Generative adversarial networks with decoder-encoder output noises.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Generative Adversarial Networks (GANs) are powerful for image generation but face training challenges.
• Standard GANs utilize random noise, requiring generators to map complex distributions, often leading to convergence issues.

Purpose of the Study:

• To introduce a novel deep model, DE-GANs, enhancing GAN performance.
• To address the training difficulties and improve the quality of generated images in GANs.

Main Methods:

• Proposed Generative Adversarial Networks with Decoder-Encoder output noises (DE-GANs).
• Integrated a pre-trained decoder-encoder architecture to transform random noise vectors into informative inputs for the GAN generator.
• Leveraged variational Bayesian inference alongside adversarial training.

Main Results:

• DE-GANs demonstrated improved generator learnability.
• The model achieved higher quality generated images compared to standard GANs.
• Extensive experiments validated the effectiveness of the DE-GANs approach.

Conclusions:

• DE-GANs effectively improve GAN performance by providing more informative inputs to the generator.
• The integration of a decoder-encoder architecture enhances the training stability and output quality of generative models.