Learning to Understand Artificial Intelligence Learning to Understand Artificial Intelligence

In the beginning was data, and the data was processed. “Every machine learning and AI project collapses without well-structured or annotated data,” says Moritz Wolter from the BnTrAInee project team. “This is the only way we can train AI to recognize patterns and thereby be of use.”

The implications of this situation can be demonstrated with the historical newspaper analysis project conducted by Privatdozent Dr. Felix Selgert, from the Department of History. “Our aim is to establish the nature of the economic situation in many areas. To do this, we study job advertisements in the Kölnische Zeitung,” says Dr. Selgert. However, unlike contemporary newspapers, those printed in the late nineteenth and early twentieth centuries were set in a comparatively unstructured fashion. Different font sizes and line spacing and unexpected column breaks make them difficult to read for the uninitiated. Computers always fail to make sense of these newspapers at the first attempt. “Like us humans, computers need practice,” says Wolter. “Without the correct data, the systems are unable to recognize patterns and develop their understanding.” This is important; otherwise researchers can quickly fall into the AI trap. This means that humans need to lay the groundwork, correct the data and point out errors and sources of error.

Not only the researchers looking to use AI in their work, but the computer science students assisting them also benefit from involvement in this program, as it provides them with the opportunity to complete the practical programming work required as part of their degree program. At BN-Trainee, they learn about the challenges experienced by and questions posed in a range of academic disciplines. This is not only a challenge, but brings great profit for both sides. They are required to work together in the development of a shared understanding of the challenges involved in their project.

“Computer scientists working with AI start off with an interest in the data, but understanding very specific research questions and the requirements that they place on the AI are also important for the analysis,” says Wolter. This leads them to develop new models or adapt existing ones, thereby benefiting subsequent teams. “Once a solution has been developed, it is comparatively easy to apply it to other challenges,” says Elena Trunz, a computer scientist from the project team. “Some methods drawn from machine learning are incredibly versatile,” she says. 

Dr. Barbara Wichtmann from the Faculty of Medicine uses artificial intelligence to evaluate MRI scans of prostate cancer. She uses the same model deployed by Professor Matthias Lang from Digital Humanities to identify various archaeological structures—charcoal pits, bomb craters and burial mounds—on aerial photographs. Another project realized within the scope of BnTrAInee uses AI to analyze the strengths of medicines, their effectiveness and dispensing practices. 

Inspiring students to work with machine learning and AI

The development of collaborative working practices is only one of three objectives of BnTrAInee. The project seeks to inspire computer science students not only to work with AI, but collaborate with researchers from other disciplines. “Computer science students are not naturally focused on the field of AI. Indeed, not everyone can deal with AI and machine learning after their training in Bonn,” says Elena Trunz from the project team. That is why we seek to foster an interest in this area and its wide range of possible applications.” 

Making experts fit for AI

The third aim of the program is to impart knowledge about AI and machine learning to researchers from a range of disciplines with no specialist understanding of it. One approach taken to this end is learning by doing. Experienced researchers from the fields of medicine, neurobiology, history or linguistics are given a three-week intensive course in the fundamentals of math, they learn procedures and the programming language Python. Working together, they solve the tasks set and discuss and analyze the best-practice approaches adopted by other participants.

A central advantage with this approach is the presence of a computer scientist. Indeed, as Elena Trunz tells us, “this is also the difference to a normal course, in which participants are often left alone to complete tasks. Indeed, the courses are extremely well received. By developing an understanding of programming, machine learning and AI functions, participants in the program are then able to run projects and data themselves on high-performance computers such as ‘Marvin’, the University of Bonn’s new mainframe computer.”

Professor Reinhard Klein sees considerable potential in the project that he manages. “I am convinced that we can roll out the benefits of AI across the University. The sort of cooperative approach that we pursue means that we can also help smaller research projects on a modest budget to perform data analyses in a cost-efficient fashion, using our project funds.”