Towards the Gradient Vanishing, Divergence Mismatching and Mode Collapse of Generative Adversarial Nets

Zhang, Zhaoyu; Luo, Changwei; Yu, Jun

doi:10.1145/3357384.3358081

Cited by 8 publications

(5 citation statements)

References 1 publication

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The vanishing gradients problem is another significant challenge encountered during the training phase of GANs. This issue emerges due to the complex architecture of GANs, where both G and D need to maintain a balance and learn collaboratively [221]. During the training process, as gradients are backpropagated through the layers of the network, they can diminish drastically, leading to stagnancy in learning.…”

Section: Vanishing Gradientsmentioning

confidence: 99%

Ten years of generative adversarial nets (GANs): a survey of the state-of-the-art

Chakraborty,

Reddy K S,

Naik

et al. 2024

Mach. Learn.: Sci. Technol.

View full text Add to dashboard Cite

Generative Adversarial Networks (GANs) have rapidly emerged as powerful tools for generating realistic and diverse data across various domains, including computer vision and other applied areas, since their inception in 2014. Consisting of a discriminative network and a generative network engaged in a Minimax game, GANs have revolutionized the field of generative modeling. In February 2018, GAN secured the leading spot on the ``Top Ten Global Breakthrough Technologies List'' issued by the Massachusetts Science and Technology Review. Over the years, numerous advancements have been proposed, leading to a rich array of GAN variants, such as conditional GAN, Wasserstein GAN, CycleGAN, and StyleGAN, among many others. This survey aims to provide a general overview of GANs, summarizing the latent architecture, validation metrics, and application areas of the most widely recognized variants. We also delve into recent theoretical developments, exploring the profound connection between the adversarial principle underlying GAN and Jensen-Shannon divergence while discussing the optimality characteristics of the GAN framework. The efficiency of GAN variants and their model architectures will be evaluated along with training obstacles as well as training solutions. In addition, a detailed discussion will be provided, examining the integration of GANs with newly developed deep learning frameworks such as Transformers, Physics-Informed Neural Networks, Large Language models, and Diffusion models. Finally, we reveal several issues as well as future research outlines in this field.

show abstract

Section: Vanishing Gradientsmentioning

confidence: 99%

Ten years of generative adversarial nets (GANs): a survey of the state-of-the-art

Chakraborty,

Reddy K S,

Naik

et al. 2024

Mach. Learn.: Sci. Technol.

View full text Add to dashboard Cite

show abstract

“…2. Inspired by ResBlock [52][53][54], DEM consists of a main convolutional stream and two parallel residual streams. To preserve the most fundamental features, the mainstream passes through two 3 × 3 convolutions and one 1 × 1 convolution, where the 3 × 3 convolutions are with the LReLU activation function.…”

Section: Detail Enhancement Module (Dem)mentioning

confidence: 99%

BDPartNet: Feature Decoupling and Reconstruction Fusion Network for Infrared and Visible Image

Wang,

Zhang,

Tao

et al. 2024

CMC

View full text Add to dashboard Cite

While single-modal visible light images or infrared images provide limited information, infrared light captures significant thermal radiation data, whereas visible light excels in presenting detailed texture information. Combining images obtained from both modalities allows for leveraging their respective strengths and mitigating individual limitations, resulting in high-quality images with enhanced contrast and rich texture details. Such capabilities hold promising applications in advanced visual tasks including target detection, instance segmentation, military surveillance, pedestrian detection, among others. This paper introduces a novel approach, a dual-branch decomposition fusion network based on AutoEncoder (AE), which decomposes multi-modal features into intensity and texture information for enhanced fusion. Local contrast enhancement module (CEM) and texture detail enhancement module (DEM) are devised to process the decomposed images, followed by image fusion through the decoder. The proposed loss function ensures effective retention of key information from the source images of both modalities. Extensive comparisons and generalization experiments demonstrate the superior performance of our network in preserving pixel intensity distribution and retaining texture details. From the qualitative results, we can see the advantages of fusion details and local contrast. In the quantitative experiments, entropy (EN), mutual information (MI), structural similarity (SSIM) and other results have improved and exceeded the SOTA (State of the Art) model as a whole.

show abstract

“…The generator accepts a random noise vector z and learns the data distribution to generate an image G(z), while the discriminator differentiates if G(z) is true or not, and finally the image is expected to be generated by the generator to fool the discriminator with a fake. However, the actual training results may not be able to achieve the ideal situation, and there will be problems such as gradient disappearance [11] and mode collapse [12]. Therefore, most of the current research focuses on two directions.…”

Section: Generative Adversarial Networkmentioning

confidence: 99%

Text to multi-object images synthesis based on non-local self-attention

Wang¹,

Yang²

2023

International Conference on Computer, Artificial Intelligence, and Control Engineering (CAICE 2023)

View full text Add to dashboard Cite

Deep convolutional network (CNN) can make the pictures generated by GAN more reasonable, but limited by the local receptive field of CNN, there are still many unreasonable places in the authenticity and semantics of the multi-object images generated from text. Therefore, a GAN-based method that incorporates a non-local self-attention mechanism is proposed. By embedding a non-local self-attention structure in the network, the network obtains global semantic information and detailed features, and uses the obtained information to perform level-by-level encoding to generate the final relatively reasonable image. The amount of parameters and calculation of the entire model is also reduced a lot. The proposed method is verified on the public COCO-stuff dataset and uses multiple indicators such as Inception Score, FID and classification accuracy score to evaluate the authenticity and diversity of the generated images. Experimental results show that the quality of the generated images is superior to that of previously proposed methods.

show abstract

Towards the Gradient Vanishing, Divergence Mismatching and Mode Collapse of Generative Adversarial Nets

Cited by 8 publications

References 1 publication

Ten years of generative adversarial nets (GANs): a survey of the state-of-the-art

Ten years of generative adversarial nets (GANs): a survey of the state-of-the-art

BDPartNet: Feature Decoupling and Reconstruction Fusion Network for Infrared and Visible Image

Text to multi-object images synthesis based on non-local self-attention

Contact Info

Product

Resources

About