Probabilistic Circuits: Deep Probabilistic Models with Tractable Inference

Explore probabilistic circuits as an expressive type of deep probabilistic model in this 56-minute talk by Robert Peharz from TU Graz, presented at the DataLearning Seminars of Imperial College London. Delve into the key features of artificial intelligence, focusing on reasoning under uncertainty and the importance of probability in decision-making. Examine the computational challenges faced by most deep learning-based probabilistic models, including GANs, VAEs, and Normalizing Flows. Learn about the framework of Probabilistic Circuits (PCs) and their unique ability to allow a wide range of exact and efficient probabilistic inference routines. Discover the "structural grammar" of PCs and understand how their structural properties relate to tractable probabilistic inference. Gain insights into recent developments in the field of PCs and their connections to neuro-symbolic reasoning.