Theses & Projects

Theses & Projects

We are continuously looking for students to work on student research projects and theses. Here, the topics cover the entire spectrum of research areas within the projects currently being dealt with by the staff. Ideas for own topics and concrete tasks are also welcome (if they go hand in hand with our research interests).

Due to the constantly changing topics and tasks proposed by the supervisors, only a few selected topics are mentioned here. 

Thesis Topic

We are looking for students who are interested in working on research in AutoML. Depending on their interests and goals, we select a topic best suited to them. In their thesis work, they will be focusing on state-of-the-art research within AI. The following topics are examples of how such a thesis might look.

Sample topics

  • Auto-PyTorch for X [BSc + MSc]

    Extending, applying and refining our AutoDL tool Auto-PyTorch for new domains like Image Classification, Segmentation, Video, Outlier Detection or NLP. We recommend a strong background in Machine Learning and your chosen application for this thesis.

    Contact: Difan Deng

  • Creating a new DAC benchmark [BSc + MSc]

    Modelling, implementing and evaluating DAC for any target algorithm. We recommend a strong background in RL, basic knowledge of DAC as well as of the target domain of your choice to be able to succeed in this topic. Possible target domains include, Machine Learning or Reinforcement Learning, MIPS or SAT solvers and Evolutionary Algorithms.

    Contact: Theresa Eimer

  • Important feature selection with fanova [BSc + MSc]

    As a Gaussian Process has poor scalability w.r.t the number of dimensions, usually it is very hard to fit it to a very high dimensional space. Not all the hyperparameters in an HPO problem are equally important: sometimes only a subset of them does influence the final performance of a HPO problem. As a popular hyperparameter importance analysis tool, fanova fits a random forest to capture the relationship between hyperparameters and performances. Similarly, BOinG also reduces the number of points to fit GP with the help of a Random Forest. Hence it would be interesting to combine fanova and BOinG to see if fanova could help to reduce the number of dimensions and hence make it applicable for a GP model. We recommend a strong background in AutoML for this thesis.

    Contact: Difan Deng

  • Meta-Policy Gradients in contextual RL [MSc]

    As hyperparameter tuning in the contextual RL setting has proven hard, using Meta-Policy gradients in the contextual setting for adapting all the hyperparameters in a single lifetime could be an alternative to established solutions. One way to start here would be to extend the work done on self-Tuning Actor-Critic (STAC) to the contextual setting by conditioning the state and action embeddings on a learnable context parameter (like the standard-deviation of the context generating distribution) and then train an agent to learn this set of parameters while interacting with the environment using Meta-Policy gradients. We recommend a strong background in Reinforcement Learning as well as prior knowledge of AutoML for this thesis.

    Contact: Aditya Mohan

  • Multi-fidelity as Meta Learning Problem [MSc]

    Using F-PACOH on multiple fidelities could be a way to transfer the inductive bias from one findelity to another and allow to evaluate on different fidelities during the estimation process. Combining this with Multi-information sources approaches (predecessors to MF) could help in also selecting relevant fidelities. Scheduling & evaluation of the validitiy are of the essence here. We recommend a strong background in AutoML for this thesis.

    Contact: Tim Ruhrkopf

Interested in working with us?

In order to get a better undstanding of how we supervise and grade thesis, refer to this page.

To do a thesis with you, we need the following information from you (though not every item may apply to every topic):

  1. Topic or areas of interest
  2. What prior knowledge do you have? Which relevant courses did you take?
  3. A rating for yourself in these areas (from -- to ++):
    • Coding in Python
    • PyTorch
    • Ability to implement a Deep Learning paper
    • Ability to implement a Reinforcement Learning paper 
    • Ability to understand and execute someone else's codebase 

This information should be sent via e-mail to the contact person given in the topic closest to your interest. The address can be found on that person's personal page.