BMBF Junior Research Group UnrEAL

19.10.2022|14:42 Uhr

Making the use of artificial intelligence safer: Funding for scientists from Wuppertal Artificial intelligence (AI) is used in many areas of our daily lives: when shopping online, it suggests products to us, the voice assistants on our smartphones are based on AI and the digital translation aids we use on vacation are also based on AI. In order to fully exploit the potential of artificial intelligence (AI) in safety-relevant applications such as automated driving, it is necessary to make them even safer - in other words, to learn to better assess possible uncertainties of AI. This is the starting point for a new research project led by Dr. Matthias Rottmann from the Stochastics working group at the University of Wuppertal.

Artificial intelligence is a key technology in many applications. For example, it is used to perceive the environment for robots or vehicles. The AI uses a camera image to make a prediction of where objects are located in a given scene. However, artificial intelligence makes mistakes in the process. It overlooks objects or makes false predictions. Its use therefore often requires a higher level of predictive reliability. Particularly in safety-relevant applications such as automated driving or medical image processing, a better assessment of the uncertainty of AI is an important factor that makes its use possible in the first place. 
“For more complex tasks, AI generally does not provide any information on the uncertainty of the predictions. In some use cases, it has also been shown that an AI makes false predictions with a supposedly high degree of certainty, for example when it encounters unknown objects. However, the evaluation of false predictions made by an AI can also be falsified by errors in the database. AI models are therefore dependent on a lot of high-quality data - this leads to high costs in data generation,” says Dr. Rottmann, summarizing the initial situation. This is where the “UnrEAL” (Uncertainty Quantification and Efficient Annotation Processes for Deep Learning) project comes in. The researchers are developing methods to recognize and improve the uncertainty of AI and ultimately to make data generation more efficient. To this end, they will first develop methods for recognizing uncertainties, detecting errors in the database and learning with little human feedback. 

In future, the results should help to build databases for AI that are both more cost-effective and of higher quality.
In order to strengthen young scientists in AI research, the Federal Ministry of Education and Research (BMBF) and the German Research Foundation (DFG) are funding 56 junior research groups across Germany. The junior research groups focus on novel and innovative AI topics. The Wuppertal scientists will receive funding of around 780,000 euros for their project over the next three years.