在我们学习过的基本生成对抗网络 (GAN) 中，有一个“生成器网络”。生成器网络的目标是什么？

（种子/随机 z 是一个叫做生成器网络 G(*) 模型的输入，产生生成/假样本 G(z)。生成/假样本 G(z) 然后作为判别器网络 D(*) 的输入。同时，真实数据样本 x 也作为判别器网络 D(*) 的输入。判别器网络 D(*) 判断它是真的还是假的。）

In the basic Generative Adversarial Network (GAN) we studied in one of the lectures (as illustrated below), there is “Generator Network”. What is the goal of the Generator Network?

(There is Seed/Random z which is an input to the a model called Generator Network G(*) which generates Generated/Fake Sample G(Z). The Generated/Fake Sample G(Z) is then an input to the Discriminator Network D(*). Along with this, Real Data Sample x is also an input to the Discriminator Network D(*). The Discriminator Network D(*) then results in whether it is Real or Fake.)

Question

在我们学习过的基本生成对抗网络 (GAN) 中，有一个“生成器网络”。 生成器网络的目标是什么？

（种子/随机 z 是一个叫做生成器网络  G(*) 模型的输入，产生生成/假样本 G(z)。  生成/假样本 G(z) 然后作为判别器网络 D(*) 的输入。同时，真实数据样本 x 也作为判别器网络 D(*) 的输入。  判别器网络 D(*) 判断它是真的还是假的。）

In the basic Generative Adversarial Network (GAN) we studied in one of the lectures (as illustrated below), there is “Generator Network”. What is the goal of the Generator Network?

(There is Seed/Random z which is an input to the a model called Generator Network G(*) which generates Generated/Fake Sample G(Z). The Generated/Fake Sample G(Z) is then an input to the Discriminator Network D(*). Along with this, Real Data Sample x is also an input to the Discriminator Network D(*). The Discriminator Network D(*) then results in whether it is Real or Fake.)

BlackTom AI · Accepted Answer

The generator maps random noise z into synthetic data samples that approximate the real-data distribution, with the explicit goal of fooling the discriminator into labeling those samples as “real.”

1) Relevant knowledge points and concepts
- GANs consist of two networks: Generator G and Discriminator D. G(z) transforms a random vector z into a sample in data space. D(x) outputs the probability that x is real.
- The training is adversarial: G tries to produce samples that D cannot distinguish from real samples, while D tries to correctly separate real from fake.
- The theoretical optimum is reached when the distribution of G(z) equals the real data distribution; at that point D cannot do better than random guessing.

2) Step-by-step reasoning process
- Step 1: Sample z from a prior (e.g., Gaussian).
- Step 2: Compute a generated sample x_fake = G(z).
- Step 3: Present x_fake and real samples x_real to D; D outputs probabilities D(x_fake) and D(x_real).
- Step 4: Update D to increase D(x_real) and decrease D(x_fake).
- Step 5: Update G to maximize D(x_fake) (i.e., make D more likely to call x_fake “real”).
- Repeat until G generates samples that D cannot reliably distinguish.

3) Intermediate “calculation” steps (loss perspective)
- Original minimax loss: min_G max_D V(D,G) = E_x[log D(x)] + E_z[log(1 − D(G(z)))].
- In practice, G often maximizes E_z[log D(G(z))] to provide stronger gradient early on. Both formulations push G to increase D(G(z)), i.e., to make fakes look real.

4) Why the correct answer is correct
- The generator’s objective is explicitly to synthesize samples that the discriminator cannot tell apart from real ones; that is the adversarial goal encoded by the loss and training loop described above.

5) Why the other options are incorrect
- “Generate a class label for any random input”: that’s a classification task, not what a GAN generator does.
- “Figure out number of feature maps for D”: this is an architecture/hyperparameter design issue, not G’s objective.
- “Pre-process inputs to make them easier to distinguish”: this is opposite to G’s role (it seeks to make fakes harder to distinguish).

Conclusion: The correct answer is “To create/synthesize samples which the Discriminator Network cannot tell apart from real samples,” which matches the provided answer.

类似问题

Which statement about Generative Adversarial Networks (GANs) is TRUE?

Reference Figure 1. Typically early in training, the value of D(G(z)) is closer to which of the following values?

Reference Figure 1. You know that your GAN is trained when D(G(z)) is close to 1

Reference Figure 1. Two cost functions are presented in the figure, which one would you use to train your GAN?

Generative adversarial networks (GANs) learn a parameterized density function, p(x).

Why did Elgammal switch from Generative Adversarial Networks (GAN) to Creative Adversarial Networks (CAN) to create variations of the images he collected for AI artwork development?

Which variation of GANs uses two discriminators: one as a typical binary classifier and the other as a score function?

What is the primary objective of the discriminator in a GAN during training?

更多留学生实用工具

智能学习助手

风格化写作助手

论文查重助手

文献引用助手

课堂转译助手

课堂笔记助手

Quiz搜索助手

学校历年真题

智能刷题助手

智能匹配练习

希望你的学习变得更简单