Coursera Learner working on a presentation with Coursera logo and
Coursera Learner working on a presentation with Coursera logo and

In this article, you will learn about generative adversarial networks or GANs. Generative adversarial networks are the generative modeling approach for deep learning techniques. This example for GANs includes convolutional neural networks.

Generative modeling is a machine learning task for unsupervised learning. It includes learning the regularities, automatic discoveries, or patterns as input data. This way, the model can use new generative data for machine learning, which involves automatic learning. You can use all the regularities or patterns in the input data in a way that generates new examples you can draw from the original dataset.

Use GANs for training the generative model and framing the problem for supervised learning with two sub-models.

·         Discriminator Model

The discriminator model classifies the examples as fake (generated) or real (domain). You need to train both these models together in an adversarial, zero-sum game. This means that the generator model generates plausible examples.

·         Generator Model

We use the generative model for training new examples.

The fields of GANs are rapidly and excitingly changing in the ability to create real examples in various domains. This is especially when the tasks relate to the image, to image photo translations such as winter to summer or night to day. This helps generate photorealistic scenes, objects, and people that you won’t recognize as fake. This article will help you discover the Generative Adversarial Networks or GANs.

What is GANs

The GAN or Generative Adversarial Network will work as an algorithmic architecture using two neural networks. Both the networks will oppose each other to generate synthetic and new data instances, passing the real data. You can use it for video generation, voice recognition, and image generation. The potential for GANs can serve both evil and good. They will distribute data and mimic each other. Their output will be remarkable for all domains such as speech, music, images, and prose.

How GANs Work

There is a neural network that helps in generating the new data instances. Experts call this neural network ‘generator.’ Through the other neural network, the discriminator evaluates them with authenticity. It means that the discriminator will decide if each data instance that it evaluates belongs to the actual training dataset.

Suppose we want to mimic the Mona Lisa. We will be generating hand-written numerals found in the MNIST dataset received from the real world. The main goal of the discriminator, while showing the instance with the help of a true MNIST dataset, is to identify the authentic ones.

In the meantime, the generator will create new and synthetic images passing to the discriminator. This will generate a new image similar to the authentic but fake. The generator will pass the hand-written digits as the goal to lie without getting caught. The discriminator will identify the images that come as fake from the generator. GAN can take the following steps:

  • The images generated by the generator will be fed alongside the stream of images into the generator that the ground-truth and actual dataset receive.
  • The generator will take the random numbers and return the image as an output.
  • The discriminator will take both the fake and the real images and return the probabilities. For instance, if the numbers are between 1 and 0. 1 will represent the prediction as authentic, and 0 will represent the fake.

This way, you will have a loop with double feedback:

  • The generator and discriminator will be in the same feedback loop.
  • The discriminator will be in the feedback loop,and images will be the ground truth.

To understand the GAN, we can consider the cop and the opposition of a counterfeit in a game of mouse and cat. In the game, the counterfeit will learn to pass the false notes, and the cop will learn to detect those notes. Both the characters will be dynamic. For instance, all the training that the cop receives will transfer to the other character in a constant escalation.

The discriminator network is the standard convolutional network for MNIST, and this will categorize all the images that it receives. The binominal classifier will label the images as fake or real. On the other hand, the generator will be the inverse convolutional network. The standard convolutional classifier will take the image, and it will downsample the image-producing probability. The generator will unsample the random noise after taking it as a vector.

The model will throw away the downsampling technique as the first data and generates new data as the second such as max-pooling. Both the nets will try to optimize different and opposing data as loss function or objective function in a zero-zum game. This will work as an actor-critic model. When the discriminator and the generator change the behavior and vice versa, the losses will be against each other.


The inverse transform technique will generate random variables to follow the given distribution, making it a uniform and random variable. This will go through an elegant transform function. This inverse transform method will extend the notion of the transforming method. Furthermore, it generates random variables. These variables will develop the functions of simpler random variables.



Weekly newsletter

No spam. Just the latest releases and tips, interesting articles, and exclusive interviews in your inbox every week.