NCPS > Theses & Projects

Theses & Projects

Below, you see a snapshot of some open thesis and project topics for bachelor's and master's students at TUHH (if you are looking for PhD positions, please check the TUHH recruitment portal and the news page on our website). Typically, we have some more topics; please talk, best after alecture or an exercise, directly to the team to find out more. Most of our topics are based on our courses (Distributed Systems, Internet of Things, TinyML, etc.). We expect a bachelor student to have taken at least one of our elective courses and a master's student to have also taken one (preferably multiple) courses with us. For each thesis, we list required background, including the course(s) taken with us and often also courses given by other institutes. Also, for master's students, it is a nice bonus to conduct a project with us before starting a thesis with us. Connecting both will ease your learning curve on the background, methods, and tools.

For each topic, we list a contact person. If you have taken courses with us and have the required background, please don't hesitate to contact them directly when a topic sounds interesting. Furthermore, you are welcome to contact Olaf or Marcus to have a chat with them about the different topics and how they match your interests and background, i.e., how they fit the courses you have taken. When you contact us, please name the course(s) you have taken with us, and please also include a list of courses you have taken (from TUNE) at TUHH and wherever else you studied.

Work atmosphere, workshops, presentations: We offer you a great work atmosphere, motivated advisors, and a coffee machine (or tea if you prefer). Next to frequent meetings and discussions with your supervisor(s), each term we also organize a series of workshops to guide you throughout the different phases of your thesis and project work, such as how to work scientifically, evaluation, writing, and presentation. Students doing their thesis and project with us are expected to join these workshops to gain the proper knowledge to conduct their work and shall also regularly join the kickoff and final presentations of other students at our institute to learn how to present their work and their results. Commonly, the workshops and presentations are once a month on Wednesday afternoon, commonly at 2pm.

Conducting a thesis with a company: This is an option, but not the default one. As a result, additional requirements apply (next to the ones stated above). We expect the topic to (a) have a strong research component like any other thesis, (b) be close to our teaching and research and the course(s) you have taken with us, and (c) at the company to be supervised by members of the R&D department. They shall have research and supervision experience, for example, from PhD studies. Please check this before contacting us, and when you contact us about doing a thesis with a company, please also let us know a bit about the topic and the company (next to listing your courses taken with us and including a list of courses you have taken from TUNE at TUHH and wherever else you studied, see above).

(Last update 14.10.2026)

Open Topics

Sparse Attention: An Evaluation of the Qwen and Minimax Approaches

Master's Thesis

Topic Description

This thesis will integrate two state-of-the-art sparse attention mechanisms, Qwen Sparse Attention (QSA) and Minimax Sparse Attention (MSA), into an existing large language model. The adapted model will then be benchmarked across multiple metrics to assess each approach’s practical advantages and limitations.

Requirements

You have a strong background in machine learning and a good understanding of transformers.
You passed modules Machine Learning I (or comparable) AND Advanced machine learning with good grades! Please attach your transcript of records (grades overview) when contacting us.

Contact: Oscar Kemme

How Many Connections Does Spatio-Temporal Prediction Really Need?

Master's Thesis

Background

Predicting how traffic flows through a city, how air quality develops, or how the weather evolves are all spatio-temporal (ST) prediction problems. A network of sensors measures a quantity over time, and the goal is to forecast its future values. Because neighbouring sensors influence each other, state-of-the-art models represent the sensor network as a graph and learn how information should flow between nodes. This is encoded in an adjacency matrix.

In many modern ST models, this adjacency matrix is not predefined but learned: it is computed as the product of two learned node embeddings (Wu et al. 2019). This closely resembles the query-key product of global attention, in which every sensor can attend to every other sensor.

The Problem

Global attention has a price. Its cost grows quadratically with the number of nodes, which becomes prohibitive for large-scale traffic datasets with several thousand sensors. Recent work addresses this in two main ways:

Linearising the softmax with a kernel in the attention formulation (Han et al. 2024).
Restricting the degrees of freedom in the adjacency matrix (Kong et al. 2025).

Both approaches achieve training time that scales linearly with the number of sensors. However, they leave a fundamental question unanswered: how much freedom does the adjacency matrix actually need?

A surprising hint comes from our own previous work (thor Straten et al. 2026): even a randomly initialised adjacency matrix performs remarkably well. This suggests that current models may be learning far more connectivity than they require.

Research Question

What is the minimum number of degrees of freedom (dof) an adjacency matrix needs for spatio-temporal prediction on large-scale traffic datasets, and, as a consequence, how few connections between sensors are sufficient to reach state-of-the-art performance?

Your Profile

Strong background in machine learning, and a good understanding of attention mechanisms and graphs.
Successful completion of the modules Machine Learning I (or comparable) and Advanced Machine Learning, with good grades.
Preferably, prior experience with spatio-temporal datasets, e.g. in traffic, air quality, or climate.

How to Apply

Please contact Dr. Yannick Wölker and attach your transcript of records (grade overview).

TU Hamburg

Institute for Networked Cyber-Physical Systems

Prof. Dr. Olaf Landsiedel

Harburger Schloßstraße 28

21079 Hamburg, Germany

E-mail : ncps@tuhh.de

Phone : +49 40 30601 3531