Unrolled Generative Adversarial Networks

By Luke Metz, Ben Poole, David Pfau, Jascha Narain Sohl-Dickstein. 2016

1,076 citations recorded by the source corpus.

We introduce a method to stabilize Generative Adversarial Networks (GANs) by defining the generator objective with respect to an unrolled optimization of the discriminator. This allows training to be adjusted between using the optimal discriminator in the generator's objective, which is ideal but infeasible in practice, and using the current value of the discriminator, which is often unstable and leads to poor solutions. We show how this technique solves the common problem of mode collapse, stabilizes training of GANs with complex recurrent generators, and increases diversity and coverage of the data distribution by the generator.

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

High-Resolution Image Synthesis with Latent Diffusion Models (2021, background): “cult to optimize [2, 27, 53] and struggle to capture the full data distribution [54].”
ColdGANs: Taming Language GANs with Cautious Sampling Strategies (2020, neutral): “We connect this to the failure mode in GANs observed by Metz et al. [25], where the generator and the discriminator oscillate during training, rather than converging to a ﬁxed point.”
Progressive Growing of GANs for Improved Quality, Stability, and Variation (2017, methodology): “Figure 9 shows non-curated images generated in the unsupervised setting, and Table 4: Results for MNIST discrete mode test using two tiny discriminators (K/4, K/2) deﬁned by Metz et al. (2016).”
Image Super-Resolution via Iterative Refinement (2021, background): “instability and mode collapse [19], [20], [21], [22].”
Imperfect ImaGANation: Implications of GANs Exacerbating Biases on Facial Data Augmentation and Snapchat Selfie Lenses (2020, background): “…et al., 2018; Goodfellow, 2016; Che et al., 2017; Arora et al., 2017) and several GAN variants have been developed to alleviate its effects (Metz et al., 2017; Srivastava et al., 2017; Arjovsky et al., 2017; Miyato et al., 2018; Tolstikhin et al., 2017; Karras et al., 2019a), a distinction…”
Improved Precision and Recall Metric for Assessing Generative Models (2019, background): “While metrics have been proposed to estimate the degree of variation, these have not seen widespread use as they are subjective [1], domain specific [19], or not reliable enough [22].”
Improved Training of Wasserstein GANs (2017, background): “GANs can produce very visually appealing samples, but are often hard to train, and much of the recent work on the subject [22, 18, 2, 20] has been devoted to finding ways of stabilizing training.”
Many Paths to Equilibrium: GANs Do Not Need to Decrease a Divergence At Every Step (2017, methodology): “To assess the effectiveness of the gradient penalty on standard datasets for the non-saturating GAN formulation, we train a non-saturating GAN, a non-saturating GAN with the gradient penalty introduced by [9] (denoted by GAN-GP), a non-saturating GAN with the gradient penalty introduced by [12] (den”
Coulomb GANs: Provably Optimal Nash Equilibria via Potential Fields (2017, methodology): “To avoid mode collapsing without batch normalization, several methods have been proposed [5, 33, 38].”
Gradient descent GAN optimization is locally stable (2017, methodology): “Specifically, our regularization term (motivated by stability analysis) captures a degree of “foresight” of the generator in the optimization procedure, similar to the unrolled GANs procedure [Metz et al., 2017].”
The Numerics of GANs (2017, background): “While simplistic, algorithms training GANs often fail to converge even on such simple examples without extensive ﬁne-tuning of the architecture and hyper parameters [13].”
Wasserstein GAN (2017, background): “This is due to the fact that mode collapse comes from the fact that the optimal generator for a ﬁxed discriminator is a sum of deltas on the points the discriminator assigns the highest values, as observed by [4] and highlighted in [11].”
Self-Attention Generative Adversarial Networks (2018, background): “…works have attempted to stabilize the GAN training dynamics and improve the sample diversity by designing new network architectures [22, 37, 10], modifying the learning objectives and dynamics [1, 27, 15, 3, 39], adding regularization methods [7, 16] and introducing heuristic tricks [26, 19].”