[1609.03126] Energy-based Generative Adversarial Networkcontact arXivarXiv Twitter

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.

1 mentions: @shinmura0
Date: 2020/02/13 14:21

Referring Tweets

@shinmura0 最後に追記。このような「デュアルオートエンコーダ」の構造は以下の論文で登場した模様。 ・t.co/V6laNMwrEut.co/waGFpByCwc 私は読んでいない(^^;

Bookmark Comments

Related Entries

Read more [1803.04765] Deep k-Nearest Neighbors: Towards Confident, Interpretable and Robust Deep Learning
4 users, 4 mentions 2019/02/04 09:47
Read more [1605.05395] Learning Deep Representations of Fine-grained Visual Descriptions
3 users, 1 mentions 2019/09/18 11:18
Read more [1707.05776] Optimizing the Latent Space of Generative Networkscontact arXivarXiv Twitter
5 users, 0 mentions 2020/01/28 18:51
Read more [1710.09412] mixup: Beyond Empirical Risk Minimizationcontact arXivarXiv Twitter
7 users, 0 mentions 2020/02/10 15:01
Read more [1606.07947] Sequence-Level Knowledge Distillationcontact arXivarXiv Twitter
4 users, 1 mentions 2020/02/18 23:21