Scalable Learning of Probabilistic Circuits

Probabilistic circuits (PCs) are a family of tractable probabilistic models capable of answering a wide range of queries exactly and in polytime. While inference is usually straightforward, learning PCs that both obey the needed restrictions for inference tractability and exploit their expressive power has proven to be a challenge. This dissertation aims to propose fast and scalable structure learning algorithms for PCs from two different standpoints: from a logical point of view, we efficiently construct a PC that takes certain knowledge in the form of logical constraints and scalably translate them into a probabilistic circuit; from the viewpoint of data guided structure search, we propose hierarchically building PCs from random hyperplanes. We empirically show that either approach is competitive against state-of-the-art methods of the same class in both performance and scalability.