Vincent Fortuin is helping making AI safer and more reliable

Branco Weiss Fellow Vincent Fortuin, a principal investigator at Helmholtz AI and the Technical University of Munich, along with 24 researchers from around the globe have published a pioneering approach in the realm of artificial intelligence (AI). Their paper, “Bayesian Deep Learning is Needed in the Age of Large-Scale AI”, challenges the status quo and presents a vision for AI that is more reliable, trustworthy, and safe.

The authors raise critical points about the current focus in AI development: They argue that the race for accuracy using massive amounts of data overlooks crucial aspects such as the AI’s ability to judge its own reliability. This insight is especially vital in fields where data are precious and rare, like in groundbreaking scientific research, where making a wrong prediction can cost more than just inconvenience. Bayesian deep learning marries the predictive power of AI with the wisdom of centuries-old Bayesian statistics, a statistical approach that dates back to the 18th century. This blend allows AI to not just make predictions, but also to understand and communicate how certain it is about those predictions. The implications of the researchers’ findings are manifold, from revolutionizing the way we discover new medicines to enhancing the reliability of the digital assistants in our homes and phones. They envision a future where AI can not only provide answers but also indicate when it might be wrong, leading to safer and more dependable technology.

Dr. Fortuin and his colleagues acknowledge the journey ahead is filled with challenges, such as making these advanced AI models work quickly and efficiently at a large scale. However, they are optimistic, outlining potential solutions and inviting the global research community to join them in exploring this uncharted territory.

Read the paper in the proceedings of the International Conference on Machine Learning in Vienna

Read an interview with Dr. Vincent Fortuin on the Helmholtz Munich website