banner18

Generative Adversarial Networks Explained

What are generative models?

 

Generative adversarial networks or GANs are models that use a non-supervised learning approach for the generation of a model where the given sample data consist of input variable X but there is no output variable given Y.

So we tend to use that only available input variable to train generative model and ith that the model recognizes the patterns and based on that generates unknown output variable. In supervised learning we are more aligned towards creating predictive models from input variables and this type of modeling is also known as discriminative modeling.  On the contrary, unsupervised models are used to create or generate new examples in input distribution.   

What are GANs?

Generative adversarial networks are deep learning-based generating models that are used for unsupervised learning it is basically a system where two competing neural networks compete with each other to create or generate variations in the data. Gan’s theory was proposed by lan Goodfellow in 2014 in his paper and a more stable model theory was given by Alec Radford in 2016 which is also known as DGCA. We can call it Deep Correlational general adverse networks and most of the GANs today use DGCA network.

 

Generative models tend towards unsupervised learning using machine learning which automatically discovers and learn the regularities i.e regular pattern in given input data. This model can be used to generate new examples that possibly could have been drawn from the given dataset.

GANs are smart ways of training a generative model. The problem can be framed as a  supervised learning problem with two sub-models: the first one is the generator model that we train to generate new examples and the other is the discriminator model that tries to classify examples as either from the domain or generated(i.e either they are real or fake). 

Some applications of GANs

Prediction of next frame in a video

The prediction of future events in a video frame is made possible with the help of GAN and DVD gan or we can call it a dual video discriminative gan that can generate 256 by 256 videos.

  • Text to image generation

Object gan (i.e gans which are object-driven) performs text to image generation in two steps. Firstly gans generates semantics layout and then following that the model synthesizes the image by a deconvoluting image in the final step.

  • Swapping faces

          The AI-powered photo manipulation and different face editing apps are in   

          Full swing. The technology used is deep fake which uses gans in the backend.

          This app is supposed to have the most advanced neural portrait editing 

          technology.            

  • Showing Artistic Skills

These networks have the capability of creating very innovative portraits from very scratch. There are several events where GANs had shown their artistic skills.

  • Generating automated machine

We could design automated avatars using gans. We could create a 3d model of them to explore the world. If we could learn proper probability function we could generate images, songs, and even 3d objects.

 
  •  May, 21, 2021
  • Hemangi Desale
POST A COMMENT
We'll never share your email with anyone else.
Save my name, email, and website in this browser for the next time I comment.
Latest Blogs