Uncertainty-Aware Learning: From Bayesian Neural Networks to Agentic Decision Making

-

B4/5 L0 A0215

This talk points out that uncertainty quantification is important for reliable AI, and that modern machine learning should be viewed through the lens of probabilistic decision making.

The talk begins with Bayesian neural networks, illustrating the role of uncertainty quantification through simple examples and emphasizing the inferential challenges that arise for larger neural architectures. It then connects these challenges to probabilistic limits via neural network Gaussian processes, explaining under what conditions bottlenecked neural architectures converge to deep Gaussian processes. The talk next treats topological deep learning as a representation challenge, where non-Euclidean domains motivate new probabilistic questions. It concludes by elevating uncertainty from a modeling feature to a systems requirement, outlining Bayesian decision-theoretic principles for agentic AI orchestration in which model components may be non-Bayesian, but collective behavior should remain uncertainty-aware and reliable.

Presenters

Theodore Papamarkou, Founder, PolyShape; Visiting Professor, School of Applied Mathematical and Physical Sciences (SEMFE), National Technical University of Athens (NTUA)

Brief Biography

Theodore Papamarkou is a founder of PolyShape, a startup focused on collective intelligence, and a visiting professor at the National Technical University of Athens. Prior to his current role, Theodore held the positions of distinguished professor at Zhejiang Normal University, of professor at The University of Manchester, of research scientist in artificial intelligence at the Oak Ridge National Laboratory, and of assistant professor at the University of Glasgow. Earlier in his career, he worked as a post-doctoral researcher at the University of Warwick, at University College London, and at the University of Cambridge. Theodore's research spans Bayesian, categorical and topological aspects of deep learning and of agentic AI. Theodore is also interested in applications of these methodological areas in biomedical research and healthcare.