Trustworthy AI
Available theses (bachelor thesis, master thesis), research and project modules as well as tutor and HiWi jobs can be found on this website or on the black board of the professorship ( across from room 3003-N and next to room 3079-N).
We will be happy to present our theses to interested students on request. The topics are primarily from our research areas or our national and international projects. However, students are also welcome to contribute their own ideas and approaches. If you are interested, please get in touch with the respective contact person.
THESES & RESEARCH/PROJECT MODULES
Research Topic „Certifiable AI in Medicine“
Motivation
It is thanks to remarkable results that AI is increasingly capturing the spotlight and making its way into various domains, including medicine. However, due to the "black box" nature of deep neural networks, there is still much research to be conducted before intelligent systems can become the norm and we can more precisely guarantee that these intelligent algorithms "act" in accordance with our European values, as stated in the EU AI Act. How can we now create methods and tools from a technical point of view that support the certification process in the best possible way?
Problem Statement
The main objective is to "make auditing simple" and provide specific guidance for domain-specific implementation of certifiable AI in medicine, considering the needs of different stakeholders and the consequences of specific design decisions. With a view towards (partial) automation, the steps are aligned with the entire lifecycle, following a process-oriented approach similar to ISO standards.
Contact
Research Topic „Trustworthy AI in Medicine“
Motivation
As widely known, intelligent systems being developed today are not inherently flawed but can exhibit "illusion" or even discrimination, which necessitates their application to be done consciously. Among the most well-known challenges are bias, robustness, and the trained model's generalization capability. However, transparent decision-making is a prerequisite for the trustworthy integration of AI-based applications in their respective contexts, particularly in safety-critical domains such as medicine. It should be noted, that among others, the utilized data, their distribution, and structure significantly contribute to gaining a deeper understanding of the model's behavior.
Problem Statement
This encompasses all research approaches that contribute in some way to improving the understanding of model behavior. These approaches can target all stages of the machine learning pipeline, ranging from data collection to model development and deployment, as well as considerations for long-term monitoring and system decommissioning. The overall objective is to enhance transparency and ensure a comprehensive understanding of how the model functions throughout its entire lifecycle.
Contact
Topic „Active Learning“
Motivation
Deep learning projects usually require a large set of annotated data. In the medical field, these annotations usually have to be created manually, which is very expensive. Within the LIFEDATA research project a framework is being developed which should reduce the manual effort by using Active Learning and Semi Supervised Learning.
Scope
The research project takes place in cooperation with Corpuls, allowing various tasks to be worked on at the university or in the company. For details please contact us.
Contact
Research Topic „Adaptive Learning for Industrial Processes“
Motivation
Within the transformation strategies towards Industry 4.0, the encapsulation of automation through Neural Networks is indispensable. Increasingly, highly adaptable and data-intensive architectures are used which engage in an interactive interplay with domain experts to achieve continual optimization. Through the dynamic models, new data and concepts can be implemented in real-time without restricting the productive process. Adaptive learning can be understood as a collective term for dynamic learning methods which in turn contain the widespread methods of active- and online learning. In the field of industrial process optimization, adaptive learning can contribute to involving human domain experts in the learning process in an interactive and effective way.
Scope
This research topic supports the current projects of the AI Production Network. Current topics mainly focus on the integration of human in the AI optimation process in automated and non-destructive quality assurance. For details please contact us.
Contact