AI Lectures: LMU mathematician expects further breakthroughs in AI research in the medium term
26 Oct 2021
In the first of LMU’s new AI Lectures, Prof. Gitta Kutyniok reported on spectacular successes but also elementary problems around artificial intelligence
Research with deep neural networks has given the development of AI a significant boost. At LMU’s first virtual AI Lecture, mathematician Prof. Gitta Kutyniok highlighted the success story of artificial intelligence. “Self-learning algorithms are already doing their work in many fields of science and technology and have long since made their way into our everyday lives in areas like speech recognition and image diagnostics,” said Kutyniok. “We owe this not only to the existence of enormous computational capacities and huge data reservoirs for training the algorithms but also to the deep neural networks that make deep learning possible.”
However, the artificial intelligence success story is also something of a paradox: “The training of self-learning algorithms is largely based on trial and error. And there is still no in-depth theoretical understanding of how AI actually arrives at its decisions,” explains Gitta Kutyniok.
We don’t always know exactly how AI makes decisions
The artificial neurons and their connections to multilayer networks are ultimately modeled on the brain. “A network like this learns from vast amounts of training data; it learns to systematize, to derive rules from that and to make decisions on that basis,” says Kutyniok.
But sometimes even minor modulations of the setup still lead to inexplicable errors today. Often, not only does the AI lack robustness but it is also impossible to follow how it makes its decisions—the knowledge of how the network “reasons” is missing. “Yet both of these aspects are crucially important, especially for applications in sensitive areas such as autonomous driving or medical diagnostics,” says Kutyniok. That said, Kutyniok expects to see decisive breakthroughs in the coming years, particularly in the mathematical foundations of AI and its reliability and explainability. Because more and more scientists are devoting themselves to those questions, the density of results is increasing. “Research is accelerating rapidly,” Kutyniok is confident.
The lecture and the subsequent discussion with Prof. Oliver Jahraus, LMU Vice President, and the audience is available on the LMU YouTube channel. Link:
More information and an interview with Prof. Gitta Kutyniok can be found here. For further dates for the AI Lectures and to register, clickhere.