Improving Multi-Agent Generative Adversarial Nets with Variational Latent Representation

• Main Authors: Huan Zhao, Tingting Li, Yufeng Xiao, Yu Wang
• Format: Article
• Language: English
• Published: MDPI AG 2020-09-01
• Series: Entropy
• Subjects: diversity; generative adversarial networks; mode collapsing; multi-agent generator; quality; variable auto-encoder
• Online Access: https://www.mdpi.com/1099-4300/22/9/1055

Generative adversarial networks (GANs), which are a promising type of deep generative<br />network, have recently drawn considerable attention and made impressive progress. However,<br />GAN models suffer from the well-known problem of mode collapse. This study focuses on this<br />challenge and introduces a new model design, called the encoded multi-agent generative adversarial<br />network (E-MGAN), which tackles the mode collapse problem by introducing the variational latent<br />representations learned from a variable auto-encoder (VAE) to a multi-agent GAN. The variational<br />latent representations are extracted from training data to replace the random noise input of the<br />general multi-agent GANs. The generator in E-MGAN employs multiple generators and is penalized<br />by a classifier. This integration guarantees that the proposed model not only enhances the quality of<br />generated samples but also improves the diversity of generated samples to avoid the mode collapse<br />problem. Moreover, extensive experiments are conducted on both a synthetic dataset and two<br />large-scale real-world datasets. The generated samples are visualized for qualitative evaluation. The<br />inception score (IS) and Fréchet inception distance (FID) are adopted to measure the performance<br />of the model for quantitative assessment. The results confirmed that the proposed model achieves<br />outstanding performances compared to other state-of-the-art GAN variants.