Automated and autonomous vehicles are predicted to be the future of transportation, but their core element of artificial intelligence (AI) has previously struggled to master changing traffic situations. Scientists from the University of Stuttgart are now conducting research into new methods of machine learning as part of the "AI Delta Learning" project (with Mercedes-Benz AG as consortium leader) in order to train the AI module to gain a better understanding of its surroundings. The aim is to transfer existing knowledge from familiar domains to changing or new traffic situations with minimal effort, and to develop autonomous vehicles through to series production.
Night and day, rain and shine, forests and fields: drivers need to adjust to very different surroundings, often within a short space of time. Previously though, it has been necessary to train an AI module for every individual situation to do what a human being can master intuitively. "Previous methods for training artificial intelligence have been too specialized and haven’t been able to give the module any real understanding of its surroundings", explain Prof. Bin Yang and Robert Marsden from the Institute of Signal Processing and System Theory (ISS), who coordinate the University of Stuttgart’s part of the project, which is carried out in the Department of Electrical Engineering and Information Technology. "This is why the neural networks which the AI is based on have difficulties with situations which haven’t been mapped in the training data." For example, a model which is trained to recognize a car during the day will find it much more difficult to recognize a vehicle at night.
Learning based on differencesEnormous training data sets have previously been required for AI to be able to function as a generalist - a costly and time-consuming endeavor. Researchers working as part of the AI Delta Learning project have therefore been looking for ways of developing AI modules for autonomous vehicles and transforming them in such a way that they can react reliably outside of predefined scenarios. The focus is on developing methods of transferring existing knowledge from familiar domains to new target domains. With the help of these methods, in future AI should only need to learn the differences - the deltas - between unknown target domains and existing knowledge. These deltas can be subdivided into six application scenarios. For example, this includes dealing with further developments in the field of vehicle sensors or with long-term trends related to transportation. Also included in the scope of the project are the consideration of short-term changes such as different times of day or weather conditions as well as extending the use of AI methods to other countries, to name but a few aspects.
What has already been learned should not be discarded when bridging the delta, but should be used as a basis to build on instead. "This is the only way that autonomous systems can reliably cover the full complexity of the world of transportation and keep pace with the ever-shorter innovation cycles and constant changes in mobility", emphasizes Dr. Mohsen Sefati (Mercedes Benz AG), coordinator and leader of the AI Delta Learning project.
The project focuses on three main areas for delta learning: transfer learning, didactics and automotive suitability. The state of technology is advanced in all three areas, ensuring that the next generation of AI algorithms is prepared for unrestricted use in autonomous vehicles.
Training with simulated dataAs part of the project, the ISS at the University of Stuttgart considers the question of how the model can be supported when learning a more general representation. Simulated data for camera images of driving scenes which were originally generated for the computer games and film industry are used for this purpose. Among other things, researchers examine an approach whereby the appearance of simulated images is transformed in such a way that they can no longer be differentiated from real images. Furthermore, they also work on methods of continuous learning, in which new knowledge is integrated into the model without old knowledge being forgotten.
About AI Delta LearningThe AI Delta Learning project is made up of 19 consortium partners from prestigious universities and research institutions, as well as automobile manufacturers, suppliers and technology providers. The "Autonomous and Networked Driving" flagship initiative initiated and developed by the VDA as part of the AI family has been awarded funding by the German Federal Ministry for Economic Affairs and Energy for a period of three years.
Prof. Bin Yang, University of Stuttgart, Institute of Signal Processing and System Theory, e-mail
Tel: +49 (0)711/685 67330
Share this page
" alt="<p> <span >If you click on this link a connection to Facebook will be established:<br/> </span> <a href=’#’ target=’_blank’ >Share this page on Facebook Regard our privacy notice.
" If you click on this link a connection to Twitter will be established: Share this page on Twitter Regard our privacy notice.
" alt="<p> <span >If you click on this link a connection to Twitter will be established:<br/> </span> <a href=’#’ target=’_blank’ >Share this page on Twitter Regard our privacy notice.
" If you click on this link a connection to LinkedIn will be established: Share this page on LinkedIn Regard our privacy notice.
" alt="<p> <span >If you click on this link a connection to LinkedIn will be established:<br/> </span> <a href=’#’ target=’_blank’ >Share this page on LinkedIn Regard our privacy notice.
" If you click on this link a connection to Xing will be established: Share this page on Xing Regard our privacy notice.
" alt="<p> <span >If you click on this link a connection to Xing will be established:<br/> </span> <a href=’#’ target=’_blank’ >Share this page on Xing Regard our privacy notice.
Share this page on E-Mail Regard our privacy notice.
" alt="<p> <span ><br/> </span> <a href=’#’ target=’_self’ >Share this page on E-Mail Regard our privacy notice.