Zurück zum Glossar

Generative Adversarial Networks (GANs): Structure, Variants, and Use Cases

A Generative Adversarial Network (GAN) is a deep learning model that generates new, realistic-looking data from an existing training dataset. Two neural networks train against each other: one generates synthetic data, while the other evaluates its authenticity. This competitive principle iteratively improves the quality of the generated outputs until synthetic and real data are barely distinguishable.

‍

What is a Generative Adversarial Network?

Architecturally, a GAN consists of two components: the Generator and the Discriminator (also: Discriminator). The Generator receives a random input – described as noise or random variation – and from this, generates synthetic data that mimics the characteristics of the training dataset. This can include images, music, or other data modalities.

‍

The Discriminator evaluates both real samples from the training set and the samples generated by the Generator. It outputs a probability value: a high value means the data is classified as real; a low value indicates a fake.

‍

How Does Adversarial Training Work?

During training, a feedback loop is created between the two networks. The Discriminator is optimized to correctly identify real data and classify synthetic data as fake. Simultaneously, the Generator is trained to deceive the Discriminator as successfully as possible.

‍

Specifically, the Generator attempts to maximize the probability of the Discriminator making a mistake. The Discriminator, in turn, tries to minimize this error probability. Both networks are iteratively adjusted using backpropagation and corresponding loss functions. The result: The Generator produces increasingly convincing data, while the Discriminator learns to recognize subtle differences between real and synthetic output.

‍

For example: In a GAN for dog images, the Generator transforms random noise into dog-like pictures. The Discriminator compares these with real dog photos from the training set. Only when the Discriminator can no longer reliably distinguish the differences are the generated images considered sufficiently realistic.

‍

GAN Variants: An Overview

Various GAN architectures exist, depending on the specific requirements:

     
  • Vanilla GAN: The basic model without specific enhancements.
    •  
  • Conditional GAN (cGAN): Incorporates additional information like class labels into the generation process, enabling more targeted data generation.
    •  
  • Deep Convolutional GAN (DCGAN): The generator and discriminator utilize Convolutional Neural Networks. The generator employs transposed convolutions for upscaling; the discriminator analyzes image details via convolutional layers.
    •  
  • StyleGAN: Generates high-resolution images up to 1024×1024 pixels by employing generator and discriminator layers for different levels of detail.
    •  
  • CycleGAN: For image-to-image translation with unpaired datasets. The generator and discriminator are trained cyclically, ensuring that a reverse translation makes the reconstruction as similar as possible to the original image.
    •  
  • LAPGAN (Laplace Pyramid GAN): A hierarchical method that generates high-quality images across multiple stages.

‍

Practical Examples and Use Cases

GANs are used in various areas:

     
  • Image Generation: New images are generated based on training data – supported by text-based prompts or by editing existing images.
    •  
  • Data Augmentation: GANs generate synthetic training data with the attributes of real data to better train other models.
    •  
  • Data Completion: Missing information in datasets is supplemented, such as dependencies between surface data and underground structures.
    •  
  • Medical Imaging: Realistic 3D organ images can be generated from X-rays and other scans, which are used for surgical planning and simulation.

‍

Conclusion

GANs are competition-based generative models where the generator and discriminator improve each other. The adversarial training process makes it increasingly difficult to distinguish synthetic data from real data. Variants like cGAN, DCGAN, StyleGAN, or CycleGAN adapt the basic architecture to specific requirements – from targeted data generation and stable image generation to image-to-image transformation.