What Is a Generative Adversarial Network (GAN)? How the Generator and Discriminator Compete

What Is a GAN, and What Are Its Core Components?

A generative adversarial network (GAN) is a machine learning model built from two paired networks that together generate realistic data by learning patterns from existing training datasetsCITE:E7. Goodfellow and colleagues, in the 2014 paper that introduced the framework, describe estimating a generative model through an adversarial process in which two models are trained simultaneously: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability a given sample came from the training data rather than from GCITE:E1. Google's machine learning documentation frames the same pairing in functional terms: the generator learns to produce the target output, while the discriminator learns to distinguish true data from the generator's outputCITE:E4.

How Do the Generator and Discriminator Compete, and What Are Their Training Objectives?

The generator tries to fool the discriminator, and the discriminator tries to keep from being fooled, according to Google's description of the two networks' rolesCITE:E5. Goodfellow and colleagues formalize this rivalry mathematically: the training procedure for G is to maximize the probability that D makes a mistake, and the overall framework corresponds to a minimax two-player gameCITE:E2. In other words, the generator's objective and the discriminator's objective are defined in direct opposition to each other, with each network's progress measured by the other network's failure rate.

What Is GAN's Theoretical Foundation, and When Is the Optimal Solution Reached?

Goodfellow and colleagues proved that, in the space of arbitrary functions G and D, a unique solution exists: G recovers the training data distribution exactly, and D equals 1/2 everywhereCITE:E3. A discriminator output of 1/2 across all inputs means the discriminator has been reduced to random guessing — it can no longer tell whether a given sample is real or generated. This equilibrium point is the theoretical target that the minimax training procedure described above is designed to reachCITE:E2CITE:E3.

How Does GAN Perform in Practical Applications Such as Image Generation?

Google's documentation states that GANs can create images that look like photographs of human faces, even though the faces don't belong to any real personCITE:E6. This is a direct illustration of the generator/discriminator pairing described above applied to image data: the generator produces the target output (a face image) and the discriminator's role is to distinguish that output from true dataCITE:E4CITE:E6.

What Does This Mean?

The practical outcome described by Google — faces that resemble photographs but belong to no real personCITE:E6 — lines up with the theoretical equilibrium Goodfellow and colleagues proved for the framework, in which the discriminator ends up unable to distinguish real data from generated dataCITE:E3. The minimax objective, in which the generator is trained specifically to maximize the discriminator's error rateCITE:E2, is the mechanism connecting the 2014 theoretical result to the kind of output shown in later applied documentation.