Learning Objectives

  • Types of Generative Modeling?
  • What is Generative Modeling
  • Generative Adversarial Network (GAN)
  • Applications of GAN
  • Components & Working of GAN

Generative Modeling

  • As the name says ‘Generative’, it’s all about creating or generating or producing any sensible data.
  • It’s complex when compared to other deep learning models, and it requires a lot of data to train the model.
  • Some examples of well-known generative modelings are GPT-1, GPT-2, GPT-3, etc.
  • It is an Unsupervised Learning Approach as there is no label or target in the dataset while using Generative Modeling.

Types of Generative Modeling

  • The below image is explained in the video in the next section.

image.png

  • Don’t worry about the terms like AutoEncoders, Boltzmann Machine, etc. You will again see them in the optional study material that will be shared with you at the end of the BootCamp.

Time stamp: 3 mins 45 secs to 5 mins 45 secs

Generative Adversarial Network (GAN)

  • GAN was released in 2014
  • Here is a quick abbreviation of what it stands for?
    • Generative: Generator / Producer / Creator - It generates data
    • Adversarial: Checker / Investigator / Opposition Party - It checks what is created or generated. Whether the created data is fake or real. The aim is to make adversarial fool. Adversarial is also known as Discriminator.
    • Network: Neural Network. We use Neural Networks in Generative and Adversarial steps.
  • It comprises of many layers of many layers of convolutional layers, batch normalization and ReLU / Sigmoid with skip connections
  • Go to https://thispersondoesnotexist.com/
  • What do you see? A normal human being, right?
  • If you didn't figure out from the website name yet, the person you saw does not actually exist! It's an image generated by a Deep Learning technique called GAN. Each time you refresh the website, you'll see a completely new person. How cool is that?

image.png

  • Let us look at some examples to understand what a GAN and its variants are capable of:

image.png

  • Given an image of the road, the network is able to fill in the details with objects such as cars etc.
  • The network is able to convert a black & white image into colour.
  • Given an aerial map, the network is able to find the roads in the image.
  • It is also able to fill in the details of a photo, given the edges.

Applications of GAN

  • Generate new data from available data, mostly creating and producing images but not limited to images only.
  • Generates music by using some clone voice.
  • Text to Image Generation
  • Creation of Anime characters in Game development and animation production.
  • Image to Image translation
  • Low resolution to high resolution
  • Prediction of next frame in a video
  • Interactive image generation and many more.

David Beckham speaks nine languages to launch Malaria Must Die Voice Petition

Do you know how? It’s the application of GANs.

Applications of GAN

Ignore the pop ups in between the video

Time stamp: 5 mins 45 secs to 9 mins 35 secs

DeepFakes - Application of GAN

  • Creating convincing image, audio and video hoaxes.
  • Yes, the things in the previous video were definitely not said by former US President Barack Obama.
  • Once upon a time, the researchers were keen on creating DeepFakes.
  • Now, they are working on detecting them!

Components of GAN

  • The network learns to generate from a training distribution through a 2-player game. The two entities are Generator and Discriminator (or Adversarial).

image.png

  • As you can identify from their names, a generator is used to generate real-looking images and the discriminator’s job is to identify which one is a fake. No need to get into their details at the moment. Just enjoy the glory of GANs!

Components & Working of GANs

Ignore the pop ups in between the video

Time stamp: 9 mins 35 secs to the end