Timezone: »
Noise injection is an effective way of circumventing overfitting and enhancing generalization in machine learning, the rationale of which has been validated in deep learning as well. Recently, noise injection exhibits surprising effectiveness when generating high-fidelity images in Generative Adversarial Networks (GANs) (e.g. StyleGAN). Despite its successful applications in GANs, the mechanism of its validity is still unclear. In this paper, we propose a geometric framework to theoretically analyze the role of noise injection in GANs. First, we point out the existence of the adversarial dimension trap inherent in GANs, which leads to the difficulty of learning a proper generator. Second, we successfully model the noise injection framework with exponential maps based on Riemannian geometry. Guided by our theories, we propose a general geometric realization for noise injection. Under our novel framework, the simple noise injection used in StyleGAN reduces to the Euclidean case. The goal of our work is to make theoretical steps towards understanding the underlying mechanism of state-of-the-art GAN algorithms. Experiments on image generation and GAN inversion validate our theory in practice.
Author Information
Ruili Feng (USTC)
Deli Zhao (Alibaba Group)
Zheng-Jun Zha (University of Science and Technology of China)
Related Events (a corresponding poster, oral, or spotlight)
-
2021 Spotlight: Understanding Noise Injection in GANs »
Thu. Jul 22nd 12:25 -- 12:30 AM Room
More from the Same Authors
-
2023 : Latent Space Editing in Transformer-Based Flow Matching »
Tao Hu · David Zhang · Meng Tang · Pascal Mettes · Deli Zhao · Cees Snoek -
2023 Poster: Cones: Concept Neurons in Diffusion Models for Customized Generation »
Zhiheng Liu · Ruili Feng · Kai Zhu · Yifei Zhang · Kecheng Zheng · Yu Liu · Deli Zhao · Jingren Zhou · Yang Cao -
2023 Poster: RLEG: Vision-Language Representation Learning with Diffusion-based Embedding Generation »
Liming Zhao · Kecheng Zheng · Yun Zheng · Deli Zhao · Jingren Zhou -
2023 Poster: Composer: Creative and Controllable Image Synthesis with Composable Conditions »
Lianghua Huang · Di Chen · Yu Liu · Yujun Shen · Deli Zhao · Jingren Zhou -
2023 Oral: Cones: Concept Neurons in Diffusion Models for Customized Generation »
Zhiheng Liu · Ruili Feng · Kai Zhu · Yifei Zhang · Kecheng Zheng · Yu Liu · Deli Zhao · Jingren Zhou · Yang Cao -
2023 Poster: Random Shuffle Transformer for Image Restoration »
Jie Xiao · Xueyang Fu · Man Zhou · Hongjian Liu · Zheng-Jun Zha -
2022 Poster: Principled Knowledge Extrapolation with GANs »
Ruili Feng · Jie Xiao · Kecheng Zheng · Deli Zhao · Jingren Zhou · Qibin Sun · Zheng-Jun Zha -
2022 Spotlight: Principled Knowledge Extrapolation with GANs »
Ruili Feng · Jie Xiao · Kecheng Zheng · Deli Zhao · Jingren Zhou · Qibin Sun · Zheng-Jun Zha -
2022 Poster: Region-Based Semantic Factorization in GANs »
Jiapeng Zhu · Yujun Shen · Yinghao Xu · Deli Zhao · Qifeng Chen -
2022 Spotlight: Region-Based Semantic Factorization in GANs »
Jiapeng Zhu · Yujun Shen · Yinghao Xu · Deli Zhao · Qifeng Chen -
2021 Poster: Uncertainty Principles of Encoding GANs »
Ruili Feng · Zhouchen Lin · Jiapeng Zhu · Deli Zhao · Jingren Zhou · Zheng-Jun Zha -
2021 Spotlight: Uncertainty Principles of Encoding GANs »
Ruili Feng · Zhouchen Lin · Jiapeng Zhu · Deli Zhao · Jingren Zhou · Zheng-Jun Zha