Tractable Probabilistic Circuits

Probabilistic circuits represent distributions through the computation graph of probabilistic inference, as a type of neural network. They move beyond probabilistic graphical models and other deep generative models by guaranteeing tractable inference for certain classes of queries: marginal probabilities, entropies, expectations, and related queries of interest. These probabilistic circuit models are now also effectively learned from data, outperforming VAE and flow-based likelihoods on MNIST-family benchmarks. They thus enable new solutions to some key problems in machine learning, including state-of-the-art neural compression results. This talk will overview these recent developments, in terms of learning, probabilistic inference, theory, and applications.