Technology Hub Machine Learning

In addition, with the new Machine Learning Technology Hub, for example, Fraunhofer is addressing current industrial challenges that enable the strategic integration of transparent and resilient artificial intelligence solutions into production, business and sales processes.

Machine Learning (ML) is a key technology for cognitive systems, intelligent products and digital assistants and plays a decisive role in the digital transformation of our economy and society. In recent years, considerable progress has been made, not least because of the plentiful availability on the internet of data annotated with explanations and add-ons. However, in industrial practice, existing ML processes often reach their limits. This because data of relevance to industry are mostly neither freely available nor extensively annotated, and companies can only begin using cognitive systems to a limited extent. What is needed are processes which integrate expert knowledge where data and artificial intelligence alone are not sufficient.

The leading edge with "Informed Machine Learning"

This is where the Fraunhofer Machine Learning Hub comes in: The research goal is to develop a new generation of reliable ML processes that use compositional approaches to systematically integrate structural and procedural expert knowledge into statistical training processes, making these operate robustly and understandably, even with little training data. Fraunhofer-shaped "Informed Machine Learning" broadens the application and usage range of Machine Learning enormously. Companies to date have had no insight into the "Black Box," but Informed Machine Learning makes it possible to transparently understand the decisions made by leaning systems and to intervene at the right points – an important prerequisite for assessing quality, reliability and risks, as well as for combining learning results with existing expertise and models. This creates new possibilities for increasing quality and efficiency as well as for the development of new products, services and future-oriented business models.

At the Fraunhofer Machine Learning Hub, the Fraunhofer institutes IAIS (management), IOSB, ITWM and SCAI consolidate their many years of scientific expertise and their expertise from the direct transfer of pioneering ML research into industry. What results from the close cooperation with IoT-COMMS und Data Spaces Technology Hubs of the Fraunhofer CCIT is a seamless chain from the recording of data, their secure storage and processing right up to their intelligent use or application.

What special role do the topics Machine Learning and Artificial Intelligence play in the cognitive internet?

Today, artificial intelligence and cognitive systems must be so efficient that they are no longer programmable but have to learn from data. With Machine Learning, we use the data and knowledge available for flexible and self-improving, intelligent systems.

What is the position of industry in this field, and how can Fraunhofer support enterprises?

The importance of Machine Learning is clearly recognized in the economy, but in many sectors on the one hand the appropriate data, and on the other hand the necessary specialists are lacking. We help enterprises assemble and create the right combinations of data, processes and business models. In our data science instruction courses at Fraunhofer, on site at the company or in interactive-online courses, we train the next generation of data scientists.

What are your long-term goals with the ML Hub and its implications for the topic of "Cognitive Internet Technologies?"

With the Machine Learning Technology Hub, we want to establish "Informed Machine Learning" on a broad footing, by which we mean approaches that not only learn from data but can also utilize existing expert knowledge and models, as are often available in the economy, to improve performance. At the same time, the transparency and reliability of the results are assured, thereby bringing about the preconditions for trustworthy cognitive internet technologies.

• "Informed Learning – Hybrid Learning" for the integration of data- and knowledge-driven methods and for the use of "a priori" knowledge
• "Informed Learning – Simulation-based Learning" for the automatic generation of plausible training data for "thin data" scenarios
• "Resource-aware Learning" for approaches for the optimal utilization of existing infrastructure, for example High Performance Computing (HPC), "IoT Edge" solutions and quantum computing