[Translate to English:]
KI Wissen

First Open Project Day of KI Wissen held

For a little over a year, the KI Wissen project has been working on how existing knowledge can be integrated into and extracted from AI systems. The goal of KI Wissen is to create hybrid AI systems, that integrate modern data-driven and traditional knowledge-based methods. KI Wissen is the youngest of the four projects of the KI Familie and thus also builds on the results of the other three projects – KI Absicherung, KI Delta Learning and KI Data Tooling. On February 22, the first results of KI Wissen were now presented at a first Open Project Day. The aim of the Open Project Day was to engage in an open exchange and to promote further cooperation with the other three KI Familie projects.

The all-day virtual event started with two welcoming remarks by Corina Apachite from Continental as the host and Ernst Stöckl-Pukall from the German Federal Ministry of Economics and Climate Action as the funding partner. Afterwards Ulrich Kreßel from Mercedes-Benz highlighted in his presentation the novel collaboration framework of “family projects” and the impact of the VDA Leitinitiative. Aligning research topics and resources for a fast track towards deployment is of particular importance for the German industry to gain leadership in the highly dynamic field of autonomous driving. Project coordinator Jörg Dietrich then gave a general introduction to the challenges, goals and methods of KI Wissen, followed by the four subproject leaders Ludger van Elst (standing in for Jörg Reichardt), Antje Loyal, Gerhard Schunk and Edmilson Freitas. After this introductory part, the new approaches developed in KI Wissen were presented in three forums, each dealing with a use case from the project: Pedestrian Detection under Occlusion, Complex Lane Change and Controlled Rule Exception. All three forums had the same structure: First the use case was explained in more detail, then five selected concepts were presented in short elevator pitches, followed by an open poll in which the audience could decide, which two approaches would be presented in more detail. This interactive format allowed the audience's interest to be specifically incorporated into the flow of the event, ensuring an insightful and interesting event for visitors. After the presentations, the concepts presented were discussed by a panel of experts, consisting of colleagues from the other KI Familie projects. This also generated important information and ideas for the presenters for their further work on the concepts and highlighted opportunities for increased collaboration between the projects.

In the first use case forum on “Pedestrian Detection Under Occlusion” the audience selected two pitches on the topic of concept extraction to be presented in detail. The expert discussion following the presentation already highlighted the collaboration potential with KI Absicherung, which has pedestrian detection as their central use case, and KI Data Tooling which studies corner cases, of which occlusion is a special case.

In the forum on “Complex Lane Change” the two selected talks dealt with integrating knowledge into the detection of traffic signs using information about the co-occurrence of certain signs and using knowledge about vehicle dynamics together with causal models to predict lane change maneuvers and plan them for the ego vehicle.

The final use case forum on “Controlled Rule Exceptions” was unique in the way that the other use cases have a traffic scenario prescribed but rule exceptions may be necessary or allowed in a large variety of traffic scenarios. Selected were two talks on “Rule conformant decision making” and “Continual Learning for Soft Knowledge Integration & Extraction”. While the former dealt with an approach to incorporate the hierarchical nature of traffic rules explicitly in the inference process, the latter presented an approach to represent the hierarchical nature of explicit and implicit traffic rules as a continual learning process.

The six presentations selected in total in the three use case forums spanned the entire range of the ways knowledge is used in KI Wissen: integration, extraction and conformity checks. Furthermore, the talks also covered the different kinds of knowledge that need to be considered in automated driving functions: mathematical and physical knowledge, normative and expert knowledge, and world knowledge. While all these aspects of the KI Wissen project were selected by the audience, a preference for AI concepts dealing with knowledge extraction and conformity emerged. This ties into the current strong interest in explainable AI models and their safety.

The Open Project Day was concluded by an industry panel moderated by Christoph Stiller with Robert Thiel from Continental, Ulrich Kreßel from Mercedes-Benz, Jörg Schrepfer from Valeo and Thorsten Bagdonat from Volkswagen discussing the relevance of the hybrid KI Wissen approach for industrialization. A particular benefit was seen with regard to the GDPR constraints and the potential to get by with less data. Nevertheless, the participants agreed that without comprehensive and high-quality data the way towards autonomous driving will be cumbersome.

Overall, the event consisted of numerous presentations and discussions that provided in-depth insights into the work of the first year of KI Wissen. The approximately 150 participants were thus given an overview of the project as such, the initial results and the further work.