Prof. Dr. Tobias Friedrich

Deep Learning for Combinatorial Optimization

MSc Seminar - Summer 2021


Over the past decade, deep learning has advanced a lot and proved to be useful for solving various problems. In this project seminar we research the use of deep learning approaches to solve combinatorial problems with some applications to computer vision. For example, in many computer vision problems, there is an underlying graph representation which allows for the development of deep learning algorithms to computer vision problems based on combinatorial graph optimizations.

Other combinatorial optimization problems focus on (but are not limited to) routing, solving NP-hard problems, optimizing deep neural networks, keypoints computation, studying the process of activation in the combinatorial setting, object detection, and optimizing deep learning frameworks that use combinatorial algorithms (such as nearest neighbours) as part of their computation.

In particular, we will work on selected topics in small groups. The goal is to develop novel applications of deep learning approaches. In the end, you will present your findings to each other as well as create a scientific report.

Outline of the Course

The main focus of the project seminar lies on active research. You will form small groups to work on the topic of your choice. For each topic, there will be a short introduction in the first few weeks. During this time you will get familiar with the underlying problem. After that, you will actively work out solutions for the problem. We will support you with weekly meetings where we discuss your ideas and approaches, the progress and further actions. While you will be guided by us, we encourage you to propose own ideas. In the end, each team will write up their findings in a scientific report and present these to the others. Summing up, the outline of the course looks roughly as follows.

  • First two weeks: Introduction to the topic and literature review.
  • Next eight weeks: Confrontation with the topic and active research on it.
  • Last two weeks: Writing your scientific report and presenting your results.

Requirements and Formalities

There are no formal requirements to participate in this course. However, we expect a certain familiarity with deep learning approaches. Furthermore, you should enjoy active and independent work on interesting topics.

Please note that there is a participant limitation. Besides registration to the Studienreferat, please also register to the moodle of the course.


The grade will be based to 80% on your scientific report and to 20% on the final presentation.


There will be an introductory meeting via Zoom in the first week of the lecture period (details will be given in the moodle of the course). There, we will discuss the topics, the teams, details and all the following dates.