Optimizing Latent Distributions for Non-Adversarial Generative Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Generative Adversarial Networks (GANs) use a discriminator to guide generator training for high-quality image synthesis.
• GAN training can be unstable, making generator convergence difficult.
• Limited real data may not fully capture true data distributions.

Purpose of the Study:

• To develop a stable, non-adversarial approach for training generative models.
• To address the limitations of insufficient real data representation in generative modeling.
• To propose an alternative to adversarial training in generative networks.

Main Methods:

• Investigated optimizing generators over a set of distributions within a Wasserstein ball centered on the training data distribution.
• Developed a tractable reformulation for the proposed objective function.
• Analyzed theoretical properties and connections/differences with GANs.

Main Results:

• Demonstrated theoretical solvability of the new objective function.
• Showcased effective image generator learning using the non-adversarial approach.
• Achieved generated image quality comparable to that of GANs on real-world datasets.

Conclusions:

• The proposed non-adversarial generative network offers a stable and effective alternative to GANs.
• This method successfully generates high-quality images even with limited training data.
• The approach provides a viable path for robust generative model training.