Generative Adversarial Networks

Ian J. Goodfellow; Jean Pouget-Abadie; Mehdi Mirza; Bing Xu; David Warde-Farley; Sherjil Ozair; Aaron Courville; Yoshua Bengio

Papers · arXiv 1406.2661

Generative Adversarial Networks

Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio

Published 2014-06-10stat.MLcs.LG arXiv PDF

Read with translation and AI

Summary in three lines

🎯 Deep generative models struggled with intractable probabilistic computations, unnormalized partition functions, and reliance on Markov chains.
🔧 Formulates generative modeling as a two-player minimax game between a generator $G$ and a discriminator $D$ , trainable via backpropagation.
📈 Reaches global optimality where $p_{g} = p_{data}$ and achieves Gaussian Parzen window log-likelihood of $225 \pm 2$ on MNIST and $2057 \pm 26$ on TFD.

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.

Reading guide

Problem and motivation

Deep learning achieved dramatic success in discriminative tasks by leveraging backpropagation, dropout, and piecewise linear units. In contrast, deep generative models had far less impact due to severe computational bottlenecks: intractable partition functions, difficult approximate inference, and the need for Markov Chain Monte Carlo (MCMC) methods during training and sample generation. MCMC algorithms frequently suffer from slow mixing between modes. The authors set out to design a generative framework that eliminates the need for Markov chains or approximate inference networks during both training and generation.

Key ideas

Adversarial Framework: Frame generative modeling as a minimax game between two multilayer perceptrons: a generator $G$ (analogous to counterfeiters) and a discriminator $D$ (analogous to police).
Sampling via Forward Propagation: The generator $G (z; θ_{g})$ directly transforms random noise $z \sim p_{z}$ into synthetic data samples using standard feedforward computation.
Direct Gradient Flow: The generator is updated solely via gradients propagated backward through the discriminator rather than directly fitting individual training examples.
Theoretical Connection to Divergences: In the non-parametric limit where $G$ and $D$ have infinite capacity, optimizing the minimax objective minimizes the Jensen–Shannon divergence between the real data distribution $p_{data}$ and the model distribution $p_{g}$ .

Method (with the key equations)

The system consists of generator $G (z; θ_{g})$ mapping latent prior $p_{z} (z)$ to data space, and discriminator $D (x; θ_{d})$ predicting the probability that $x$ originates from $p_{data}$ rather than $p_{g}$ .

The minimax value function is: $G min D max V (D, G) = E_{x \sim p_{data} (x)} [lo g D (x)] + E_{z \sim p_{z} (z)} [lo g (1 - D (G (z)))]$

For any fixed generator $G$ , the optimal discriminator is derived analytically as: $D_{G}^{*} (x) = \frac{p _{data} ( x )}{p _{data} ( x ) + p _{g} ( x )}$

Substituting $D_{G}^{*}$ into $V (D, G)$ yields the virtual training criterion $C (G) = max_{D} V (G, D)$ , which reduces to: $C (G) = - lo g (4) + 2 \cdot J S D (p_{data} ∥ p_{g})$ The global minimum is uniquely achieved when $p_{g} = p_{data}$ , yielding $C (G) = - lo g 4$ and $D^{*} (x) = \frac{1}{2}$ everywhere.

Training Algorithm:

Alternating optimization: train $D$ for $k$ steps ( $k = 1$ used experimentally) using gradient ascent on minibatches of size $m$ : $\nabla_{θ_{d}} \frac{1}{m} i = 1 \sum m [lo g D (x^{(i)}) + lo g (1 - D (G (z^{(i)})))]$
Update $G$ with one gradient descent step. Because $lo g (1 - D (G (z)))$ can saturate early in training when $D$ easily rejects poor generator samples, $G$ is practically updated by maximizing $lo g D (G (z))$ instead.

Results (with the most important numbers)

The authors evaluated adversarial nets on MNIST, the Toronto Face Database (TFD), and CIFAR-10. Because GANs do not offer a closed-form likelihood, performance was quantitatively assessed by fitting a Gaussian Parzen window to generated samples and calculating test set log-likelihood:

MNIST: Adversarial nets achieved a mean log-likelihood of $225 \pm 2$ , outperforming Deep Belief Networks ( $138 \pm 2$ ), Stacked Contractive Autoencoders ( $121 \pm 1.6$ ), and Deep Generative Stochastic Networks ( $214 \pm 1.1$ ).
TFD: Adversarial nets attained $2057 \pm 26$ , competitive with Stacked CAEs ( $2110 \pm 50$ ) and exceeding DBNs ( $1909 \pm 66$ ) and Deep GSNs ( $1890 \pm 29$ ).
Qualitative Generation: Visual inspection of random samples confirmed high quality without memorization (verified by showing nearest training neighbors) and demonstrated smooth latent space interpolations in $z$ .

Limitations and open questions

No Explicit Density Function: The model implicitly samples $p_{g} (x)$ , making exact density evaluation intractable and necessitating heuristic evaluations like Parzen window estimates.
Mode Collapse ("The Helvetica Scenario"): If $G$ is trained excessively without sufficient discriminator updates, $G$ can collapse many distinct $z$ values onto identical $x$ outputs, failing to capture the full diversity of $p_{data}$ .
Training Instability: $D$ and $G$ must remain carefully balanced throughout training to prevent degenerate dynamics or gradient vanishing.

Key terms

Generative Adversarial Nets: A framework for training generative models where a generator network and a discriminator network compete in a two-player minimax game.
Generator ($G$): A differentiable neural network that maps latent prior noise $z$ into the data space to approximate the true data distribution $p_{data}$ .
Discriminator ($D$): A neural network classifier trained to estimate the probability that an input sample came from the real dataset rather than the generator.
Jensen–Shannon Divergence (JSD): A symmetric, bounded measure of distance between two probability distributions; the GAN minimax objective is minimized when the JSD between $p_{data}$ and $p_{g}$ is zero.
Helvetica Scenario: The failure mode (now commonly known as mode collapse) where the generator maps many distinct latent noise vectors $z$ to the exact same data point $x$ due to lack of diversity.
Gaussian Parzen Window: A non-parametric kernel density estimation method used by the authors to approximate test set log-likelihood from generator samples.
Non-saturating game: The practical alternative generator objective of maximizing $lo g D (G (z))$ instead of minimizing $lo g (1 - D (G (z)))$ to prevent gradient vanishing early in training.

Questions

Why replace the original minimax loss $\log(1 - D(G(\bm{z})))$ for the generator?

Early in training, when the generator produces poor samples, the discriminator easily classifies them as fake with high confidence. This causes $lo g (1 - D (G (z)))$ to saturate, yielding very small gradients. Maximizing $lo g D (G (z))$ provides much stronger gradients early on while preserving the same fixed point.

What is the theoretical optimal value of the game?

When $G$ reaches the global optimum ( $p_{g} = p_{data}$ ), the optimal discriminator predicts $D^{*} (x) = \frac{1}{2}$ everywhere, and the value function achieves $- lo g 4$ .

Why were GANs advantageous compared to contemporary generative models like RBMs or GSNs?

They do not require Markov chains for sampling, do not require intractable partition function approximations, avoid explicit inference steps during training, and can be trained purely using backpropagation with modern activations like ReLUs and maxout.

What was the hyperparameter setting for discriminator steps $k$ in the paper?

The authors used $k = 1$ discriminator update per generator update in their experiments as the least computationally expensive option.

The overview on this page was written by AI from the paper and may contain mistakes. Read the paper for the details.

Read with translation and AI