Foretelling the Future of AI through Mathematical Foundations - From Learning Architectures to Trustworthy Agents
AAAI 2027 2nd MATH4AI Workshop
February 22-23, 2027 | Montréal, Canada | Palais des congrès
AAAI 2027 2nd MATH4AI Workshop
February 22-23, 2027 | Montréal, Canada | Palais des congrès
Mathematics has played an essential role throughout the development of artificial intelligence. Mathematical logic and theories of computation supported the emergence of symbolic AI, while probability, statistics, optimization, and linear algebra laid the foundations for modern machine learning. Today, ideas from approximation theory, geometry, dynamical systems, numerical analysis, and scientific computing are increasingly being used to study deep learning and to design new AI architectures.
However, the rapid progress of AI has outpaced our mathematical understanding of these systems. Modern models can achieve impressive results, but many basic questions remain open. We still do not fully understand why certain architectures work so well, how their capabilities depend on model structure and training, when they will generalize, or why they sometimes fail in unexpected ways. Mathematical analysis is therefore needed not only to explain existing models, but also to reveal their limitations and guide the development of future AI.
These questions become even more important as AI systems move beyond prediction and generation toward autonomous agents. Such agents may reason over several steps, plan toward a goal, use external tools, maintain memory, learn from feedback, and interact with people, environments, or other agents. This raises a new set of mathematical problems. How should reasoning, memory, goals, and uncertainty be represented? When is an agent's learning or decision process stable? How do errors accumulate across a long sequence of actions? How can we evaluate, control, or verify an agent's behavior? What guarantees can be established when multiple agents cooperate or compete?
This workshop will explore how mathematical foundations can help us understand and shape future AI, from learning architectures to autonomous agents. It will bring together researchers interested in the theory of deep learning, mathematics-inspired model design, scientific machine learning, foundation models, and trustworthy agentic systems. The aim is to connect mathematical analysis with the practical questions arising from increasingly capable AI systems.
The workshop will be hosted for one-day and in-person and will feature a mix of invited talks from leading experts, contributed presentations from submitted papers, poster sessions, and panel discussions. This format is designed to maximize interaction between participants and foster meaningful dialogues that can lead to new collaborations and research directions. The two poster sessions and panel discussions will boost the discussion among attendees about the mathematical foundations which can inspire new types of deep learning algorithms or architectures and also design of future architectures and agents that are reliable, efficient, and trustworthy.