# Generative Adversarial Networks

**Authors:** Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio  
**arXiv:** [1406.2661](https://arxiv.org/abs/1406.2661)  
**Date:** 2014-06-10  
**Summary page:** https://www.summarizepaper.com/en/paper/1406.2661v1/

## Key points

- Authors propose a novel framework for estimating generative models using an adversarial process
- Framework involves training two models simultaneously: a generative model G and a discriminative model D
- Training procedure for G is designed to maximize the probability of D making mistakes
- Approach can be viewed as a minimax two-player game between G and D
- Unique solution exists where G recovers training data distribution and D is equal to 1/2 everywhere
- Proposed framework does not require Markov chains or unrolled approximate inference networks
- Entire system can be trained using backpropagation with multilayer perceptrons
- Experiments conducted to evaluate the potential of the framework, involving qualitative and quantitative assessments of generated samples
- Results demonstrate promising outcomes in terms of generating high-quality samples resembling real data

## Summary

In the paper titled "Generative Adversarial Networks," authors Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio propose a novel framework for estimating generative models using an adversarial process. The framework involves training two models simultaneously: a generative model G that captures the data distribution and a discriminative model D that estimates the probability of a sample belonging to the training data rather than G.

The training procedure for G is designed to maximize the probability of D making mistakes. This approach can be viewed as a minimax two-player game between G and D. In this game, there exists a unique solution in the space of arbitrary functions G and D, where G successfully recovers the training data distribution and D is equal to 1/2 everywhere. Unlike traditional methods that rely on Markov chains or unrolled approximate inference networks, the proposed framework does not require their usage during either training or sample generation. Instead, the entire system can be trained using backpropagation when G and D are defined by multilayer perceptrons. To evaluate the potential of this framework, experiments were conducted involving qualitative and quantitative assessments of the generated samples. The results demonstrate promising outcomes in terms of generating high-quality samples that resemble real data. Overall, this paper presents an innovative approach to estimating generative models through an adversarial process. By simultaneously training both a generative model and a discriminative model without relying on complex techniques like Markov chains or unrolled approximate inference networks, it offers a unique solution with effective results in generating realistic samples as validated by experimental results.