Attentive evolutionary generative adversarial network
- PDF / 4,917,286 Bytes
- 15 Pages / 595.276 x 790.866 pts Page_size
- 0 Downloads / 247 Views
Attentive evolutionary generative adversarial network Zhongze Wu 1 & Chunmei He 2 & Liwen Yang 1 & Fangjun Kuang 3
# Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Generative adversarial network (GAN) is an effective method to learn generative models from real data. But there are some drawbacks such as instability, mode collapse and low computational efficiency in the existing GANs. In this paper, attentive evolutionary generative adversarial network (AEGAN) model is proposed in order to improve these disadvantages of GANs. The modified evolutionary algorithm is designed for the AEGAN. In the AEGAN the generator evolves continuously to resist the discriminator by three independent mutations at every batch and only the well-performing offspring (i.e.,the generators) can be preserved at next batch. Furthermore, a normalized self-attention (SA) mechanism is embedded in the discriminator and generator of AEGAN to adaptively assign weights according to the importance of features. We propose careful regulation of the generators evolution and an effective weight assignment to improve diversity and long-range dependence. We also propose a superior training algorithm for AEGAN. With the algorithm, the AEGAN overcomes the shortcomings of traditional GANs brought by single loss function and deep convolution and it greatly improves the training stability and statistical efficiency. Extensive image synthesis experiments on CIFAR-10, CelebA and LSUN datasets are presented to validate the performance of AEGAN. Experimental results and comparisons with other GANs show that the proposed model is superior to the existing models. Keywords Generative adversarial network . Attention mechanism . Evolutionary algorithm . Image generation
1 Introduction Generative Adversarial Network (GAN) achieves great success on synthesizing real-world images and modelings. Since GAN is proposed by Good fellow et al. [1], it is widely applied in various areas, including image vision calculating [2, 3], language processing [4] and message security [5]. GAN is designed to learn the possible distribution of real data and synthesize similar data based on the same distribution. Various derived models based on GANs are proposed by many researchers. Mao, Xudong et al. [6] propose LSGAN, which presents least squares loss function to overcome the Zhongze Wu and Chunmei He contributed equally to this work. * Chunmei He [email protected] 1
School of Physics and optoelectronics, Xiangtan University, Hunan411105, Xiangtan, China
2
School of Computer Science & School of Cyberspace Security, Xiangtan University, Hunan411105, Xiangtan, China
3
School of Information Engineering, Wenzhou Business College, Zhejiang325035, Wenzhou, China
problems of vanishing gradients and training unstable in GAN. I. Gulrajani et al. [7] propose Wasserstein GAN gradient penalty (WGAN-GP) and use Wasserstein distance and gradient-penalty (GP) in adversarial training goals. Radford et al. [8] propose Deep Convolution GAN (DCGAN), which has stronger
Data Loading...