4
$\begingroup$

enter image description here

What is the meaning of $V(D,G)$? How do we get these expectation parts? I was trying to understand it following this article: Understanding Generative Adversarial Networks (D.Seita), but, after many tries, I still can't understand how he got from $\sum_{n=1}^{N} \log D(x)$ to $\mathbb{E}(\log(D(x))$.

$\endgroup$
  • $\begingroup$ Welcome to SE:AI! $\endgroup$ – DukeZhou Apr 12 at 21:11
  • $\begingroup$ I only glanced at the article, but it sounds like 1/2 of the data is coming from each source, so y_i was replaced with 1/2. To get expectation of a random variable you sum over all values of the variable and multiply each value by the probability of it occurring. Since p(.) = 0.5, then by summing over the value of each variable D and multiplying by the probability, it yields an expectation. $\endgroup$ – Hanzy Apr 13 at 1:44
  • $\begingroup$ I also noticed this is similarly asked in the comment section below the article. I think it’s the first question, which the author (sort of) explains, although briefly. $\endgroup$ – Hanzy Apr 13 at 1:45
  • $\begingroup$ @Hanzy but i still dont understand why he could replace sum of values with expectation? If there is really half 0s and half 1s wouldt then the summation be equal to N/2 rather than 1/2? $\endgroup$ – i_rezic Apr 13 at 8:09
1
$\begingroup$

To understand this equation first you need to understand the context in which it is first introduced. We have two neural networks (i.e. $D$ and $G$) that are playing a minimax game. This means that they have competing goals. Let's look at each one separately:

Generator

Before we start, you should note that throughout the whole paper the notion of the data-generating distribution is used; in short the authors will refer to the samples through their underlying distributions, i.e. if a sample $a$ is drawn from a distribution $p_a$, we'll denote this as $a \sim p_a$. Another way to look at this is that $a$ follows distribution $p_a$.

The generator ($G$) is a neural network that produces samples from a distribution $p_g$. It is trained so that it can bring $p_g$ as close to $p_{data}$ as possible so that samples from $p_g$ become indistinguishable to samples from $p_{data}$. The catch is that it never gets to actually see $p_{data}$. Its inputs are samples $z$ from a noise distribution $p_z$.

Discriminator

The discriminator ($D$) is a simple binary classifier that tries to identify which class a sample $x$ belongs to. There are two possible classes, which we'll refer to as the fake and the real. Their respective distributions are $p_{data}$ for the real samples and $p_g$ for the fake ones (note that $p_g$ is actually the distribution of the outputs of the generator, but we'll get back to this later).

Since it is a simple binary classification task, the discriminator is trained on a binary cross-entropy error:

$$ J^{(D)} = H(y, \hat y) = H(y, D(x)) $$

where $H$ is the cross-entropy $x$ is sampled either from $p_{data}$ or from $p_g$ with a probability of $50\%$. More formally:

$$ x \sim \begin{cases} p_{data} \rightarrow & y = 1, & \text{with prob 0.5}\\ p_g \;\;\;\,\rightarrow & y = 0, & \text{otherwise} \end{cases} $$

We consider $y$ to be $1$ if $x$ is sampled from the real distribution and $0$ if it is sampled from the fake one. Finally, $D(x)$ represents the probability with which $D$ thinks that $x$ belongs to $p_{data}$. By writing the cross-entropy formula we get:

$$ H(y, D(x)) = \mathbb{E}_y[-log \; D(x)] = \frac{1}{N} \sum_{i=1}^{N}{ \; y_i \; log(D(x_i))} $$

where $N$ is the size of the dataset. Since each class has $N/2$ samples we can split this sum into two parts: $$ = - \left[ \frac{1}{N} \sum_{i=1}^{N/2}{ \; y_i \; log(D(x_i))} + \frac{1}{N} \sum_{i=N/2}^{N} \; (1 - y_i) \; log((1 - D(x_i))) \right] $$

The first of the two terms represents the the samples from the $p_{data}$ distribution, while the second one the samples from the $p_g$ distribution. Since all $y_i$ are equally likely to occur, we can convert the sums into expectations:

$$ = - \left[ \frac{1}{2} \; \mathbb{E}_{x \sim p_{data}}[log \; D(x)] + \frac{1}{2} \; \mathbb{E}_{x \sim p_{g}}[log \; (1 - D(x))] \right] $$

At this point, we'll ignore $2$ from the equations since it's constant and thus irrelevant when optimizing this equation. Now, remember that samples that were drawn from $p_g$ were actually outputs from the generator (obviously this affects only the second term). If we substitute $D(x), x \sim p_g$ with $D(G(z)), z \sim p_z$ we'll get:

$$ L_D = - \left[\; \mathbb{E}_{x \sim p_{data}}[log \; D(x)] + \; \mathbb{E}_{z \sim p_{z}}[log \; (1 - D(G(z)))] \right] $$

This is the final form of the discriminator loss.

Zero-sum game setting

The discriminator's goal, through training, is to minimize its loss $L_D$. Equivalently, we can think of it as trying to maximize the opposite of the loss:

$$ \max_D{[-J^{(D)}]} = \max_D \left[\; \mathbb{E}_{x \sim p_{data}}[log \; D(x)] + \; \mathbb{E}_{z \sim p_{z}}[log \; (1 - D(G(z)))] \right] $$

The generator however, wants to maximize the discriminator's uncertainty (i.e. $J^{(D)}$), or equivalently minimize $-J^{(D)}$.

$$ J^{(G)} = - J^{(D)} $$

Because the two are tied, we can summarize the whole game through a value function $V(D, G) = -J^{(D)}$. At this point I like to think of it like we are seeing the whole game through the eyes of the generator. Knowing that $D$ tries to maximize the aforementioned quantity, the goal of $G$ is:

$$ \min_G\max_D{V(D, G)} = \min_G\max_D \left[\; \mathbb{E}_{x \sim p_{data}}[log \; D(x)] + \; \mathbb{E}_{z \sim p_{z}}[log \; (1 - D(G(z)))] \right] $$

Disclaimer:

This whole endeavor (on both my part and the authors' part) was to provide a mathematical formulation to training GANs. In practice there are many tricks that are invoked to effectively train a GAN, that are not depicted in the above equations.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.