Energy-based generator matching: A neural sampler for general state space

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Woo, Dongyeop, Kim, Minsu, Kim, Minkyu, Seong, Kiyoung, Ahn, Sungsoo
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866909912807768064
author Woo, Dongyeop
Kim, Minsu
Kim, Minkyu
Seong, Kiyoung
Ahn, Sungsoo
author_facet Woo, Dongyeop
Kim, Minsu
Kim, Minkyu
Seong, Kiyoung
Ahn, Sungsoo
contents We propose Energy-based generator matching (EGM), a modality-agnostic approach to train generative models from energy functions in the absence of data. Extending the recently proposed generator matching, EGM enables training of arbitrary continuous-time Markov processes, e.g., diffusion, flow, and jump, and can generate data from continuous, discrete, and a mixture of two modalities. To this end, we propose estimating the generator matching loss using self-normalized importance sampling with an additional bootstrapping trick to reduce variance in the importance weight. We validate EGM on both discrete and multimodal tasks up to 100 and 20 dimensions, respectively.
format Preprint
id arxiv_https___arxiv_org_abs_2505_19646
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Energy-based generator matching: A neural sampler for general state space
Woo, Dongyeop
Kim, Minsu
Kim, Minkyu
Seong, Kiyoung
Ahn, Sungsoo
Machine Learning
We propose Energy-based generator matching (EGM), a modality-agnostic approach to train generative models from energy functions in the absence of data. Extending the recently proposed generator matching, EGM enables training of arbitrary continuous-time Markov processes, e.g., diffusion, flow, and jump, and can generate data from continuous, discrete, and a mixture of two modalities. To this end, we propose estimating the generator matching loss using self-normalized importance sampling with an additional bootstrapping trick to reduce variance in the importance weight. We validate EGM on both discrete and multimodal tasks up to 100 and 20 dimensions, respectively.
title Energy-based generator matching: A neural sampler for general state space
topic Machine Learning
url https://arxiv.org/abs/2505.19646