Combined use of knowledge and data: Hybrid AI for greater security and transparency

Large AI models that generate texts or images based on huge amounts of data and machine learning continue to cause a stir. However, data-based AI approaches also reveal limitations, for example due to their resource-intensive training or the lack of transparency of the AI system. Hybrid AI offers a promising alternative for sensitive areas of application such as medicine: it combines data-driven AI approaches with human knowledge and promises energy-efficient, robust and explainable AI systems. In the second issue of AI at a Glance, Plattform Lernende Systeme explains what hybrid AI is and the potential and challenges associated with it.

Data-driven approaches have dominated AI development for some years now. If these reach their limits, a combination with knowledge-based AI processes can offer solutions. This is known as hybrid AI. "While machine learning approaches are purely data-driven, hybrid AI allows existing knowledge to be taken into account during learning. In principle, this corresponds to human learning processes: what you already know, you don't have to keep relearning. Hybrid AI leads to greater data economy and more robust models," says Ute Schmid, Professor of Applied Computer Science at the University of Bamberg and head of the Technological Enablers and Data Science working group of Plattform Lernende Systeme.

Combining the advantages of data- and knowledge-based AI processes

Data-driven AI systems can perform not only specific tasks, but also a wide variety of tasks, and are extremely powerful. However, they also have weaknesses: Distortions in data sets, for example, can be reproduced and lead to discriminatory decisions. It is difficult to understand how the results come about. Many problems cannot be solved by ever larger models, data volumes or computing capacities. Added to this is the high resource consumption of purely data-driven AI approaches.

In the classic knowledge-based AI approach, human knowledge is systematized in such a way that computers can process it. These AI methods deliver explainable, comprehensible results and are best suited to well-defined problems that do not change significantly over time. However, knowledge-based AI systems are difficult to scale and their performance is inadequate when processing large empirical (real-time) data streams. This occurs, for example, when autonomous vehicles or AI-supported robots perceive their environment via sensors and act and learn on the basis of this information.

Data-saving, robust and explainable

Hybrid AI systems can combine data and knowledge in different ways. A knowledge-based system can be embedded in a data-driven component - or vice versa. Alternatively, knowledge-based and data-based systems can inform each other: The results of one system provide input for the other. Ideally, neural networks retain their trainability and effectiveness, while knowledge-based components contribute explainability and the simple integration of human knowledge. Hybrid AI is used in the diagnosis of tumors, to predict tipping points in climate research or in robotics for the rehabilitation of stroke patients. "Humans are capable of learning. However, they usually act on the basis of learned knowledge and reflexes. Accordingly, we should take an 'integrative' approach to the development of robotic systems by embedding hybrid AI approaches in rule-based basic functions and function-defining structures. This saves resources and increases the safety of the systems," says Elsa Kirchner, Professor of Medical Technology Systems at the University of Duisburg-Essen and head of the Learning Robotic Systems working group of Plattform Lernende Systeme.

In addition to the many advantages, the use of hybrid AI is also associated with challenges. For example, there is a lack of guidelines for the implementation and benchmarks in the research and development of hybrid systems. Their creation is complex and requires interdisciplinary collaboration between researchers and domain experts.

About the format: AI at a Glance

AI at a Glance provides a concise, well-founded and scientifically based overview of current developments in the field of Artificial Intelligence and highlights potentials, risks and open questions. The analyses are produced with the support of experts from Plattform Lernende Systeme and are published by the office. The second issue of the series focuses on "Hybrid AI. Combined use of knowledge and data". It is available to download free of charge.

Further information:

Linda Treugut / Birgit Obermeier
Press and Public Relations

Lernende Systeme – Germany's Platform for Artificial Intelligence
Managing Office | c/o acatech
Karolinenplatz 4 | D - 80333 Munich

T.: +49 89/52 03 09-54 /-51
M.: +49 172/144 58-47 /-39
presse@plattform-lernende-systeme.de