Particle Dynamics for Latent-Variable Energy-Based Models

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Tang, Shiqin, Zhuang, Shuxin, Feng, Rong, Yu, Runsheng, Li, Hongzong, Zhang, Youzhi
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866908599449550848
author Tang, Shiqin
Zhuang, Shuxin
Feng, Rong
Yu, Runsheng
Li, Hongzong
Zhang, Youzhi
author_facet Tang, Shiqin
Zhuang, Shuxin
Feng, Rong
Yu, Runsheng
Li, Hongzong
Zhang, Youzhi
contents Latent-variable energy-based models (LVEBMs) assign a single normalized energy to joint pairs of observed data and latent variables, offering expressive generative modeling while capturing hidden structure. We recast maximum-likelihood training as a saddle problem over distributions on the latent and joint manifolds and view the inner updates as coupled Wasserstein gradient flows. The resulting algorithm alternates overdamped Langevin updates for a joint negative pool and for conditional latent particles with stochastic parameter ascent, requiring no discriminator or auxiliary networks. We prove existence and convergence under standard smoothness and dissipativity assumptions, with decay rates in KL divergence and Wasserstein-2 distance. The saddle-point view further yields an ELBO strictly tighter than bounds obtained with restricted amortized posteriors. Our method is evaluated on numerical approximations of physical systems and performs competitively against comparable approaches.
format Preprint
id arxiv_https___arxiv_org_abs_2510_15447
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Particle Dynamics for Latent-Variable Energy-Based Models
Tang, Shiqin
Zhuang, Shuxin
Feng, Rong
Yu, Runsheng
Li, Hongzong
Zhang, Youzhi
Machine Learning
Latent-variable energy-based models (LVEBMs) assign a single normalized energy to joint pairs of observed data and latent variables, offering expressive generative modeling while capturing hidden structure. We recast maximum-likelihood training as a saddle problem over distributions on the latent and joint manifolds and view the inner updates as coupled Wasserstein gradient flows. The resulting algorithm alternates overdamped Langevin updates for a joint negative pool and for conditional latent particles with stochastic parameter ascent, requiring no discriminator or auxiliary networks. We prove existence and convergence under standard smoothness and dissipativity assumptions, with decay rates in KL divergence and Wasserstein-2 distance. The saddle-point view further yields an ELBO strictly tighter than bounds obtained with restricted amortized posteriors. Our method is evaluated on numerical approximations of physical systems and performs competitively against comparable approaches.
title Particle Dynamics for Latent-Variable Energy-Based Models
topic Machine Learning
url https://arxiv.org/abs/2510.15447