我们提出了一种通过对抗过程估计生成模型的新框架。在该框架中，我们同时训练两个模型：一个生成模型 G，用于捕获数据分布；一个判别模型 D，用于估计样本来自训练数据而非 G 的概率。G 的训练过程是最大化 D 犯错误的概率。该框架对应于一个极小极大两人博弈。在任意函数 G 和 D 的空间中，存在唯一解，其中 G 恢复训练数据分布，而 D 处处等于 1/2。当 G 和 D 由多层感知机定义时，整个系统可以通过反向传播进行训练。在训练或生成样本期间，无需任何马尔可夫链或展开的近似推理网络。实验通过对生成样本的定性和定量评估，展示了该框架的潜力。
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.
核心贡献 · Key contributions
提出生成器与判别器网络的对抗训练框架。 Proposes adversarial training framework with generator and discriminator networks.
证明生成器匹配数据分布、判别器输出 1/2 时达到全局最优。 Shows global optimum when generator matches data distribution, discriminator outputs 1/2.
训练和采样无需马尔可夫链或近似推断。 Eliminates need for Markov chains or approximate inference during training or sampling.
在 MNIST、TFD 和 CIFAR-10 上展示有竞争力的样本质量。 Demonstrates competitive sample quality on MNIST, TFD, and CIFAR-10.
引入极小极大博弈公式用于生成模型估计。 Introduces minimax game formulation for generative model estimation.
局限 · Limitations
没有概率密度 p_g(x)的显式表示。 No explicit representation of probability density p_g(x).
训练需要 G 和 D 仔细同步以避免模式崩溃。 Training requires careful synchronization between G and D to avoid mode collapse.
理论保证假设无限容量；实际 MLP 可能存在临界点。 Theoretical guarantees assume infinite capacity; practical MLPs may have critical points.
通过 Parzen 窗估计似然方差大，高维性能差。 Likelihood estimation via Parzen window has high variance and poor high-dimensional performance.
框架限于可微的生成器和判别器函数。 Framework limited to differentiable generator and discriminator functions.
论文章节 · Sections（共 10）
摘要Abstract
1 引言1 Introduction
2 相关工作2 Related work
3 对抗网络3 Adversarial nets
4 理论结果4 Theoretical Results
4.1 pg=pdata 的全局最优性4.1 Global Optimality of pg=pdatasubscript𝑝𝑔subscript𝑝datap_{g}=p_{\text{data}}