What Are GANs (Generative Adversarial Networks)?

The Background

Ian Goodfellow, a PhD student at the Université de Montréal, introduced GANs (Generative Adversarial Nets) to the world in 2014 through a research paper titled “Generative Adversarial Nets.” Instead of relying solely on existing datasets, GANs could generate new data that mimicked real datasets. This innovation would revolutionize data collection, making it faster and cheaper. Furthermore, GANs could create synthetic data that resembled sensitive information without compromising security, which was a game-changer for fields requiring high levels of privacy.

How Do GANs Work?

Imagine GANs as a clever duo of neural networks: the Generator and the Discriminator. These two networks engage in a dynamic game of cat and mouse, each trying to outsmart the other.

The Generator

The Generator’s job is to create new data. It starts with a random vector, called a latent vector, which is just a bunch of numbers with no meaningful pattern. The Generator then transforms this random data through a series of layers, applying patterns it has learned from the real dataset. These layers are like stages in a factory, each one refining the raw material a little more. By the time the data emerges from the final layer, it closely resembles the original dataset.

The Discriminator

On the other side of the equation is the Discriminator. Its role is to evaluate the data produced by the Generator. The Discriminator is trained on both real data and the synthetic data created by the Generator. It analyzes this data through multiple layers of its own neural network, ultimately producing a probability score between 0 and 1. This score indicates how similar the synthesized data is to the real data — the higher the score, the closer the resemblance.


The real magic of GANs happens during training. The Generator and Discriminator are in constant competition. The Generator strives to create data that the Discriminator can’t distinguish from real data, while the Discriminator continually improves its ability to tell the difference. This tug-of-war results in both networks becoming increasingly sophisticated. The Generator gets better at creating realistic data, and the Discriminator becomes more adept at spotting fakes.

Loss Function

The GAN’s ability can be mathematically represented by a loss function which measures how well the Discriminator can differentiate between real and fake data:

loss discriminator = -[log(D(x))+log(1−D(G(z)))]

In this formula, D(x) is the Discriminator’s probability score for real data, and D(G(z)) is its score for the generated data. The Generator aims to minimize this loss, pushing it to create data that more closely resembles the original.

Here is a simple diagram of how GANs work. Start at Random Noise and follow the cycle.


The implications of GANs are vast and exciting. They can be used to create art, music, and videos that are eerily similar to existing works, offering new tools for artists and creatives. They can generate synthetic data to enhance training datasets, fill in missing data, or even create entirely new datasets for various applications.

Currently, one of the largest uses of GANs by the general public is to create art. For example, deep fake videos, AI-generated or edited photos, and style transfer are widely known. In various industries, professionals use GANs to create or enhance datasets. Various AI-related companies are also using GANs to produce text or audio with the help of an NLP (Natural Language Processor).

Here’s an example of a blurry image versus one that a GAN has restored. The discriminator compares the restored image to high-quality images from its training since it doesn’t have the high-quality version of this exact picture. This way, the GAN uses its prior knowledge to determine if the restored image is clear.


  • Generative Adversarial Networks (GANs) have revolutionized how we generate and enhance data.
  • There are two nueral networks within each GAN: one that generates data (the Generator), while the other evaluates its authenticity (the Discriminator).
  • Through competition and refinement, GANs produce synthetic data that closely resembles real datasets, speeding up data collection, preserving privacy, and synthesizing brand-new data.
  • Their applications span from art creation to data enhancement across industries.
  • GANs showcase the power of adversarial training, pushing the boundaries of AI’s capabilities.

Extra Resources

Please leave any suggestions or questions in the comments! Stay tuned for more . . .

One Reply to “What Are GANs (Generative Adversarial Networks)?”

Leave a Reply

Your email address will not be published. Required fields are marked *