Empirical Analysis Of Overfitting And Mode Drop In Gan Training

Yazici, Yasin; Foo, Chuan-Sheng; Yap, Kim-Hui; Chandrasekhar, Vijay

doi:10.1109/icip40778.2020.9191083

Cited by 29 publications

(19 citation statements)

References 5 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Although GAN is used to obtain reliable data relatively easily, the performance of GAN itself is also closely related to how much input data it contains 22 . The key problem with small datasets is that the discriminator overfits the training examples 32 . In small data sets feedback to the generator becomes meaningless and training starts to diverge.…”

Section: Discussionmentioning

confidence: 99%

See 1 more Smart Citation

High‐resolution knee plain radiography image synthesis using style generative adversarial network adaptive discriminator augmentation

Ahn

Choi

et al. 2022

Journal Orthopaedic Research

View full text Add to dashboard Cite

In this retrospective study, 10,000 anteroposterior (AP) radiography of the knee from a single institution was used to create medical data set that are more balanced and cheaper to create. Two types of convolutional networks were used, deep convolutional GAN (DCGAN) and Style GAN Adaptive Discriminator Augmentation (StyleGAN2-ADA). To verify the quality of generated images from StyleGAN2-ADA compared to real ones, the Visual Turing test was conducted by two computer vision experts, two orthopedic surgeons, and a musculoskeletal radiologist. For quantitative analysis, the Fréchet inception distance (FID), and principal component analysis (PCA) were used. Generated images reproduced the features of osteophytes, joint space narrowing, and sclerosis. Classification accuracy of the experts was 34%, 43%, 44%, 57%, and 50%. FID between the generated images and real ones was 2.96, which is significantly smaller than another medical data set (BreCaHAD = 15.1). PCA showed that no significant difference existed between the PCs of the real and generated images (p > 0.05). At least 2000 images were required to make reliable images optimally. By performing PCA in latent space, we were able to control the desired PC that show a progression of arthritis. Using a GAN, we were able to generate knee X-ray images that accurately reflected the characteristics of the arthritis progression stage, which neither human experts nor artificial intelligence could discern apart from the real images. In summary, our research opens up the potential to adopt a generative model to synthesize realistic anonymous images that can also solve data scarcity and class inequalities.

show abstract

Section: Discussionmentioning

confidence: 99%

“…22 The key problem with small datasets is that discriminator overfits the training examples. 32 In small data sets feedback to the generator becomes meaningless and training starts to diverge. Medical data is particularly costly and timeconsuming to integrate and standardize in one place compared to other data.…”

Section: Ganspacementioning

confidence: 99%

High‐resolution knee plain radiography image synthesis using style generative adversarial network adaptive discriminator augmentation

Ahn

Choi

et al. 2022

Journal Orthopaedic Research

View full text Add to dashboard Cite

show abstract

“…However, the theoretical analysis conducted in (Arora et al 2017) showed that the training objective can approach its optimum value even though the generated distribution is far from the target distribution. Moreover, GAN models suffer from well-known problems, such as mode dropping or mode collapse (Yazici et al 2020), which affect the quality of the generated sample. Understanding whether GANs can reproduce the same level of similarity exhibited by the original Monte Carlo data set, is essential in the context of scientific simulations.…”

Section: Discussionmentioning

confidence: 99%

Validation of Deep Convolutional Generative Adversarial Networks for High Energy Physics Calorimeter Simulations

Rehm¹,

Vallecorsa²,

Borras³

et al. 2021

Preprint

View full text Add to dashboard Cite

In particle physics the simulation of particle transport through detectors requires an enormous amount of computational resources, utilizing more than 50% of the resources of the CERN Worldwide Large Hadron Collider Grid. This challenge has motivated the investigation of different, faster approaches for replacing the standard Monte Carlo simulations. Deep Learning Generative Adversarial Networks are among the most promising alternatives. Previous studies showed that they achieve the necessary level of accuracy while decreasing the simulation time by orders of magnitudes. In this paper we present a newly developed neural network architecture which reproduces a three-dimensional problem employing 2D convolutional layers and we compare its performance with an earlier architecture consisting of 3D convolutional layers. The performance evaluation relies on direct comparison to Monte Carlo simulations, in terms of different physics quantities usually employed to quantify the detector response. We prove that our new neural network architecture reaches a higher level of accuracy with respect to the 3D convolutional GAN while reducing the necessary computational resources. Calorimeters are among the most expensive detectors in terms of simulation time. Therefore we focus our study on an electromagnetic calorimeter prototype with a regular highly granular geometry, as an example of future calorimeters.

show abstract

“…[1] studies the relationship between GAN memorization and discriminator size, concluding that the distributions learnt by GANs have significantly less support than real distributions. [28] then shows that with fixed latent codes during the training process the authors can achieve GAN memorization.…”

Section: Related Workmentioning

confidence: 97%

When do GANs replicate? On the choice of dataset size

Feng

Guo

Benitez-Quiroz

et al. 2021

2021 IEEE/CVF International Conference on Computer Vision (ICCV)

View full text Add to dashboard Cite

Do GANs replicate training images? Previous studies have shown that GANs do not seem to replicate training data without significant change in the training procedure. This leads to a series of research on the exact condition needed for GANs to overfit to the training data. Although a number of factors has been theoretically or empirically identified, the effect of dataset size and complexity on GANs replication is still unknown. With empirical evidence from BigGAN and StyleGAN2, on datasets CelebA, Flower and LSUN-bedroom, we show that dataset size and its complexity play an important role in GANs replication and perceptual quality of the generated images. We further quantify this relationship, discovering that replication percentage decays exponentially with respect to dataset size and complexity, with a shared decaying factor across GAN-dataset combinations. Meanwhile, the perceptual image quality follows a U-shape trend w.r.t dataset size. This finding leads to a practical tool for one-shot estimation on minimal dataset size to prevent GAN replication which can be used to guide datasets construction and selection.

show abstract

Empirical Analysis Of Overfitting And Mode Drop In Gan Training

Cited by 29 publications

References 5 publications

High‐resolution knee plain radiography image synthesis using style generative adversarial network adaptive discriminator augmentation

High‐resolution knee plain radiography image synthesis using style generative adversarial network adaptive discriminator augmentation

Validation of Deep Convolutional Generative Adversarial Networks for High Energy Physics Calorimeter Simulations

When do GANs replicate? On the choice of dataset size

Contact Info

Product

Resources

About