Improved Techniques for Training GANs
By Tim Salimans, Ian J. Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen. 2016
10,497 citations recorded by the source corpus.
We present a variety of new architectural features and training procedures that we apply to the generative adversarial networks (GANs) framework. We focus on two applications of GANs: semi-supervised learning, and the generation of images that humans find visually realistic. Unlike most work on generative models, our primary goal is not to train a model that assigns high likelihood to test data, nor do we require the model to be able to learn well without using any labels. Using our new techniques, we achieve state-of-the-art results in semi-supervised classification on MNIST, CIFAR-10 and SVHN. The generated images are of high quality as confirmed by a visual Turing test: our model generates MNIST samples that humans cannot distinguish from real data, and CIFAR-10 samples that yield a human error rate of 21.3%. We also present ImageNet samples with unprecedented resolution and show that our methods enable the model to learn recognizable features of ImageNet classes.
Subfields: Biophysics, Computer Vision and Pattern Recognition, Generative Models, Optimization & Training
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Biophysics
In Computer Vision and Pattern Recognition
- Luc Van Gool
- Dacheng Tao
- Ming-Hsuan Yang
- Trevor Darrell
- Xiaogang Wang
- Jitendra Malik
- Andrew Zisserman
- Bernt Schiele
In Generative Models
- Stefano Ermon
- Jun-Yan Zhu
- Richard Zhang
- Arash Vahdat
- Yoshua Bengio
- Jiaming Song
- Eli Shechtman
- Jan Kautz
In Optimization & Training
- Yoshua Bengio
- Quoc V. Le
- Simon Lacoste-Julien
- Frank Hutter
- Sergey Levine
- Roger Baker Grosse
- Peter Richtárik
- Stefano Ermon
Citations in the corpus
- Semi-Supervised Sequence Modeling with Cross-View Training (2018, background): “…Laine and Aila, 2017; Miyato et al., 2017b) rely on adding continuous noise and applying image-specific transformations like cropping to inputs, GANs (Salimans et al., 2016; Wei et al., 2018) are very difficult to train on text due to its discrete nature, and mixup (Zhang et al., 2018; Verma et…”
- Denoising Diffusion Probabilistic Models (2020, methodology): “On CIFAR10, we calculated Inception and FID scores on 50000 samples using the original code from the OpenAI [48] and TTUR [19] repositories, respectively.”
- Unsupervised Data Augmentation for Consistency Training (2019, background): “Besides the direct extensions, it was found that training neural classifiers to classify out-of-domain examples into an additional class [54] works very well in practice.”
- Self-Training With Noisy Student Improves ImageNet Classification (2019, methodology): “Finally, frameworks in semi-supervised learning also include graph-based methods [84, 73, 77, 33], methods that make use of latent variables as target variables [32, 42, 78] and methods based on low-density separation [21, 58, 15], which might provide complementary benefits to our method.”
- A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark (2019, neutral): “, 2017), IS (Salimans et al., 2016), precision-recall (Sajjadi et al.”
- S4L: Self-Supervised Semi-Supervised Learning (2019, background): “Many of the initial results on semi-supervised learning with deep neural networks were based on generative models such as denoising autoencoders [31], variational autoencoders [15] and generative adversarial networks [27, 33].”
- Are GANs Created Equal? A Large-Scale Study (2017, background): “Inception Score (IS) [24] is based on the fact that a good model should generate samples for which, when evaluated by the classifier, the class distribution has low entropy.”
- Zero-Shot Text-to-Image Generation (2021, methodology): “, 2020), the last of which reports the best Inception Score (Salimans et al., 2016) and Fréchet Inception Distance (Heusel et al.”
- Evaluating Lossy Compression Rates of Deep Generative Models (2020, methodology): “In order to stabilize the training dynamic, we used the gradient penalty (GP) (Salimans et al., 2016).”
- Training Generative Adversarial Networks with Limited Data (2020, background): “42 and inception score (IS) [37] from 9.”
- Your GAN is Secretly an Energy-based Model and You Should use Discriminator Driven Latent Sampling (2020, methodology): “We evaluate the quality and diversity of generated samples via the Inception Score [53] and Fréchet Inception Distance (FID) [54].”
- Your Classifier is Secretly an Energy Based Model and You Should Treat it Like One (2019, background): “Given that we cannot compute normalized likelihoods, we present inception scores (IS) (Salimans et al., 2016) and Frechet Inception Distance (FID) (Heusel et al.”
- Flow Contrastive Estimation of Energy-Based Models (2019, neutral): “24) on Conv-small used in [54, 42] FCE-init 9.”
- Generative Modeling by Estimating Gradients of the Data Distribution (2019, methodology): “For quantitative evaluation, we report inception [48] and FID [20] scores on CIFAR-10 in Tab.”
- Generating Diverse High-Fidelity Images with VQ-VAE-2 (2019, background): “The two most common metrics for comparing GANs are Inception Score [34] and Fréchet Inception Distance (FID) [13].”
- Learning Non-Convergent Non-Persistent Short-Run MCMC Toward Energy-Based Model (2019, background): “Table 1 (a) compares the Inception Score (IS) [41, 3] and Fréchet Inception Distance (FID) [17] with Inception v3 classifier [43] on 40,000 generated examples.”
- BIVA: A Very Deep Hierarchy of Latent Variables for Generative Modeling (2019, background): “The range of applications that come with generative models are vast, where audio synthesis [55] and semi-supervised classification [38, 31, 44] are examples hereof.”
- A Style-Based Generator Architecture for Generative Adversarial Networks (2018, background): “Similar to the inception score [49], the exponentiation brings the values from logarithmic to linear domain so that they are easier to compare.”
- Large Scale GAN Training for High Fidelity Natural Image Synthesis (2018, background): “, 2018) achieves an Inception Score (Salimans et al., 2016) of 52.”