Energy-based generative adversarial networks

Junbo Zhao, Michael Mathieu, Yann LeCun

Research output: Contribution to conferencePaper

Abstract

We introduce the “Energy-based Generative Adversarial Network” model (EBGAN) which views the discriminator as an energy function that attributes low energies to the regions near the data manifold and higher energies to other regions. Similar to the probabilistic GANs, a generator is seen as being trained to produce contrastive samples with minimal energies, while the discriminator is trained to assign high energies to these generated samples. Viewing the discriminator as an energy function allows to use a wide variety of architectures and loss functionals in addition to the usual binary classifier with logistic output. Among them, we show one instantiation of EBGAN framework as using an auto-encoder architecture, with the energy being the reconstruction error, in place of the discriminator. We show that this form of EBGAN exhibits more stable behavior than regular GANs during training. We also show that a single-scale architecture can be trained to generate high-resolution images.

Original languageEnglish (US)
StatePublished - Jan 1 2019
Event5th International Conference on Learning Representations, ICLR 2017 - Toulon, France
Duration: Apr 24 2017Apr 26 2017

Conference

Conference5th International Conference on Learning Representations, ICLR 2017
CountryFrance
CityToulon
Period4/24/174/26/17

Fingerprint

Discriminators
energy
Image resolution
Logistics
Classifiers
Energy
Generative
reconstruction
logistics

ASJC Scopus subject areas

  • Education
  • Computer Science Applications
  • Linguistics and Language
  • Language and Linguistics

Cite this

Zhao, J., Mathieu, M., & LeCun, Y. (2019). Energy-based generative adversarial networks. Paper presented at 5th International Conference on Learning Representations, ICLR 2017, Toulon, France.

Energy-based generative adversarial networks. / Zhao, Junbo; Mathieu, Michael; LeCun, Yann.

2019. Paper presented at 5th International Conference on Learning Representations, ICLR 2017, Toulon, France.

Research output: Contribution to conferencePaper

Zhao, J, Mathieu, M & LeCun, Y 2019, 'Energy-based generative adversarial networks', Paper presented at 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, 4/24/17 - 4/26/17.
Zhao J, Mathieu M, LeCun Y. Energy-based generative adversarial networks. 2019. Paper presented at 5th International Conference on Learning Representations, ICLR 2017, Toulon, France.
Zhao, Junbo ; Mathieu, Michael ; LeCun, Yann. / Energy-based generative adversarial networks. Paper presented at 5th International Conference on Learning Representations, ICLR 2017, Toulon, France.
@conference{76256ca3b51d4ed79ba46d0a3c30e37d,
title = "Energy-based generative adversarial networks",
abstract = "We introduce the “Energy-based Generative Adversarial Network” model (EBGAN) which views the discriminator as an energy function that attributes low energies to the regions near the data manifold and higher energies to other regions. Similar to the probabilistic GANs, a generator is seen as being trained to produce contrastive samples with minimal energies, while the discriminator is trained to assign high energies to these generated samples. Viewing the discriminator as an energy function allows to use a wide variety of architectures and loss functionals in addition to the usual binary classifier with logistic output. Among them, we show one instantiation of EBGAN framework as using an auto-encoder architecture, with the energy being the reconstruction error, in place of the discriminator. We show that this form of EBGAN exhibits more stable behavior than regular GANs during training. We also show that a single-scale architecture can be trained to generate high-resolution images.",
author = "Junbo Zhao and Michael Mathieu and Yann LeCun",
year = "2019",
month = "1",
day = "1",
language = "English (US)",
note = "5th International Conference on Learning Representations, ICLR 2017 ; Conference date: 24-04-2017 Through 26-04-2017",

}

TY - CONF

T1 - Energy-based generative adversarial networks

AU - Zhao, Junbo

AU - Mathieu, Michael

AU - LeCun, Yann

PY - 2019/1/1

Y1 - 2019/1/1

N2 - We introduce the “Energy-based Generative Adversarial Network” model (EBGAN) which views the discriminator as an energy function that attributes low energies to the regions near the data manifold and higher energies to other regions. Similar to the probabilistic GANs, a generator is seen as being trained to produce contrastive samples with minimal energies, while the discriminator is trained to assign high energies to these generated samples. Viewing the discriminator as an energy function allows to use a wide variety of architectures and loss functionals in addition to the usual binary classifier with logistic output. Among them, we show one instantiation of EBGAN framework as using an auto-encoder architecture, with the energy being the reconstruction error, in place of the discriminator. We show that this form of EBGAN exhibits more stable behavior than regular GANs during training. We also show that a single-scale architecture can be trained to generate high-resolution images.

AB - We introduce the “Energy-based Generative Adversarial Network” model (EBGAN) which views the discriminator as an energy function that attributes low energies to the regions near the data manifold and higher energies to other regions. Similar to the probabilistic GANs, a generator is seen as being trained to produce contrastive samples with minimal energies, while the discriminator is trained to assign high energies to these generated samples. Viewing the discriminator as an energy function allows to use a wide variety of architectures and loss functionals in addition to the usual binary classifier with logistic output. Among them, we show one instantiation of EBGAN framework as using an auto-encoder architecture, with the energy being the reconstruction error, in place of the discriminator. We show that this form of EBGAN exhibits more stable behavior than regular GANs during training. We also show that a single-scale architecture can be trained to generate high-resolution images.

UR - http://www.scopus.com/inward/record.url?scp=85071037867&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85071037867&partnerID=8YFLogxK

M3 - Paper

AN - SCOPUS:85071037867

ER -