By Aylin Krieger
16 Nov 2026
The Peer-to-Peer Network on Artificial Intelligence is a central component of the UBRA Lead Project Artificial Intelligence. It serves as a strategic platform to connect researchers across member institutions, bridging institutional boundaries to leverage synergies between theoretical research and practical application. Ultimately, the network aims to strengthen Bremen's position as a leading location for responsible AI research.

On 6 November 2026, Prof. Dr. Martin Mundt, Head of the Open World Lifelong Machine Learning Lab (OWL-ML Lab) at the University of Bremen, delivered the central presentation. As a member of the university's high-profile area "Minds, Media, Machines", the lab's mission is to develop adaptive AI systems that continue learning throughout their lifecycles.
Prof. Dr. Mundt's team addresses a fundamental weakness in current models: most AI systems remain "static" and lack adaptability. In conventional continual learning, a model's output is influenced not only by the data it was trained on but also by the order in which it encountered that data. This dependency often leads to biases in the AI's responses later on.
The OWL-ML Lab aims to make learning processes more adaptive by developing modular AI models that can be continuously extended and updated. Prof. Dr. Mundt emphasised that Lifelong ML is primarily an "algorithms problem". It requires new mathematical models and optimisation methods that enable AI to continuously integrate knowledge without forgetting what has already been learned. A central technical approach involves differentiating within neural networks between fast adaptation to current observations and slow consolidation of knowledge into model parameters. The lab advocates for models that are modularly updatable, moving away from static approaches towards sustainable updates.

A critical challenge in this development is the phenomenon of "shortcuts". AI models tend to learn superficial statistical correlations that yield correct results in the short term but lack robustness. The OWL-ML Lab investigates these mechanisms in both machines and human learners to develop algorithms that are more inclusive, fairer, and more sustainable than previous AI models.
The societal relevance of this approach was illustrated through two concrete application areas:
- Language Change: Words shift in meaning over decades and centuries. Current AI models often exhibit a "recency bias", fixating on recent data and failing to correctly interpret historical texts.
- Deepfake Detection: As generators for fake images evolve, detectors must also continue learning. The OWL-ML Lab's research highlights a critical risk: improving performance through continual learning can compromise system fairness, potentially leading to discrimination based on factors such as skin tone. This underscores the necessity of optimising AI algorithms specifically for social justice.
The presentation was followed by an engaged discussion in which participants applied the OWL-ML Lab's approaches to their own research fields. The Peer-to-Peer format provides exactly this kind of framework for collaboratively exploring complex topics. It fosters networking and open exchange within the Alliance, laying the groundwork for new ideas, joint research approaches, and valuable connections between institutions.
About the Peer-to-Peer Network
The Peer-to-Peer Network on Artificial Intelligence meets on the first Friday of each month from 12:00 to 1:30 p.m. The sessions are open to researchers from all U Bremen Research Alliance member institutions. If you are interested, please contact the Coordinator, Dr. Andrea Franke. More information about the Peer-to-Peer Network.
Publications:
- Paper: Busch, Kamath, Mitchell, Stammer, Kersting, Mundt, Where is the Truth? The Risk of Getting Confounded in a Continual World· ICML 2025 · arXiv:2402.06434
- Paper Dorovatas, Schwerin, … , Choi, Mundt, Aljundi (24 authors), Position: Modular Memory is the Key to Continual Learning Agents· ICML 2026 (spotlight) · arXiv:2603.01761
- Paper on deepfake detection: Das, Böse, Paul, Storch, Mundt, Time to Fake It: Continual Machine Learning Intensifies Unfairness in Deepfake Detection · https://arxiv.org/abs/2610.21974
- Paper on the Language Change Project: Hegde, Paul, Frey, Brack, Kersting, Mundt, Schramowski, Chronoberg: A 250-Year Diachronic Corpus with Temporal Affective Lexicons for Language Model Analysis · EMNLP 2026: https://aclanthology.org/2026.emnlp-main.2515/
