanding of ML (all courses have practical exercises too) but it’s hard to judge without an explicit comparison. The second emphasis in Tübingen is the social component of ML. There are lots of seminars discussing the ethics of ML or the intersection between ML and other fields such as medicine. From my personal experience, I would estimate that around 75 percent of the lectures that I attended fulfill the idea of excellence, i.e. they teach a mixture between old but relevant and new material, require a lot of effort but yield great understanding. Unfortunately, I had some courses that were rather shallow, didn’t update their content even though new research was available, or were clearly too easy. Since most lectures are in their first iterations and five new professorships in the realm of ML have been filled in the last two years and the university is still hiring more I expect the average quality of the lectures to rise further.

The student population that I know so far (only the first generation, so small sample size) is roughly 50 percent German and 50 percent international but I expect them to become more international in the future. From my perspective they are on average rather high-performing, ambitious students confirming the self-understanding of excellence.

The core lectures of the course are Deep Learning, Probabilistic Machine Learning, and Statistical Machine Learning. If you want to peak into the lectures you can find them on YouTube. There are around 20 further ML related lectures including Mathematics of ML, Data Literacy, Time Series, Self-Driving Cars, Neural Data Analysis, and Efficient ML in Hardware - just to name a few (For a full list look at the module handbook). Additionally, you can choose between around 50 different general CS lectures to broaden your perspectives.

Since only 24 credit points of 120 are mandatory lectures the program allows for individual specializations. Currently, one can specialize in applied or theoretical ML but it is impossible to focus exclusively on, for example, Reinforcement Learning. Given the impressive amount of new ML-related professors and groups, I expect that specialization will be easier in the future.

Regarding Corona, die University of Tübingen has adapted quite well and all lectures are now online. If you want a sample of the average quality of lectures, I would recommend looking at the YouTube channel.

The grading scheme in Tübingen is similar to other programs in Germany, i.e. it is rather hard but possible to achieve the best grade of 1.0. It is also realistic that you fail a class if you have not prepared for it and there is little grade inflation compared to e.g. the US.

If you already want to publish at conferences during your Master’s program, most supervisors will support you if you are willing to put in the effort. My supervisor, for example, told me that the aim of my Master’s thesis was to submit it to ICML if the results were sufficiently good (I started with a CS master and switched to ML if you were wondering about the timeline). However, whether you want to go through this effort and try to publish is obviously up to you and your supervisor, I can only say that most potential senior researchers would be up for it and willing to support the effort.

Research Directions in Tübingen

The amount of ML research that is done in Tübingen is huge. There is Deep Learning, Probabilistic ML, Statistical ML, Computer Vision, Robotics, some Reinforcement Learning (RL), Self-Driving Cars, Robustness and Adversarial examples, some Natural Language Processing (NLP), Fairness, and Ethics in AI (from a technical and humanities perspective), ML in Climate Science, a very large Neuroscience section, Causality and much more. I think the fields that are currently a bit underrepresented are NLP, RL, and AI-safety. Some years ago, Tübingen didn’t have a large focus on Deep Learning but they have upgraded and adapted since then and I would expect them to be a global tier 2 when it comes to Deep Learning.

Some of the fields that Tuebingen is internationally known for include Probabilistic Numerics (Philipp Hennig), Empirical Inference with a focus on causality and Kernel methods (Bernhard Schölkopf), Robustness and Optimization (Matthias Hein), Self-driving cars and Computer Vision (Andreas Geiger), and the neuroscience groups (Bethge lab, Peter Dayan). If you are interested in the intersection between ML and neuroscience I would suggest doing the ML master. If your focus lies with the foundations of neuroscience there are other master programs in Tübingen that might be a better fit.

Personally, I first did a bit of everything for a year and then specialized in the overlap of probability theory and Deep Learning by working on Bayesian Neural Networks. I think for most subfields in ML it can be said that somebody in Tübingen is working on them and if you are interested in specializing very early I would recommend clicking through the links below.

If you want to check out who does ML research in Tübingen, have a look at the research groups, people’s page of the IMPRS, the website of the MPI-IS, and the ML in science cluster of excellence.

Options outside of University

Tübingen is right at the heart of the cyber-valley initiative. This essentially just means the province and local industry come together to boost the ML competence in the region. They fund new professorships, research groups, buildings, etc. In short, Tübingen spends a lot of money on ML. Being an Excellence Cluster is not just a label but comes with a 50 Million Euro grant over 7 years that started in 2019. Its aim is to attract global talent at the intersection of ML and other sciences (e.g. climate science, ML for social good or ML in medicine). The benefits of such a cluster are indirect but noticeable. Many of the cluster people offer seminars which means that you can discuss the implications of ML with domain experts (e.g. a philosopher or geologist) and gain new perspectives. Additionally, the cluster organizes workshops and small conferences that are usually free for students where you can broaden your perspectives.

If you are leaning more to the research side, you can try to become a student assistant, write your thesis or do research projects with the university or at the Max Planck Institute for Intelligent Systems (MPI-IS) and thereby have direct access to top researchers in their respective field (this is not exaggerated, just look at the latest news).

If you care more about industry experience there are also lots of options. You can do internships or collaborations with large companies like Bosch or IBM. Bosch and Amazon are both building an AI campus for 700 and 200 researchers respectively that should be finished in the next couple of years. Even though their buildings are not built yet, they already do collaborations with the university.

Some personal notes

Even though this sounds a lot like a promotional piece, I honestly think that Tübingen is the place to be for ML, at least in continental Europe. However, if you want to do research in the fields of NLP, RL, or AI safety other universities might be a better fit. Even though I am not sure if there are Master’s programs with a strong focus on AI safety in Europe or even globally.

Regarding the courses: I have taken most of the ones that are already available and I think the majority of lectures and seminars are good with some exceptions. To figure out which ones you should avoid, I would recommend asking more experienced students. The vast amount of different options definitely is a benefit, especially since they are likely to become even more in the future.

Tübingen as a town might not be for everyone. At the end of the day, it is not a large city but a town of 90k inhabitants (35k of which are students) that has fewer options (nightlife or food diversity) than a larger city could provide. However, the university provides a lot of options for physical activity and there are other ways to spend your free time. If you really want a “big city feeling” though, you will likely not find it in Tübingen.

From an Effective Altruism (EA) perspective, Tübingen is pretty nice. There is an active EA chapter, we are currently founding an AI-safety reading group, and there is a small LessWrong chapter. There are also many other university groups and NGOs, like a debating club or Global Marshall Plan, that might be interesting from an EA perspective.

Decision Guide

You should consider the ML Master’s program in Tübingen if you

  • Want to have a 2-year/120 ECTS tuition-free master
  • Want theoretical and practical courses in your program
  • Want to have the option to cooperate with industry (e.g. Amazon, Bosch) or academic (e.g. MPI-IS, Cluster of excellence) collaborations
  • Want the option to explore the intersection between ML and other sciences (e.g. ethics in AI or ML in Medicine)

You should not choose Tübingen if you

  • Want to have a big city feeling to your place of study
  • Want to focus primarily on topics of Natural Language Processing or Reinforcement Learning

If you have any further questions about the town or program, want to get advice on how to improve your chances of getting in, or just want to leave some feedback don’t hesitate to contact me via the channels listed on my blog. If you want to know more about the research I do you can find short summary posts on my blog.



New Comment