Abstract
This paper establishes a rigorous connection between circuit representationsand tensor factorizations, two seemingly distinct yet fundamentally relatedareas. By connecting these fields, we highlight a series of opportunities thatcan benefit both communities. Our work generalizes popular tensorfactorizations within the circuit language, and unifies various circuitlearning algorithms under a single, generalized hierarchical factorizationframework. Specifically, we introduce a modular "Lego block" approach to buildtensorized circuit architectures. This, in turn, allows us to systematicallyconstruct and explore various circuit and tensor factorization models whilemaintaining tractability. This connection not only clarifies similarities anddifferences in existing models, but also enables the development of acomprehensive pipeline for building and optimizing new circuit/tensorfactorization architectures. We show the effectiveness of our framework throughextensive empirical evaluations, and highlight new research opportunities fortensor factorizations in probabilistic modeling.