Mar, 2024

概率生成式电路 —— 揭秘

TL;DRZhang et al. introduced probabilistic generating circuits (PGCs) as a probabilistic model that unifies probabilistic circuits (PCs) and determinantal point processes (DPPs); the main insight of the paper is that negative weights are responsible for the power of PGCs, and PGCs are essentially PCs with negative weights; PGCs are defined only for binary random variables, while categorial variables with larger image size can be modeled as PCs with negative weights, allowing for tractable marginalization, making PCs with negative weights a stricter subset of PGCs.