GAN

Tags

machine learningneural networksgenerative modeling

Generative adversarial network (GAN) architecture. A GAN has two parts. The discriminator $D$ acts as a classifier that learns to distinguish fake data produced by the generator $G$ from real data. $G$ incurs a penalty when $D$ detects implausible results. This signal is backpropagated through the generator weights such that $G$ learns to produce more realistic samples over time, eventually fooling the discriminator if training succeeds.

Edit

Open in Overleaf

Download

PNG PNG (HD) PDF SVG TeX

Code

gan.tex (33 lines)

\documentclass[tikz]{standalone}

\usepackage{mathtools}

\usetikzlibrary{calc,positioning}

\begin{document}
\begin{tikzpicture}[
    ->, thick,
    node/.style={circle, fill=teal!60},
    label/.style={below, font=\footnotesize},
  ]

  \node[node] (zin) {$\vec z_\text{in}$};
  \node[node, right=5em of zin] (fake) {$\vec x_\text{fake}$};
  \draw (zin) -- node[above] {$G(\vec x)$} node[label] {generator} (fake);

  \draw[<-] (zin) -- node[above] {$p_\theta(\vec z)$} node[label] {latent noise} ++(-3,0);
  \node[node, above=of fake] (real) {$\vec x_\text{real}$};
  \draw[<-] (real) -- node[above] {$p_\text{data}(\vec x)$} ++(-3,0);
  \node[node, right=6em of fake] (D) at ($(fake)!0.5!(real)$) {$\vec x$};
  \node[right=7em of D] (out) {real?};
  \draw (D) -- node[above] {$D(\vec x)$} node[label] {discriminator} (out);

  \coordinate[right=2.5em of fake, circle, fill, inner sep=0.15em] (pt1);
  \coordinate[right=2.5em of real, circle, fill, inner sep=0.15em] (pt2);

  \draw[-, dashed] (pt1) edge[bend left] coordinate[circle, fill=orange, inner sep=1mm, pos=0.7] (pt3) (pt2);
  \draw (fake) -- (pt1) (real) -- (pt2) (pt3) -- (D);

\end{tikzpicture}
\end{document}