Research Training AI Systems for Use in Dangerous Situations

Systems based on artificial intelligence must also function reliably in infrequent and dangerous situations. Researchers at Heidelberg University along with scientists from France and Japan are studying how to “train” AI systems for such circumstances. Prof. Dr Carsten Rother of the Interdisciplinary Center for Scientific Computing (IWR) is heading up the Heidelberg arm of the collaboration, which is being funded with approximately 300,000 euros by the German Research Foundation for a period of three years.

“AI-based systems will find their way into many areas of our life in the near future, coexisting with humans in dynamically changing environments,” states Prof. Rother. “In spite of considerable progress in the field of machine learning, these types of systems do not always perform reliably in all scenarios, especially in dangerous situations that don’t arise every day.” Given this, the researchers are planning to develop dedicated software and hardware in order to generate new data for machine learning based on images and video footage. This will allow researchers to specifically generate rare scenarios to test and improve their artificial intelligence applications.

Also participating in the joint project, entitled “Understanding and Creating Dynamic 3D Worlds towards Safer AI”, are the École des Ponts ParisTech and Japan’s Kyoto University. It is one of nine trilateral projects in the field of artificial intelligence being funded in the amount of approximately seven million euros by the Agence Nationale de la Recherche and the Japan Science and Technology Agency along with the German Research Foundation. Scientists from Germany, France and Japan are participating in each of these projects.