PacGAN: The Power of Two Samples in Generative Adversarial Networks

By Zinan Lin, Ashish Khetan, Giulia C. Fanti, Sewoong Oh. 2017

380 citations recorded by the source corpus.

Generative adversarial networks (GANs) are innovative techniques for learning generative models of complex data distributions from samples. Despite remarkable improvements in generating realistic images, one of their major shortcomings is the fact that in practice, they tend to produce samples with little diversity, even when trained on diverse datasets. This phenomenon, known as mode collapse, has been the main focus of several recent advances in GANs. Yet there is little understanding of why mode collapse happens and why recently-proposed approaches mitigate mode collapse. We propose a principled approach to handle mode collapse called packing. The main idea is to modify the discriminator to make decisions based on multiple samples from the same class, either real or artificially generated. We borrow analysis tools from binary hypothesis testing—in particular the seminal result of (Blackwell, 1953)—to prove a fundamental connection between packing and mode collapse. We show that packing naturally penalizes generators with mode collapse, thereby favoring generator distributions with less mode collapse during the training process. Numerical experiments on benchmark datasets suggests that packing provides significant improvements in practice as well.

Subfields: Computer Vision and Pattern Recognition, Statistical and Nonlinear Physics, 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

In Statistical and Nonlinear Physics

In Generative Models

Citations in the corpus

Progressive Growing of GANs for Improved Quality, Stability, and Variation (2017, background): “ have found it best to insert it towards the end (see Appendix A.1 for details). We experimented with a richer set of statistics, but were not able to improve the variation further. In parallel work, Lin et al. (2017) provide theoretical insights about the beneﬁts of showing multiple images to the d”
VAEBM: A Symbiosis between Variational Autoencoders and Energy-based Models (2020, background): “Following Lin et al. (2018), we report the number of covered modes and the KL divergence from the categorical distribution over 1000 categories from generated samples to true data (Table 5).”
Improved Precision and Recall Metric for Assessing Generative Models (2019, background): “The baseline configuration (A) has a dedicated minibatch standard deviation layer that aims to increase variation in the generated images [12, 16].”
Adversarial score matching and improved sampling for image generation (2020, neutral): “We proposed a hybrid approach between 3-Stacked MNIST Modes (Max 1000) KL DCGAN (Radford et al., 2015b) 99.0 3.40 ALI (Dumoulin et al., 2016) 16.0 5.40 Unrolled GAN (Metz et al., 2016) 48.7 4.32 VEEGAN (Srivastava et al., 2017) 150.0 2.95 PacDCGAN2 (Lin et al., 2017b) 1000.0 0.06 WGAN-GP (Kumar et a”
Mutual Information Neural Estimation (2018, background): “Number of captured modes and Kullblack-Leibler divergence between the training and samples distributions for DC-GAN (Radford et al., 2015), ALI (Dumoulin et al., 2016), Un-rolled GAN (Metz et al., 2017), VeeGAN (Srivastava et al., 2017), PacGAN (Lin et al., 2017). and reverse (decoder) models, respe”
Discriminator Rejection Sampling (2018, methodology): “We investigate the impact of DRS on a low-dimensional synthetic data set consisting of a mixture of twenty-ﬁve 2D isotropic Gaussian distributions (each with standard deviation of 0.05) arranged in a grid (Dumoulin et al., 2016; Srivastava et al., 2017; Lin et al., 2017).”
Pros and Cons of GAN Evaluation Measures (2018, methodology): “The reverse KL divergence over the modes has been used in [50] to measure the quality of mode collapse as follows.”
The relativistic discriminator: a key element missing from standard GAN (2018, methodology): “To alleviate this problem, for 256x256 images, we packed the discriminator [Lin et al., 2017] (i.”
Maximum Entropy Generators for Energy-Based Models (2019, neutral): “Our en-Maximum (Du-moulin et al., 2016), Unrolled GAN (Metz et al., 2016), Vee-GAN (Srivastava et al., 2017), PacGAN (Lin et al., 2017), WGAN-GP (Gulrajani et al., 2017).”
CTAB-GAN: Effective Table Data Synthesizing (2021, methodology): “CTGAN [24] integrates PacGAN [14] structure in its discriminator and uses WGAN loss plus gradient penalty [8] to train a conditional GAN framework.”
Steering Self-Supervised Feature Learning Beyond Local Pixel Statistics (2020, neutral): “This technique was also proposed by [36].”
Privacy-Preserving Inference in Machine Learning Services Using Trusted Execution Environments (2019, background): “The potency of a GAN trained with conditional information has been demonstrated in works such as [21], [24].”
Revisiting Precision and Recall Definition for Generative Model Evaluation (2019, methodology): “Note that a similar notion was proposed by the authors of PACGAN (Lin et al., 2018) under the name of mode collapse region (denoted as MCR ( P, Q ) ).”
DLow: Diversifying Latent Flows for Diverse Human Motion Prediction (2020, neutral): “s been used for diverse feature [23] and data [64] selection. Diversity has also been a key aspect in generative modeling. A vast body of work has tried to alleviate the mode collapse problem in GANs [11,12,65,4,25,16,47,72] and the posterior collapse problem in VAEs [78,66,36,8,48,29]. Normalizing ”
Diverse Image Generation via Self-Conditioned GANs (2020, methodology): “For Stacked-MNIST dataset, we use the DCGAN architecture [35] identical to that used in prior work[25].”
Modeling Tabular data using Conditional GAN (2019, methodology): “Here we compare it with three alternatives, WGANGP only, vanilla GAN loss only, and vanilla GAN + PacGAN.”
Instance-Conditioned GAN (2021, positive): “Although many approaches attempt to mitigate this problem – e.g. [20, 32, 35, 38] –, complex data distributions such as the one in ImageNet [45] remain a challenge for unconditional GANs [33, 36].”