Probabilistic Circuit

Probabilistic circuits (PCs) are a class of deep generative models designed for efficient probabilistic inference, addressing the trade-off between model expressiveness and computational tractability. Current research focuses on enhancing PC expressiveness through novel architectures like sum-of-squares and probabilistic neural circuits, improving training efficiency via techniques such as GPU acceleration and latent variable distillation, and extending their applicability to diverse data types including graphs and continuous variables. PCs offer a powerful framework for tractable probabilistic modeling with applications in various fields, including image and language modeling, causal inference, and robust machine learning.