Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU) is considered a pioneer in the field of artificial intelligence. As early as 1975, it established the first chair for pattern recognition in Germany - a step that proved to be groundbreaking for research and teaching in computer science.

At a time when computers still weighed 25 kilograms, had a maximum of 64 kilobytes of memory and cost a fortune, scientists at FAU were already researching topics such as the automatic classification of images and machine speech recognition - the foundations of modern AI systems.

Today, FAU is one of the leading research institutions in the field of machine learning. The "Pattern Recognition Lab" trains more than 3,000 students every semester - a clear contrast to the early years when only a few interested people attended the courses.

"Today, we benefit from the foresight that our predecessors demonstrated 50 years ago," says University President Prof. Dr. Joachim Hornegger, who headed the Chair between 2005 and 2015. "What was basic research back then now forms the basis for applications that shape our everyday lives."

The current chair holder, Prof. Dr. Andreas Maier, is continuing this tradition and continues to see FAU as an important driving force in AI research and the training of young talent.

Prof. Hornegger, it is sometimes said that the Chair of Pattern Recognition was the first AI chair in Germany. Is that true?

Prof. Hornegger: Yes, absolutely. Pattern recognition is an important area of artificial intelligence - I always say: the part of AI that really works (laughs). In fact, FAU was the first university in Germany to dedicate a chair to this field. The first holder was Heinrich Niemann. For us at the time, he was the pope of pattern recognition, even if he would never call himself that - he was far too modest for that. But it is no exaggeration to say that he made history in AI research in this country.

It almost sounds like pattern recognition is old hat these days...

Prof. Maier: Not at all. Without pattern recognition, there would be no ChatGPT! The statistical methods that people like Heinrich Niemann helped to develop and that have been researched here for 50 years are to a certain extent the basis for the current AI boom. This is because large language models are nothing more than giant statistical inference machines: they look for patterns in sentences and draw conclusions about which word is likely to come next.

Prof. Hornegger: At the time, however, nobody could have imagined how quickly the field would develop. The data volumes and storage sizes available to us today were unimaginable. When I was a student, there were no USB sticks, but floppy disks; they could hold 1.4 megabytes of data. To store a one-second image sequence, you would have needed a few dozen of them.
At the time, Heinrich Niemann was working on the analysis of language and images: How can you reliably identify sounds, even if they come from different speakers? What is depicted in a photo and what features can an algorithm - we're talking about a classifier - use to recognize it? Something like this.

So the computer had to know, for example, that a table has four legs?

Prof. Hornegger: That's right. At the time, it was about explicit knowledge representation. What components does a car consist of, which features must be present, which of them can be concealed and how? These were some of the questions that the chair dealt with back then. Today, the algorithms learn this automatically, thanks to the huge flood of data available to them and the computing power they can use.

Prof. Maier: This explicit description - a table has four legs - has become less and less important. Someone in the speech recognition department at IBM is said to have once said: Whenever we fire a linguist, the error rate of our algorithms goes down. A linguist may know how a word sounds under ideal conditions. But in reality, it often sounds different. Instead, current methods are based on statistics, i.e. the knowledge that certain sounds often occur together. This makes the systems much more flexible.

Because a table can also have three legs - or perhaps two of the four legs are covered? And because all tables nevertheless have things in common - patterns that are not so easy to put into words?

Prof. Maier: That's right. The algorithms learn for themselves which features are important and in which combination they must occur for it to be likely to be a table.

Prof. Hornegger: At that time, there was a lack of both computing power and training data for such approaches. Instead, Mr. Niemann and I later tried to find the most efficient solutions for certain problems - such as detecting contours in a photo. And yet the idea of neural networks was already emerging at the time. And algorithms had already been developed that could be used to train such networks. However, these methods only achieved their breakthrough when modern, extremely fast graphics processors appeared. And when the Internet suddenly made masses of data available to feed the models.

Prof. Maier: In the meantime, even this information is no longer sufficient to significantly improve the AI processes. Today, we are therefore starting to train the algorithms with synthetic data. For example, we generate an image of a tumor on the computer, which we vary in a variety of ways. We then feed this to the artificial intelligence so that it can then better recognize such tumours in real radiology images.

Prof. Hornegger: However, there are also problems that an AI will not be able to solve. If we have a text in a completely unknown language, the processes cannot translate it. For example, we once had a project that dealt with the vocalizations of orcas. The AI recognizes patterns that occur again and again. However, without further information, it does not know what these patterns mean. This only works if there is also behavioral data. Then the algorithm may learn that orca mothers make these sounds when they are separated from their child.

Many people are afraid of developments in the field of artificial intelligence. Can you understand that?

Prof. Hornegger: I understand that. AI will change entire job profiles. Even computer science - GPT-5 can now write programs in just a few minutes that would have taken months in the past. But fear is a bad advisor. On the other hand, artificial intelligence will also create freedom that we will hopefully be able to use for more creative tasks. There have always been changes in the past.

Prof. Maier: I have the impression that we computer scientists are not heard enough in public on this issue. The researchers I know are all concerned with the question of what social implications the technology will have. At the same time, I have the impression that almost all of them are enthusiastic about the possibilities offered by AI. We must not forget that artificial intelligence can also achieve a great deal of good for humanity. These opportunities are often neglected in the discourse.