Least Squares Generative Adversarial Networks
By Xudong Mao, Qing Li, Haoran Xie, Raymond Y. K. Lau, Zhen Wang, Stephen Paul Smolley. 2016
5,091 citations recorded by the source corpus.
Unsupervised learning with generative adversarial networks (GANs) has proven hugely successful. Regular GANs hypothesize the discriminator as a classifier with the sigmoid cross entropy loss function. However, we found that this loss function may lead to the vanishing gradients problem during the learning process. To overcome such a problem, we propose in this paper the Least Squares Generative Adversarial Networks (LSGANs) which adopt the least squares loss function for the discriminator. We show that minimizing the objective function of LSGAN yields minimizing the Pearson X2 divergence. There are two benefits of LSGANs over regular GANs. First, LSGANs are able to generate higher quality images than regular GANs. Second, LSGANs perform more stable during the learning process. We evaluate LSGANs on LSUN and CIFAR-10 datasets and the experimental results show that the images generated by LSGANs are of better quality than the ones generated by regular GANs. We also conduct two comparison experiments between LSGANs and regular GANs to illustrate the stability of LSGANs.
Subfields: Computer Vision and Pattern Recognition, Generative Models
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Computer Vision and Pattern Recognition
- Luc Van Gool
- Dacheng Tao
- Ming-Hsuan Yang
- Trevor Darrell
- Xiaogang Wang
- Jitendra Malik
- Yi Yang
- Andrew Zisserman
In Generative Models
- Jun-Yan Zhu
- Stefano Ermon
- Richard Zhang
- Eli Shechtman
- Alexei A. Efros
- Jiaming Song
- Yoshua Bengio
- Ming-Hsuan Yang
Citations in the corpus
- Are GANs Created Equal? A Large-Scale Study (2017, background): “[18] propose a least-squares loss for the discriminator and show that minimizing the corresponding objective (LS GAN in Table 1) implicitly minimizes the Pearson χ(2) divergence.”
- Training Generative Adversarial Networks with Limited Data (2020, methodology): “In Figure 8a, we employ multiplicative Gaussian dropout for all layers of the discriminator, similar to the approach employed by Karras et al. [13] in the context of LSGAN loss [18].”
- Large Scale GAN Training for High Fidelity Natural Image Synthesis (2018, background): “One line of work is focused on changing the objective function (Arjovsky et al., 2017; Mao et al., 2016; Lim & Ye, 2017; Bellemare et al., 2017; Salimans et al., 2018) to encourage convergence.”
- Progressive Growing of GANs for Improved Quality, Stability, and Variation (2017, methodology): “…recently that formulation has been improved (Hjelm et al., 2017) and a number of more stable alternatives have been proposed, including least squares (Mao et al., 2016b), absolute deviation with margin (Zhao et al., 2017), and Wasserstein distance (Arjovsky et al., 2017; Gulrajani et al., 2017).”
- Generative Adversarial Networks (2021, neutral): “Text-to-image synthesis using GANs involves generating images from textual descriptions.”
- Semantic Image Synthesis With Spatially-Adaptive Normalization (2019, methodology): “We train the generator with the same multi-scale discrim inator and loss function used in pix2pixHD except that we replace the least squared loss term [33] with the hinge loss term [30,35,51].”
- Diagnosing and Enhancing VAE Models (2019, background): “…from (Lucic et al., 2018) involving numerous competing GAN models, including MM GAN (Goodfellow et al., 2014), WGAN (Arjovsky et al., 2017), WGAN-GP (Gulrajani et al., 2017), NS GAN (Fedus et al., 2017), DRAGAN (Ko-dali et al., 2017), LS GAN (Mao et al., 2017) and BEGAN (Berthelot et al., 2017).”
- Improved Training of Wasserstein GANs (2017, methodology): “For each architecture, we train models using four different GAN procedures: WGAN with gradient penalty, WGAN with weight clipping, DCGAN [21], and Least-Squares GAN [17].”
- Cross-Modal Contrastive Learning for Text-to-Image Generation (2021, background): “A large amount of work has focused on designing the adversarial objective to improve training [12, 1, 31, 47, 29, 54].”
- Self-Supervised GANs via Auxiliary Rotation Loss (2018, background): “As a result, many techniques to stabilize GAN training have been proposed [4, 5, 6, 7, 8, 9, 10].”
- High-Resolution Image Synthesis and Semantic Manipulation with Conditional GANs (2017, methodology): “Implementation details We use LSGANs [36] for stable training.”
- Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks (2017, methodology): “First, for LGAN (Equation 1), we replace the negative log likelihood objective by a least-squares loss [35].”
- VQVAE Unsupervised Unit Discovery and Multi-scale Code2Spec Inverter for Zerospeech Challenge 2019 (2019, positive): “For the adversarial loss, we found LSGAN is stabler, thus LSGAN with β = 1 is used in every model.”
- StackGAN++: Realistic Image Synthesis with Stacked Generative Adversarial Networks (2017, methodology): “EBGAN-PT [56], LSGAN [26], and WGAN-GP [13] on the”
- Multimodal Unsupervised Image-to-Image Translation (2018, methodology): “We use the LSGAN objective proposed by Mao et al. [38].”
- Generating Adversarial Examples with Adversarial Networks (2018, methodology): “For GANs training, we use the least squares objective proposed by LSGAN (Mao et al., 2016), as it has been shown to produce better results with more stable training.”
- InfoMax-GAN: Improved Adversarial Image Generation via Information Maximization and Contrastive Learning (2020, methodology): “Other approaches employing information-theoretic principles include Variational GAN (VGAN) [50], which uses an information bottleneck [58] to regularize the discriminator representations; with [6, 39, 44] extending to minimise divergences apart from the original JS divergence.”
- Continual learning with hypernetworks (2019, methodology): “We use the Pearson Chi 2 Least-Squares GAN loss from Mao et al. (2017) in our experiments.”