A Style-Based Generator Architecture for Generative Adversarial Networks

Karras, Tero; Laine, Samuli; Aila, Timo

doi:10.1109/tpami.2020.2970919

Cited by 2,552 publications

(5,104 citation statements)

References 16 publications

Supporting

Mentioning

5,057

Contrasting

Unclassified

Order By: Relevance

“…Example problems that have, sometimes surprisingly so, turned out to be amenable to ML approaches include text [64] or object [65,66] recognition in images, mapping pictures to textual descriptions of their content [67], machine translation of natural language [68], scoring possible moves in the game of Go [69] and Starcraft [70], and many more. Increasingly, we also see ML methods being applied to problems that do not strictly follow this pattern, such as synthesis of realistic-looking portraits [71].…”

Section: On Machine Learningmentioning

confidence: 99%

SO(8) supergravity and the magic of machine learning

Comşa

Firsching

Fischbacher

2019

J. High Energ. Phys.

106

View full text Add to dashboard Cite

Using de Wit-Nicolai D = 4 N = 8 SO(8) supergravity as an example, we show how modern Machine Learning software libraries such as Google's TensorFlow can be employed to greatly simplify the analysis of high-dimensional scalar sectors of some M-Theory compactifications. We provide detailed information on the location, symmetries, and particle spectra and charges of 192 critical points on the scalar manifold of SO(8) supergravity, including one newly discovered N = 1 vacuum with SO(3) residual symmetry, one new potentially stabilizable non-supersymmetric solution, and examples for "Galois conjugate pairs" of solutions, i.e. solution-pairs that share the same gauge group embedding into SO(8) and minimal polynomials for the cosmological constant. Where feasible, we give analytic expressions for solution coordinates and cosmological constants.As the authors' aspiration is to present the discussion in a form that is accessible to both the Machine Learning and String Theory communities and allows adopting our methods towards the study of other models, we provide an introductory overview over the relevant Physics as well as Machine Learning concepts. This includes short pedagogical code examples. In particular, we show how to formulate a requirement for residual Supersymmetry as a Machine Learning loss function and effectively guide the numerical search towards supersymmetric critical points. Numerical investigations suggest that there are no further supersymmetric vacua beyond this newly discovered fifth solution.At the moment, the N = 8 Supergravity Theory is the only candidate in sight. There are likely to be a number of crucial calculations within the next few years that have the possibility of showing that the theory is no good. If the theory survives these tests, it will probably be some years more before we develop computational methods that will enable us to make predictions and before we can account for the initial conditions of the universe as well as the local physical laws. These will be the outstanding problems for theoretical physics in the next twenty years or so.But to end on a slightly alarmist note, they may not have much more time than that.At present, computers are a useful aid in research, but they have to be directed by human minds. If one extrapolates their recent rapid rate of development, however, it would seem quite possible that they will take over altogether in theoretical physics. So, maybe the end is in sight for theoretical physicists, if not for theoretical physics. S. Hawking, Conclusion of his 1981 Inaugural lecture [1]"Is the end in sight for theoretical physics?"

show abstract

Section: On Machine Learningmentioning

confidence: 99%

SO(8) supergravity and the magic of machine learning

Comşa

Firsching

Fischbacher

2019

J. High Energ. Phys.

106

View full text Add to dashboard Cite

show abstract

“…These results show that our proposed generalization is an effective method to maintain performance in a classification network that suffers from performance degradation due to differences in the intensity of medical images. Recently, structure of generator that greatly improves the quality of the generated image [39,40] and model with advanced few-shot capability [41] are proposed. As future work, Applying these methods to our generalization module would allow the robustness and accuracy of our framework.…”

Section: Resultsmentioning

confidence: 99%

Generalization of intensity distribution of medical images using GANs

Lee

Shin

2020

Hum. Cent. Comput. Inf. Sci.

View full text Add to dashboard Cite

Computer-aided diagnosis (CAD) based on deep learning has already been studied extensively [1]. In particular, many successful studies have applied convolutional neural networks (CNNs) [2, 3] to medical image processing. Studies of the classification of pathology [4-6], lesion segmentation [7-9] and body detection [10-13] using CNNs have been carried out with good performance. However, CNNs learn the intensity of the images. If test an image with a completely different intensity from the learned image, the performance of the CNN greatly degrades. This problem is even more prominent in the medical imaging domain. Unlike the domain of real-world images, medical images are grayscale, and features of lesions in the images can be very detailed and complex. Unfortunately, medical images show different intensity distributions depending on the characteristics of the imaging equipment and the operating methods of the radiologist, and it is very difficult to extract the features of lesions that are valid for all intensity.

show abstract

“…There is a huge potential of development due to the modular nature of the proposed solution. One of the possible improvements to the current pipeline is using the approach published in [9]. The study introduces a new generator architecture that learns how to deeply understand the data, i.e.…”

Section: Discussionmentioning

confidence: 99%

GANKIN: generating Kin faces using disentangled GAN

Ghatas

Hemayed

2020

SN Appl. Sci.

View full text Add to dashboard Cite

Kin image generation from parents' images is a high-level prediction and generation problem. This study presents a new method to predict and generate a kin face using parents' faces, i.e. Tri-subject prediction or two-to-one prediction. Training one end-to-end conditional GAN from scratch can run into mode-collapse and may not converge, we overcome this by using a modular pipeline of unconditional smaller models. We start by extracting father and mother features using a pre-trained model FaceNet, then we use extracted features to predict kin features. After that, we use a linear regression model to convert the predicted features to a latent vector that can be passed to an unconditioned generative adversarial network (GAN) which generates the required kin face. Also, we present a disentangling method to help choosing the age and gender of the generated kin. The modular nature of the proposed model allows validating and improving each part of the pipeline separately. The model achieves promising results compared to the state-of-the-art.

show abstract

A Style-Based Generator Architecture for Generative Adversarial Networks

Cited by 2,552 publications

References 16 publications

SO(8) supergravity and the magic of machine learning

SO(8) supergravity and the magic of machine learning

Generalization of intensity distribution of medical images using GANs

GANKIN: generating Kin faces using disentangled GAN

Contact Info

Product

Resources

About