Combine the advantages and eliminate the disadvantages - evolutionarily, we humans have learned how to create knowledge and shape progress. With fire and water to the steam engine, from the discovery of electronic conductivity and magnetic force to electric motors. From a network of four computers to the Aparnet. Or from the Hebb's learning rule and Backpropagationon the renaissance of artificial Neural networks.
Today, AI-based processes are paving the way for autonomous driving. In this context, the AI Knowledge research project funded by the BMWI is investigating and developing methods for integrating existing knowledge into the data-driven AI functions of autonomous vehicles.
After we reported on the receipt of the funding in March, in this article we explain in more detail what approach underlies the research project and how exactly challenges of autonomous driving are solved.
AI knowledge: a hybrid approach for novel AI models in autonomous driving
The most widely used AI methods are based on large amounts of training data. In the training phase, expert intervention is dispensed with and the KI optimised exclusively through data in a continuous process. This is because the collection and processing of data is very time-consuming and ultimately very expensive. In addition, data-based AI models have a black-box character whose decision-making cannot always be directly understood. To solve these problems, previous research approaches have concentrated on optimising the data required for training.
The AI Knowledge research project is taking a new approach to eliminating these weaknesses: A hybrid approach is used by linking data-based procedures with knowledge-based methods.
Research is therefore being conducted into how known knowledge relevant to the traffic context can be integrated into AI systems. This will define a completely new training and validation basis for AI models of autonomous driving.
Project structure of the research initiative
The project structure of the research initiative is divided into four modules and focuses on the following core innovations:
Relevant knowledge for the traffic context is identified, systematised and processed. This knowledge base in the traffic context is differentiated according to three types: mathematical-physical conditions, social norms and world knowledge in the road traffic context.
Various methods for extracting knowledge are being investigated, adapted and further developed so that they can be usefully applied to the use case of autonomous driving. These include, for example, concept extraction from models or models with a direct structured output. Extraction thus serves to recognise and use newly acquired knowledge of an AI and to compare it with existing knowledge.
This involves the development of methods that check the outputs of AI systems for their conformity in relation to existing knowledge and thus support the plausibilisation and validation of AI functions of autonomous vehicles.
Integration and demonstration
The functions, components and methods developed in the project are defined in three use cases and demonstrated in a driving simulator for autonomous driving.
Thus, a comprehensive ecosystem is emerging that fundamentally addresses the challenges of autonomous driving systems through dedicated research and development: a small data base, increasing the stability of trained AI to data perturbations, data efficiency, plausibility and validation of AI-supported functions, and increasing functional goodness.
[at] is technology provider of the project community
AI Knowledge is a project of the AI family, which includes AI Delta Learning, AI Safeguarding and AI Data Tooling. The research project initiated by the VDA flagship initiative "autonomous and connected driving" is funded by the Federal Ministry for Economic Affairs and Energy for a period of 36 months and relies on 16 highly specialised project partners from Germany. [at] supports the innovative project environment AI Knowledge as a technology expert.
Based on auto-encoders, our specialists try to develop a more differentiated picture of black-box models and to make decisions made locally explainable. Specifically, on the basis of image and/or sensor data from the autonomous vehicle and with the help of modern text classification methods, we will attempt to divide driving situations into dangerous, precarious and normal situations.
For this, existing knowledge about data before and/or during model training is extracted from the data set itself. The result is a conformity assessment by measuring the similarity between prediction and extracted knowledge. This in turn leads to the fact that these results can be used to trigger conformance warnings, training data selection, model improvement.