Search papers, labs, and topics across Lattice.
This paper introduces Geometry-Aware Probabilistic Circuits (GAPCs) that incorporate Voronoi tessellations into the sum nodes of probabilistic circuits to capture the local geometry of the data manifold. The authors formalize the incompatibility of naive Voronoi tessellations with tractable inference in PCs and propose two solutions: an approximate inference framework with bounds and a structural condition for exact inference. They also introduce a differentiable relaxation for Voronoi tessellations, enabling gradient-based learning, and demonstrate improved density estimation performance.
Voronoi tessellations can be incorporated into probabilistic circuits to capture data geometry, but only with careful architectural constraints or approximate inference.
Probabilistic circuits (PCs) enable exact and tractable inference but employ data independent mixture weights that limit their ability to capture local geometry of the data manifold. We propose Voronoi tessellations (VT) as a natural way to incorporate geometric structure directly into the sum nodes of a PC. However, na\"ively introducing such structure breaks tractability. We formalize this incompatibility and develop two complementary solutions: (1) an approximate inference framework that provides guaranteed lower and upper bounds for inference, and (2) a structural condition for VT under which exact tractable inference is recovered. Finally, we introduce a differentiable relaxation for VT that enables gradient-based learning and empirically validate the resulting approach on standard density estimation tasks.
