...

Why are Generative Adversarial Networks (GAN's) Revolutionary?

Generative Adversarial Networks have been applauded by AI researchers all along with the world as is one of the biggest advancements in the machine learning field over the last ten years. Therefore making it a must-know concept for every Artificial Intelligence. This article includes:

  • Overview
  • How GANs work
  • Applications
  • The concern of Malicious Use

Overview

GANs are algorithmic structures that consist of two neural networks, namely the Generator and the Discriminator which are contesting against each other in order to generate new data samples after being fed upon the training samples. It follows the generative approach of unsupervised learning using deep learning methods. The Generative neural network is fed upon the real data and is used to generate synthetic similar data whereas the Discriminative neural network evaluates the output of the generator, and classifies the content into the actual data as well as the fake one, which was generated by the generator.

 

Image result for generative adversarial networks

Generative Adversarial Networks Architecture

 

How GANs work?

One neural network Generator generates new data that is to be similar to the true data and then the discriminator authenticates the generated data, i.e. classifies the data into actual data and the false data. The objective of the Generator is to fool the Discriminator, which means to generate the samples as similar to the real data as it can get.

Therefore there is a double feedback loop going on. The Discriminator gets its feedback from the user who knows the ratio of real and fake values that are fed to the Discriminator whereas the Generator gets its feedback from the Discriminator. So here, the discriminator is basically acting as a classifier. In other words, it is trained upon the feature and tells if a particular category has similar features or not. While the Generator can be seen as the opposite of the discriminator, it takes a category and predicts the features data belonging to its share.

For instance, let there be a GAN generating images of handwritten numbers as in the MNIST dataset, the steps it may involve would be:

  • Random numbers are given as input to the Generator neural network.
  • The Generator generates fake samples based on those inputs.
  • The output from the Generator along with the true data values are provided to the Discriminator.
  • The discriminator returns a number between 1(for real image) and 0(for fake image), which represents the probability of the dataset being authentic.

 

Image result for GAN on mnist dataset handwritten digits

GAN working on MNIST dataset of handwritten digits

 

Generative Adversarial Networks Applications

  • GANs have applications in almost every field in which machines have a future scope, from the fashion industry to medicine and from the stock markets to education fields. The most unique being the field of art. In the year 2018, a trio of three French students sold portraits made by an Artificial Intelligence for $432,000. However, they used Robbie Barrat’s code from an opensource.

 

  • Other applications include the modern video gaming industry. With the games like Final Fantasy VIII, Final Fantasy IX, and Residence Evil using extensive Generative Adversarial Networks.

 

  • GANs are also supposed to have high application in astronomy and other scientific fields. It can be used to map better extraterrestrial images. Applications in the context of present and proposed CERN experiments have demonstrated the potential of these methods for accelerating simulation and/or improving simulation fidelity.

 

Concern for Malicious Use

There is a possible grey area onto which GANs can be used. Concerns have been made about the possible misuse of GANs for purposes such as fake human images for the purpose of identity theft, creation of unique social media profiles for people who don’t even exist.

 In 2019 the state of California considered and passed on October 3, 2019, the bill AB-602, which bans the use of human image synthesis technologies to make fake pornography without the consent of the people depicted, and bill AB-730, which prohibits the distribution of manipulated videos of a political candidate within 60 days of an election. Both bills were authored by Assemblymember Marc Berman and signed by Governor Gavin Newsom. The laws will come into effect in 2020.

 

Yashita_Vaid

Technical writer at Aipoint. She holds immense knowledge in Deep Learning and python.