Tractable and Expressive Generative Modeling with Probabilistic Flow Circuits

View at publisher Summarise with VIRAT
Authors: Sahil Sidheekh, Sriraam Natarajan
Cited by: 0
References: 25

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

In this chapter, we introduce Probabilistic Flow Circuits (PFCs), a new class of models that bridges the gap between the tractable inference capabilities of Probabilistic Circuits (PCs) and the expressive power of normalizing flows. By equipping each leaf node in a PC with a univariate flow transformation, PFCs achieve significantly enhanced flexibility while preserving the tractability of PCs for tasks such as marginal and conditional inference and MAP estimation. Empirical evaluations on synthetic data and real-world benchmarks demonstrate that PFCs outperform conventional PCs in density estimation, deliver improved sample quality, and retain the ability to answer complex queries exactly – highlighting their potential as a powerful hybrid approach for real-world generative modeling.

Fields of Study & Subjects

Bayesian Modeling and Causal Inference Gaussian Processes and Bayesian Inference Logic Reasoning and Knowledge

Publication details

Journal / source
Wiley eBooks
Pages
183–222
Publication date
20 March 2026
Language
EN
Document type
Book chapter