Summer Semester 2027 – fixed term of 3–6 months

Would you like to explore a fundamental research question at the intersection of representation learning, synthetic data, and industrial computer vision? In this thesis, you will investigate how to determine whether a synthetic dataset contains the information needed to train a model for use in real-world applications.

Common dataset comparison metrics primarily measure visual or statistical similarities. However, they do not necessarily capture whether the shared information is relevant to a specific task, such as object detection. Existing methods for assessing transferability also often assume that a suitable pretrained representation is already available. This is not always the case, particularly in specialized industrial applications.

The goal of this thesis is to learn and evaluate latent representations that allow synthetic and real-world datasets to be compared based on their task-relevant information.

YOUR TASKS:

• Investigate supervised and self-supervised approaches to learning task-aware representations from labeled synthetic data and unlabeled or partially labeled real-world data.
• Develop distance metrics in latent space to compare datasets and evaluate how well these metrics predict downstream model performance in real-world applications.
• Compare your approach with established methods, analyze its limitations, and implement a prototype for assessing synthetic datasets.
• Validate your findings experimentally and investigate the underlying assumptions and potential failure cases.
• Document your findings and derive recommendations for using and selectively improving synthetic datasets in industrial AI development.

The outcome of your thesis will be a validated methodology and a functional prototype for task-aware synthetic data assessment, together with a critical analysis of the underlying assumptions, limitations, and potential failure cases. Your findings could help improve the selection, reuse, and refinement of synthetic datasets.

YOUR PROFILE:

• Currently pursuing an advanced master’s degree in Computer Science, Artificial Intelligence, Data Science, Mathematics, Robotics, Physics, Engineering, or a related field
• Good knowledge of machine learning and deep learning, with practical experience in computer vision and neural networks
• Knowledge of representation learning, self-supervised learning, or synthetic data is an advantage
• Good Python programming skills and practical experience with PyTorch or a comparable deep learning framework
• Ability to understand scientific literature, formulate hypotheses, design controlled experiments, and critically evaluate results
• Ability to quickly grasp complex technical problems and work independently and systematically
• Ability to communicate findings clearly and collaborate with an interdisciplinary team
• Strong interest in representation learning, synthetic data, sim-to-real transfer, and applied industrial AI research
• Good written and spoken English skills

You don't meet every requirement yet? We still look forward to receiving your application. We value curiosity, a willingness to learn, and the ability to approach complex questions systematically. You will have the opportunity to develop any missing skills as part of your thesis.