Overview Of Generative Adversarial Networks


Once trained, Neural Networks are fairly good at recognizing voices, images, and objects in every frame of a video – even when you are playing the video. Let’s say you are not able to buy an expensive painting from a famous painter; can you create/generate an artificial painter who can paint like any famous artist by learning from his/her past collections? The answer is Yes – using Generative adversarial networks (GAN) you can. Generative Adversarial Networks (GANs) are a class of algorithms used in unsupervised learning -- you don’t need labels for your dataset in order to train a GAN.

So what is a GAN?

Let’s take the example of a painter to understand GAN. The intuitive idea for a GAN is that there is an expert deep neural network and a learner deep neural network. Then we make them fight against each other, endlessly attempting to out-do one another. In the process, they both become stronger.

The expert has information about the original painting. The learner learns in multiple passes where it generates the neural network and gives its output to the expert to check correctness. The job of the expert is to find the real vs fake. As the process continues, the expert and learner are trained in an alternating fashion.  So the expert and learner - they are dependent on each other for efficient training. If one of them fails, the whole system fails.


Generative adversarial networks consist of two models: a generative model and a discriminative model. Discriminative model is our expert here and generative model is the learner.

The whole foundation of GANs is the equilibrium between the two networks. They help each other adapt and become stronger. They are both addressing different problems, i.e, the discriminator is a classifier network and generator is a regressor network. They have different architectures and are trained with different loss functions. Hence they cannot use each other's weights.

Generative adversarial networks

The discriminator model is a classifier that determines whether a given image looks like a real image from the dataset or like an artificially created image. This is basically a binary classifier that will take the form of a normal convolutional neural network (CNN).

The generator model takes random input values and transforms them into images through a deconvolution neural network. After many training iterations, the weights and biases in the discriminator and the generator are trained through backpropagation.

  1. Do a forward pass through the Generator network. This outputs the “fake” image since it is created from the Generator.
  2. Use the fake image as the input to do a forward and backward pass through the Discriminator network. We set the labels for logistic regression to 0 to represent that this is a fake image. This trains the Discriminator to learn what a fake image looks like. We save the gradient produced in backpropagation for the next step.
  3. Do a forward and backward pass through the Discriminator using a real image. The label for logistic regression will now be 1 to represent the real images, so the Discriminator can learn to recognize a real image.
  4. Update the Discriminator by adding the result of the gradient generated during backpropagation on the fake image with the gradient from backpropagation on the real image.
  5. Now that the Discriminator has been updated for this data batch, we still need to update the Generator. First, do a forward and backward pass with the same data batch on the updated Discriminator, to produce a new gradient. Use the new gradient to do a backward pass.

Eventually, the Generator is producing near-perfect counterfeits and the Discriminator has turned into a Master Detective looking for the slightest mistakes. If you’re interested only in generating new images, you can throw out the discriminator after training.

Mathematical Representation

To represent this mathematically - The generator (G) and discriminator (D) are both feedforward neural networks which play a min-max game between one another and is represented as below,

Generative adversarial networks


Pdata(x) -> the distribution of real data
X -> sample from pdata(x)
P(z) -> distribution of generator
Z -> sample from p(z)
G(z) -> Generator Network
D(x) -> Discriminator Network

So overall, the discriminator is trying to maximize our function V and the task of generator is exactly the opposite - to minimize the function V

  • This new framework comes with advantages and disadvantages relative to other modeling frameworks. The disadvantages are primarily that D must be synchronized well with G during training - in particular, G must not be trained too much without updating D.
  • The advantages are that Markov chains are never needed, only backpropagation is used to obtain gradients, no inference is needed during learning, and a wide variety of functions can be incorporated into the model.
  • Adversarial models may also gain some statistical advantage from the generator network not being updated directly with data examples, but only with gradients flowing through the discriminator. This means that components of the input are not copied directly into the generator’s parameters.
  • Another advantage of adversarial networks is that they can represent very sharp, even degenerate distributions, while methods based on Markov chains require that the distribution be somewhat blurry in order for the chains to be able to mix between modes.

GAN in real life

Adobe Research is using GAN for designing products, generating novel imagery from scratch based on users' scribbles.

For more details visit here.

Facebook has built a real-time style transfer model running on mobile devices. The style-transfer tool in the camera is the result of a marriage between two technologies: the Caffe2go runtime and style-transfer models. It is a new creative-effect camera in the Facebook app that helps people turn videos into works of art in the moment. That technique is called “style transfer.” It takes the artistic qualities of one image style, like the way Van Gogh paintings look and applies it to other images and videos. You can see the video here.

Below are some examples of images generated by GANs - Generated bedrooms. Source: “Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks” here.

Generative adversarial networks


There are many applications coming up with new functionality that will be highly productive but it’s a scary picture. You can duplicate almost hypothetically anything if you had a fully functional generator. Some examples - fake news, fake voice, fake music, create books which seem written by authentic authors and much more.

All major corporations are spending a lot of research in this area and we will soon see its effects/side-effects in our day-to-day life.