KI-Wissen

KI-Wissen

Development of methods for the integration of knowledge into machine learning

KI-Wissen

KI-Wissen (english: AI-Knowledge) aims at extending conventional data-driven Machine Learning approaches via the integration of various types of knowledge. In this way, the project contributes to an increase in functional quality, data efficiency, plausibility and safeguarding of AI-based functions for autonomous driving.

Project description

KI-Wissen is one out of four projects within the AI Family of “VDA Leitinitiative autonomes und vernetztes Fahren (VDA-LI)”. The projects of the AI Family address different challenges for the development of artificial intelligence and Machine Learning methods, solutions and tools in the automotive domain.
Conventional Machine Learning approaches, especially deep learning-based models, often achieve excellent performance in various tasks like object detection or prediction. However, they usually require a large set of training samples in order to learn the desired functionality. In many real-world scenarios, the existence of representative samples is very limited - not least for ethical reasons, as in various critical traffic situations. However, controlled behavior in these situations is an essential prerequisite for trustworthy and reliable autonomous systems. The project, therefore, follows the approach of using different knowledge sources and representations in order to make data-driven AI models more reliable. Besides the integration of additional knowledge, the extraction of behavioral patterns is an essential component to make the employed systems more transparent.

Research contribution

The project is divided into the sub-topics of knowledge integration, knowledge extraction and knowledge conformity. While in the first area methods are developed that take into account relevant knowledge within the traffic context, the focus in the second area is on approaches that can recognize behavioral patterns in the models and thus make them interpretable. In the subsequent third area the previously extracted concepts are compared against existing knowledge in order to verify the plausibility of the outcome.

As one of six research partners fortiss focuses on the extension of training methods on one hand. The goal is to integrate knowledge directly into the learning process such that the learned models generalize to different scenarios and thus show a higher invariance to changes in the input stream. Furthermore, fortiss explores methods that disentangle the information contained in the input data, such as shape, color and orientation, for better interpretability. In the context of checking the conformity to intended behavior, fortiss investigates the estimation of uncertainties in the output of models.

The planned research activities on knowledge augmented Machine Learning thus represent a significant contribution towards more robust and reliable AI systems. 

Funding

 

KI-Wissen is funded by the Federal Ministry of Economy Affairs and Climate Action (BMWK) within the program "New Vehicle and System Technologies"(„Neue Fahrzeug- und Systemtechnologien“).

Funding management (Projektträger): TÜV Rheinland Consulting GmbH

Project duration

01.01.2021 - 31.03.2024

Dr. Julian Wörmann

Your contact

Dr. Julian Wörmann

+49 89 3603522 438
woermann@fortiss.org

More information