A Diffusive Classification Loss for Learning Energy-based Generative Models

RuiKang OuYang & Louis Grenioux, José Miguel Hernández-Lobato · 2026

Read A Diffusive Classification Loss for Learning Energy-based Generative Models by RuiKang OuYang & Louis Grenioux, José Miguel Hernández-Lobato (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-21025.txt.

Description

RuiKang OuYang, Louis Grenioux, and José Miguel Hernández-Lobato propose a new loss for training energy-based generative models. Their 2026 paper works with score-based generative models. Such models usually learn the score, the gradient that points toward likely data. An alternative uses a series of time-dependent energy-based models. There, the score comes from the negative input-gradient of the energy. That route has its own training hurdles.

The authors introduce a diffusive classification loss for this energy-based path. Framing training as classification can stabilize learning. It can also sidestep costly sampling during training. Better energy-based models widen the toolkit for image and data generation. They also connect to physics, where energies are native language. The paper offers a fresh objective for a classic model family.

Catalog record

CREATORRuiKang OuYang & Louis Grenioux, José Miguel Hernández-Lobato
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE202 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-21025

Browse the full library · About the Tower of Babel