On November 13, we hosted Dr. Nikolay Malkin, Chancellor’s Fellow in Informatics at the University of Edinburgh and fellow of CIFAR’s Learning in Machines and Brains programme, for an engaging hybrid session exploring recent advances in structured probabilistic inference.
Dr. Malkin presented algorithmic approaches that tackle a core challenge in modern Bayesian machine learning: learning and reasoning with complex, compositional data such as graphs, sequences, and symbolic programs. The talk showcased how reinforcement learning, variational inference, and Monte Carlo methods can be combined to handle uncertainty and structure jointly—a critical capability for modern AI systems.
Key topics included:
- Modeling Bayesian posteriors over high-dimensional and structured variables
- Induction and discovery of compositional structure in generative models
- Neurosymbolic methods for uncertainty-aware reasoning in language and formal systems
Dr. Malkin highlighted conceptual connections between these methodologies and their diverse applications—from generative modeling and scientific discovery to practical challenges in inverse imaging, remote sensing, biological and chemical structure discovery, and robot control.
The hybrid format enabled both in-person attendees at the ICC Conference Room and remote participants to engage in a stimulating discussion about the intersection of probabilistic inference, symbolic reasoning, and machine learning.