Class-Splitting Generative Adversarial Networks

By Guillermo Luis Grinblat, Lucas C. Uzal, Pablo M. Granitto. 2017

35 citations recorded by the source corpus.

Generative Adversarial Networks (GANs) produce systematically better quality samples when class label information is provided., i.e. in the conditional GAN setup. This is still observed for the recently proposed Wasserstein GAN formulation which stabilized adversarial training and allows considering high capacity network architectures such as ResNet. In this work we show how to boost conditional GAN by augmenting available class labels. The new classes come from clustering in the representation space learned by the same GAN model. The proposed strategy is also feasible when no class information is available, i.e. in the unsupervised setup. Our generated samples reach state-of-the-art Inception scores for CIFAR-10 and STL-10 datasets in both supervised and unsupervised setup.

Subfields: Artificial Intelligence, 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 Artificial Intelligence

In Computer Vision and Pattern Recognition

In Generative Models

Citations in the corpus

Progressive Growing of GANs for Improved Quality, Stability, and Variation (2017, background): “quality is high. 6.5CIFAR10 INCEPTION SCORES The best inception scores for CIFAR10 (10 categories of 32 32 RGB images) we are aware of are 7.90 for unsupervised and 8.87 for label conditioned setups (Grinblat et al., 2017). The large difference between the two numbers is primarily caused by “ghosts””
Generative Feature Replay For Class-Incremental Learning (2020, background): “Generative adversarial networks (GANs) [11] are able to generate realistic and sharp images conditioned on object categories [12, 39], text [42, 57], another image (image translation) [18, 58] and style transfer [10].”
Improving MMD-GAN Training with Repulsive Loss Function (2018, background): “It has been widely accepted that the quality of generated samples can be significantly improved by integrating labels (Odena et al. (2017); Miyato & Koyama (2018); Zhou et al. (2018)) or even pseudo-labels generated by k-means method (Grinblat et al. (2017)) in the training of discriminator.”
Transferring GANs: generating images from limited data (2018, background): “Examples of conditions are class labels [32, 35, 14], text [37, 46], another image (image translation [23, 50] and style transfer [11]).”
AGAN: Towards Automated Design of Generative Adversarial Networks (2019, background): “ntiﬁed might be applicable to a wide range of data sets. Method Inception Score FID Real data 11:24 :12 7:8 DCGAN style DCGAN 7:84 0:07 [42] DFM 8:51 0:13 SN-GAN 8:69 :09 47:5 ResNet WGAN-GP 9:05 :13 [43] 55:1 [10] SN-GAN 9:10 :04 40:1 Splitting GAN 9:50 :13 CAGAN 9:51 :14 Ours AGAN-A 9:23 :08 52:7 ”
Activation Maximization Generative Adversarial Nets (2017, background): “By further enhancing the discriminator with more filters in each layer, AM-GAN also outperforms the orthogonal work (Guillermo et al., 2017) that enhances the class label information via class splitting.”
Instance-Conditioned GAN (2021, unknown): “Previous works have attempted to improve the image generation quality and diversity of GANs by partitioning the data manifold through clustering techniques [2, 19, 33, 36, 42, 46], or by leveraging mixture models in their design [14, 17, 23].”