Generative Adversarial Networks – Reinforcement Learning Framework

Filed Under: Machine Learning
General Adversarial Networks

A type of deep neural network known as the generative adversarial networks (GAN) is a subset of deep learning models that produce entirely new images using training data sets using two of its components.

This is actually a neural network that incorporates data from preparation and uses current data and information to produce entirely new data.

Learning is strictly indirect in this form of network. It is modified and dynamically altered by its initial discrimination.

As opposed to all other neural network models, this type of network is typically more demanding and complex.

The main objective of this model is to produce new literal scratch info, but this can be separated from domain to domain.

Basically, it outputs something entirely different from the data initially fed into it such as producing an image of a zebra from a horse’s data. It can quickly and more reliably teach a robot to learn in the form of reinforcement learning.

Neurons that compete with each other to process certain data and output another collection of data are generative adversarial neural networks. The generator and the discriminator are both sets of neural networks.

Generative Adversarial Networks Research Paper

Download Link:

A field of successful study was the learning of reusable function representations from large unlabeled datasets.

The nearly infinite number of unlabeled images and videos can be leveraged in the sense of computer vision to learn strong intermediate representations, which can then be used for a range of supervised learning activities such as image recognition.

For this purpose, Generative Adversarial Networks are an optimal candidate.

Generative Adversarial Networks Architecture

A GAN’s fundamental composition consists of two elements, a generator, and a discriminator. Generators that are then discriminated against, create the images.

Below is a detailed description:

Generator: This first component of the GAN is the one that produces new images from the initially fed training data.

GAN Architecture
GAN Architecture

Discriminator: The real images are fed as training data into the discriminator, which then distinguishes between the real and the false. Here it is possible to denote the likelihood of an image being (aka the input x) real as output D(X):

P(input class = actual picture).

The classifiers of the deep neural networks behave the same way this discriminator operates.

In exchanging measures, we train the two networks and lock them into a rivalry to establish themselves.

The discriminator sequentially distinguishes the slight contrast between the genuine and the generated, and the generator allows images that can not be separated by the discriminator.

Backprop: We use backpropagation to balance the weights in the neural networks based on the error of how real/fake the generated image was.

Types of Generative Adversarial Networks

  • acGAN — Face Aging With Conditional Generative Adversarial Networks
  • AdvEntuRe— Adversarial Training for Textual Entailment with Knowledge-Guided Examples
  • AMC-GAN — Video Prediction with Appearance and Motion Conditions
  • BAGAN —Data Augmentation with Balancing GAN
  • CAN — Creative Adversarial Networks
  • CipherGAN — Unsupervised Cipher Cracking Using Discrete GANs
  • D-GAN — Differential Generative Adversarial Networks: Synthesizing Non-linear Facial Variations with Limited Number of Training Data
  • DAGAN — Data Augmentation Generative Adversarial Networks
  • CycleGAN — Unpaired image-to-image Translation using Cycle-Consistent Adversarial Networks

and lots more…

Fascinating Applications of Generative Adversarial Networks

Let’s take a look at some of the very interesting and really cool applications of the Generative Adversarial Networks. If you want to work on some projects of your own, and are looking for data, here are some of the top machine learning datasets.

Human Face Generation

If you go right now to you’ll see a photograph of a person that has been generated entirely by a cycleGAN. In fact, refresh this page right now. The image below will be generated on the fly with machine learning!

General Adversarial Networks
This person does not exist

Similarly, there are other sites like and

Living Mona Lisa

A few-shot GAN network was used to bring Mona Lisa to life:

Mona Lisa Back To Life using General Adversarial Networks
Mona Lisa Back To Life


Nowadays, deepfakes have become quite popular. It uses cycleGAN to not only swap the face of a person but the whole body. Aside from several ethical concerns, since this technology is now available to the mainstream, video evidence of a crime may lose its credibility:

trump_fallon_side.gif (340×200)

Ending notes

If you liked reading this article, follow me as an author. Also, let me know if there is anything specific you would want me to cover. I hope to do a further article on implementing a Generative Adversarial Networks architecture. Until then, keep coding!

Generic selectors
Exact matches only
Search in title
Search in content