Notebook

Appearance

Neural Network ​

In Deep_learning, we use an Artificial Neural Network (ANN) as architecture to solve Machine Learning problems. An ANN imitates a brain by connecting perceptrons (i.e. artificial neurons) together.

Generative Adversarial Network ​

Example of GAN

In a Generative Adversarial Network (GAN), 2 neural networks compete with each other.

The generator will attempt to create a synthetic output (e.g. an image) from random noise. The discriminator compare the result with a real dataset (usually unsupervised) and try to determine the fake result.

Backpropagation process apply on both models so they can improve: the generator create better samples, the discriminator become better at finding them.

Variants ​

Here some prominent GAN variants:

  • DCGAN (Deep Convolutional GAN): Uses convolutional layers for better image generation.
  • WGAN (Wasserstein GAN): Improves training stability with a different loss function.
  • CycleGAN: Translates images between different domains (e.g., horse to zebra).
  • StyleGAN: Known for generating high-quality, photorealistic faces.
  • BigGAN: Scales up GANs for generating high-resolution images.

Resources ​