Tractable Probabilistic Circuits - Seminar Series

Explore the world of tractable probabilistic circuits in this comprehensive seminar presented by Associate Professor Guy Van den Broeck from UCLA. Delve into the innovative approach of representing distributions through the computation graph of probabilistic inference, akin to neural networks. Discover how probabilistic circuits surpass traditional probabilistic graphical models and deep generative models by ensuring tractable inference for specific query types, including marginal probabilities, entropies, and expectations. Learn about the recent advancements in effectively training these models from data, outperforming VAE and flow-based likelihoods on MNIST-family benchmarks. Gain insights into the practical applications of probabilistic circuits, including state-of-the-art neural compression techniques. Throughout this 1-hour 18-minute seminar, examine the latest developments in learning, probabilistic inference, theory, and real-world applications of tractable probabilistic circuits.