To train AI models that perform well in real-world applications, you need high-quality data—and as much of it as possible. In principle, AI can be used in all industries, fields, and environments; people are experimenting with it, testing it, and sometimes discarding it. In some fields, such as medicine or disaster response, there are high hopes for AI, yet it is precisely in these areas where training data is often scarce—and, most importantly, where serious errors must be avoided. For many problems, there is simply too little training data available to develop reliable and robust AI systems. Data privacy and other access restrictions do not even play a decisive role here: Collecting and making data usable for AI training can be very time-consuming and expensive. Synthetic data is a key concept here. It is intended to enable the rapid and cost-effective generation of training datasets for AI where “real” data is lacking. But is it really that simple? How far has research in this field progressed today? And what needs to be considered when embarking on this path?
Sören Pirk is a professor of computer science at Christian-Albrechts-University in Kiel, where he heads the Visual Computing and Artificial Intelligence (VCAI) lab. In this Digitalgespräch, the expert on generative AI and synthetic data explains what synthetic data is, how it is created, and what it is used for. He explores exciting research questions and describes the biggest challenges in working with this type of data. Together with hosts Marlene Görger and Petra Gehring, Pirk discusses the limits of its applicability, what characterizes “good” datasets, how rules and best practices are evolving in this field—and how the community, now that synthetic data is receiving so much attention, is looking at data that isn’t synthetic.
Further Information:
To the Webseite of the Visual Computing and Artificial Intelligence (VCAI) lab: https://www.vcai-lab.org/
To the Webseite of the Project „WildfireTwins“: https://www.wildfiretwins.org/
