GAN
这篇论文画风都和前面的几篇完全不一样,因为他的任务是生成,本篇文章将作为可选的拓展阅读。
比如说生成一些虚拟的 vtuber 他也是完全办得到的,首先留下经典的文章
Generative Adversarial Networks
GAN 的基本原理其实非常简单,这里以生成图片为例进行说明。假设我们有两个网络,G(Generator)和 D(Discriminator)。正如它的名字所暗示的那样,它们的功能分别是:
- G 是一个生成图片的网络,它接收一个随机的噪声 z,通过这个噪声生成图片,记做 G (z)。
- D 是一个判别网络,判别一张图片是不是 “真实的”。它的输入参数是 x,x 代表一张图片,输出 D(x)代表 x 为真实图片的概率,如果为 1,就代表 100% 是真实的图片,而输出为 0,就代表不可能是真实的图片。
在训练过程中,生成网络 G 的目标就是尽量生成真实的图片去欺骗判别网络 D。而 D 的目标就是尽量把 G 生成的图片和真实的图片分别开来。这样,G 和 D 构成了一个动态的 “博弈过程”。
最后博弈的结果是什么?在最理想的状态下,G 可以生成足以 “以假乱真” 的图片 G (z)。对于 D 来说,它难以判定 G 生成的图片究竟是不是真实的,因此 D (G (z)) = 0.5。
其实就是一个骗一个验证。
提到他我就想提到另一个算法叫 VAE,你可以对比一下两者的相同点和不同点
同时,数学推导要从极大似然估计考虑起!