Abstract
With the rapid development of artificial intelligence and deep learning, computer vision-oriented generative models have been widely used. Among them, the generation of adversarial network has the most far-reaching influence. To solve the problem of dynamic instability in the training of generative adversarial networks, this paper proposes a rapid construction method of generative adversarial networks based on hidden layer characterization. The generation process of the adversarial network is divided into two independent generation processes to generate the representation of the experience hidden layer and the final result respectively, so as to stably generate the training dynamics of the adversarial network and capture more data patterns. This method can effectively and stably generate the training dynamics of the adversarial network. Finally, theoretical analysis proves that this method can stably generate the training dynamics of adversarial network and reduce the difficulty of adversarial training. Large-scale experiments on multiple data sets demonstrate the effectiveness of the proposed method.
Publisher
Darcy & Roy Press Co. Ltd.
Reference8 articles.
1. Rikiya Y, Mizuho N, Kinh R, et al. Convolutional neural networks: An overview and application in radiology[J]. Insights Into Imaging, 2018, 9(4): 611-629.
2. Dongsheng A, Yang G, Min Z, et al. Ae-ot-gan: Training gans from data specific latent distribution[C]. European Conference on Computer Vision, 2020, 548-564.
3. Metz L, Poole B, Pfau D, et al. Unrolled generative adversarial networks[C]. 5th International
4. Conference on Learning Representations, Toulon, France, 2017, 1-1.
5. Martin A, Soumith C, Leon B. Wasserstein generative adversarial networks [C] . International con-ference on machine learning, Toulon, France, 2017, 214-223.