Research

Generative Adversarial Nets: The Paper That Introduced GANs

GANs train two neural networks in competition: a generator creates synthetic samples, while a discriminator learns to distinguish them from real data.

Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua BengioJun 10, 2014Advances in Neural Information Processing Systems 27 (NeurIPS 2014)12 min readUniversité de Montréal
FOUNDATIONALGenerative AIIntermediateDeep LearningMachine Learning

Key takeaway

The generator maps random noise to synthetic data. The discriminator predicts whether an input is real or generated. Their competing objectives create a minimax game; ideally, generated samples become indistinguishable from real data. The original work demonstrated this approach on MNIST, the Toronto Face Database and CIFAR-10.

Updated on Nov 17, 2026

01

Abstract

We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to 1/2 everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples.

02

Methodology

The Core Idea

Think of a counterfeiter and a detective. The generator (G) is the counterfeiter: it transforms random noise into synthetic data. The discriminator (D) is the detective: it receives real and generated examples and estimates whether each came from the training set. As D improves, G must produce more convincing samples; as G improves, D must become more discerning.

Figure 1

03

Architecture

Model Architecture

The original paper uses two multilayer perceptrons. Generator: samples a latent vector z from a simple prior and maps it to a data sample, x = G(z). Discriminator: takes x and outputs D(x), the estimated probability that x came from real training data. During training, D sees both real examples and generated examples. At inference time, generation requires only a forward pass through G.

Original GAN Training Illustration

The original paper’s Figure 1 explains the adversarial dynamics: the discriminator learns to separate the real distribution from the generator distribution, while gradients from the discriminator guide the generator toward regions that look more like real data. At equilibrium, the distributions match and the discriminator outputs 1/2.\n\nFigure source: Goodfellow et al. (2014), Figure 1.

04

Training

How to Implement a Basic GAN

A minimal implementation uses two neural networks and alternating optimization:\n\n1. Sample a minibatch of real examples and latent noise.\n2. Generate fake samples with G.\n3. Train D to label real samples as real and generated samples as fake.\n4. Sample a new noise minibatch.\n5. Freeze D’s parameters for the generator update, pass G(z) through D, and train G to make D predict real.\n6. Repeat while monitoring sample diversity and both networks’ losses.\n\nFor a modern image baseline, use convolutional networks (for example, a DCGAN-style architecture), stable normalization and carefully balanced learning rates. This is a practical extension, not the exact MLP architecture used in the original experiments.

05

Results

Experiments and Findings

The authors evaluated the approach on MNIST, the Toronto Face Database (TFD) and CIFAR-10. The generator used rectifier and sigmoid activations; the discriminator used maxout activations and dropout. They reported Parzen-window log-likelihood estimates, while explicitly noting that this evaluation method had high variance and limitations in high-dimensional spaces. The paper demonstrated that adversarial training could produce competitive samples without Markov-chain sampling.

06

Limitations

Limitations and Training Challenges

GAN training is a balancing problem: if the discriminator becomes too strong or the generator changes too quickly, useful gradients can disappear. The original paper identifies a lack of an explicit probability-density representation and the need to keep the two networks synchronized. It also describes a failure mode in which many latent inputs map to too few outputs—an early description of what is now commonly called mode collapse. The ideal convergence result relies on assumptions that practical neural networks do not fully satisfy.

07

Conclusion

Why This Paper Matters

The paper established adversarial learning as a major branch of generative modeling. Later work extended the idea with conditional GANs, convolutional architectures, image-to-image translation and high-fidelity image synthesis. GANs are no longer the only dominant generative approach, but their training framework remains an essential part of modern generative-AI history.

Continue reading

Related research

Browse all research