Mathematical Foundations of AI

Chalmers and the University of Gothenburg host a vibrant community applying AI to a variety of domains, ranging from drug discovery to genomics. However, foundational research must be strengthened if we are to keep pace with this fast-growing technology.

Furthermore, AI is evolving without fundamental principles, stacking together pieces as problems emerge. This approach – while effective in the short term – risks hindering progress and even causing harm in the longer term. We must move beyond reactive, empirical patching to understand exactly how AI models behave, and when and why modern models misbehave. Understanding the "why" behind AI's successes and failures is essential for transforming AI into a field with theoretical guarantees, enabling the design of the next generation of trustworthy systems. To achieve this, we need to build the mathematical foundations of the field.

Leveraging the multi-disciplinary community that Chalmers and the University of Gothenburg host, this theme brings together researchers from computer science, electrical engineering, physics and mathematics, to share their research and perspectives on understanding AI systems. By mapping neural networks into statistical physics of complex systems models, applying mathematical representation theory to deep learning architectures, or employing techniques from other disciplines that only a multidisciplinary environment can allow, we can decode the underlying mechanics of these algorithms.

The theme targets researchers who often work in fragmented clusters without unified communication. Theoretical physicists, mathematicians, engineers, and computer scientists frequently tackle similar foundational problems but lack a shared scientific vocabulary. By strengthening and expanding the AI Foundations community, the initiative aims to enhance cohesion and international visibility. We unite these groups through regular seminars, community retreats, and an international workshop, fostering the cross-disciplinary tools needed to build the mathematical foundations for verifiable, trustworthy AI systems.