Strudel: A fast and accurate learner of structured-decomposable probabilistic circuits

Probabilistic circuits (PCs) represent a probability distribution as computational graph. Enforcing structural properties on these graphs guarantees that several inference scenarios become tractable. Among properties, structured decomposability is particularly appealing one: it enables the efficient and exact computations of complex logical formulas, can be used to reason about expected output certain predictive models under missing data. This paper proposes Strudel, simple, fast accurate learning algorithm for structured-decomposable PCs. Compared prior work PCs, Strudel delivers more single PC in fewer iterations, dramatically scales when building ensembles It achieves this scalability by exploiting another property called determinism, sharing same graph across mixture components. We show advantages standard density estimation benchmarks challenging scenarios.