Introduction

Generative Adversarial Networks (GANs) are deep learning models that can generate new data, such as pictures, that look very similar to real ones. A GAN has two main parts:

1. Generator : makes fake images that look real.

2. Discriminator : checks if an image is real or fake.

These two networks are trained together in what we call a game-theoretic setup. This means they are like two players in a game. The generator’s goal is to make the discriminator believe its fake images are real, while the discriminator’s goal is to correctly guess which images are real and which are fake. Over time, both networks become better the generator makes more realistic images, and the discriminator becomes better at detecting fakes.

Architecture and Training

In most GANs used for making images, both the generator and discriminator are built using Convolutional Neural Networks (CNNs). CNNs are good at working with pictures because they can detect patterns like edges, colors, and textures.

How the Generator Works

The generator starts with a random vector of numbers, usually called noise. Let’s call this vector zzz. The generator turns this noise into an image using several transposed convolutional layers, which slowly build up image features.

Mathematically, we can say: Architecture and Training

In most GANs used for making images, both the generator and discriminator are built using Convolutional Neural Networks (CNNs). CNNs are good at working with pictures because they can detect patterns like edges, colors, and textures.

How the Generator Works

The generator starts with a random vector of numbers, usually called noise. Let’s call this vector zzz. The generator turns this noise into an image using several transposed convolutional layers, which slowly build up image features.

Mathematically, we can say: Architecture and Training

In most GANs used for making images, both the generator and discriminator are built using Convolutional Neural Networks (CNNs). CNNs are good at working with pictures because they can detect patterns like edges, colors, and textures.

How the Generator Works

The generator starts with a random vector of numbers, usually called noise. Let’s call this vector zzz. The generator turns this noise into an image using several transposed convolutional layers, which slowly build up image features.

Mathematically, we can say: G(z; θ_g) → x_fake

Here:

• G is the generator function

• θ_g are the generator’s weights (parameters)

• x_fake is the generated image

How the Discriminator Works

The discriminator takes an image as input (real or fake) and outputs a number between 0 and 1, which shows the probability that the image is real:

D(x; θ_d) -> y

Here:

• D is the discriminator function

• θ_d are its weights

• y is the predicted probability of being real

Training Process

The generator and discriminator are trained together with opposite goals. The discriminator wants:

D(x_real) = 1
D(G(z)) = 0

The generator wants:

D(G(z)) = 1

The GAN loss function is written as:

min_G max_D V(D, G) = E_{x ~ p_data(x)} [log D(x)] + E_{z ~ p_z(z)} [log(1 - D(G(z)))]

• min_G means the generator tries to minimize the value

• max_D means the discriminator tries to maximize the value

• E_{x ~ p_data(x)} is the expected value over real images

• E_{z ~ p_z(z)} is the expected value over random noise

This creates a feedback loop: the generator improves at making realistic images, and the discriminator improves at detecting fakes.

Applications of GANs

GANs can make very realistic images of faces, animals, buildings, and landscapes. Some examples:

1. Creating realistic faces that do not belong to real people

2. Generating cities or landscapes that look real but are fake

3. Image-to-image translation, like turning sketches into photos or daytime images into nighttime images

They are used in:

• Movies and games for special effects

• Medicine for generating safe, synthetic medical images

• Self-driving cars for creating training images for AI

GANs are exciting because they can create new data, not just recognize it. They are useful in many areas but also come with challenges. As researchers improve GANs, they will become easier to train, more stable, and safer to use. GANs show how AI can learn to create, not just classify.