Technology

A hybrid approach to AI

At the moment, AI methods are developed mainly on the basis of data. In the training phase, no expert intervention takes place and the AI is optimised exclusively through data in a continuous process. This approach has two significant weaknesses: Enormous amounts of data are needed for training and validation, which are time-consuming and expensive to collect and process. In addition, data-based AI functions are often black box models whose decision making cannot always be directly reconstructed.

Previous research approaches to solving these problems have focused on optimising the data necessary for training and validating AI. KI Wissen approaches the challenge in a new way and explores how existing knowledge relevant to the traffic context – in the form of traffic rules, mathematical-physical facts and also social norms – can be integrated into AI systems. By using this hybrid approach, i.e. the linking of data-based procedures with knowledge-based methods, the project is completely redefining the basis for the training and validation of AI functions. It now includes not only data, but information, i.e. data and knowledge. The further development from data-based to information-based AI carried out in the project addresses central challenges on the way to autonomous driving: the generalisation of AI to phenomena with a small data basis, the increase in the stability of the trained AI to disturbances in the data, the data efficiency, the plausibility check and the validation of AI-supported functions as well as the increase in functional quality.

The project will create a comprehensive ecosystem for integrating knowledge into the training and validation of AI functions. To achieve this goal, the partners in KI Wissen will explore and develop a variety of methods for the integration of knowledge as well as approaches for the extraction and conformity of knowledge.

Types of knowledge

Knowledge relevant to the traffic context is identified, systematised and processed. Specifically, three types of knowledge are considered in the project: mathematical-physical knowledge as well as world and expert knowledge.

Mathematical-physical knowledge: This type of knowledge is unique in the way that there are no exceptions to it. An example are kinematic equations, which are determined by the physics of driving a vehicle. For example, cars cannot move sideways without also moving forward or backward at the same time.
Expert knowledge (or: normative knowledge) describes the desired state of the world, in particular through which actions intended goals can be achieved. In KI Wissen, the road traffic regulations are considered as an example of this type of knowledge. An essential property of normative knowledge is that the rules described can be violated and allow for exceptions.
World knowledge is knowledge that is accessible to every individual through general life experience. An example of world knowledge is the frequency of certain kinds of road users that can be expected in any situation. For example, cyclists are rarely seen on motorways while children playing in residential areas are fairly common.

In the project, a flexible, modular and universally applicable representation of knowledge is being developed for use in various training procedures and architectures. This representation will make it possible to integrate knowledge, to extract it in a comprehensible form from the output of a system as well as to use it for examining the conformity in the validation process.

Knowledge integration

In KI Wissen, various methods for integrating knowledge into AI functions are developed and investigated. This is followed by the integration of knowledge into training procedures of existing systems and the development of new architectures incorporating knowledge for applications in autonomous driving. The main objective is to reduce the amount of data, needed for training, and the computing power required, while at the same time increasing the generalisation capability and functional quality of AI functions.

Knowledge extraction

Various methods for extracting knowledge are being investigated, adapted and (further) developed so that they can be applied in the context of autonomous driving. The focus is on the following methods: Methods for extracting concepts from the model, measures for deriving new knowledge from extracted concepts, and models that directly provide structured output. On the one hand, the extraction of knowledge can be used to recognise new, acquired knowledge of an AI and make it usable - most likely in the context of world knowledge - on the other hand, the extracted knowledge can also be used to compare it with existing knowledge and thus contribute to the explainability and analysability of AI systems.

Knowledge conformity

Another focus in the project is the development of methods that examine the outputs of AI systems for their conformity in relation to existing knowledge. The goal is to increase the plausibility and improve the safeguarding of the AI functions of autonomous vehicles. The results of this check can be incorporated (offline) into the further training of the AI functions and improve them or (online) ensure the safety of the AI functions at runtime. In this way, safety-relevant incorrect decisions of an AI function are detected at an early stage.

Integration and demonstration

The functions, components and methods developed in the project are evaluated in the context of an overall system using three selected use cases. This way, the usability aspect is emphasised. In particular, sub-components will not only be evaluated individually, but also in the context of this overall system architecture with regard to their respective suitability.

Ecosystem for knowledge

The project aims to build a comprehensive ecosystem for knowledge as a basis for efficient training and the validation of AI components. The results of the project will

help to extend the scope of validity to phenomena for which there is little data,
increase the stability of the trained AI to disturbances in the data, and
increase the traceability of AI decisions and enable higher training efficiency.

KI Wissen thus makes a significant contribution to the successful implementation of AI functions in autonomous vehicles.