You are on page 1of 15

A Technical Seminar

On
Generative Adversarial Neural
Networks

By Bhavesh Dhera
15241A05I4
Agenda
1. Abstract
2. Introduction - Generative Algorithms
3. Introduction
4. Introduction (cont.)
5. How do GANs work?
6. GAN Use cases
7. Sample Outputs generated using GANs
8. ThisPersonDoesNotExist
9. Applications
10. Conclusion
11. References
Abstract
• Generative Adversarial Neural Networks (GANs) is a
framework for creating generative models via an adversarial
process that uses two models: a generative model ‘G’ that
captures the data distribution, and a discriminative model ‘D’
that estimates the probability that the sample came from
training data and not from ‘G’. The Generator ‘G’ learns to
produce realistic-looking data while the Discriminator ‘D’
learns to distinguish generated examples from the real ones.
Introduction
• There are 3 classes of Generative Algorithms:
– Predict associated features when given a label (Naïve
Bayes)
– Predict associated features when given a hidden
representation (Generative Adversarial Networks)
– Given some features, predict the rest of the features
(Inpainting, Imputation)
Introduction (cont.)
• A Generative Adversarial Network (GAN) is a class
of machine learning systems in which two neural networks, the
Generator and the Discriminator compete with each other in
a zero-sum game framework which can generate photographs
that look at least superficially authentic to human
observers, having many realistic characteristics.
• The Generator ‘G’ generates candidates while the
Discriminator ‘D’ evaluates them. The Generator tries to
maximize the Discriminators error rate, i.e., tries to fool it
more often, by generating more and more authentic looking
images.
Introduction (cont.)
• The Discriminator ‘D’ is trained initially until it achieves acceptable
accuracy. The Generator ‘G’, on the other hand, is seeded with
randomized I/P that is sampled from a predefined latent space.
• Backpropagation is used in both the networks allowing ‘G’ to
generate more authentic images and ‘D’ to get better at flagging
synthetic images.
• The Discriminator, which serves essentially as a classifier, is a
‘Convolutional Neural Network’, whereas a Generator is a
‘Deconvolutional Neural Network.’
How do GANs work?
• The three important parts of a GAN are:
– Generative: To learn a generative model which describes how
data is generated in terms of a probabilistic model.
– Adversarial or Discriminative ‘D’: The training of a model is
done in an adversarial setting.
– Networks: Using Deep Neural Networks (DNNs) as the AI
algorithms for training.
• The GANs are formulated as a minimax game; D tries to maximize
its reward V(D, G) and G tries to minimize D’s reward, that is to
say, maximize its loss, which can be described mathematically as:
GAN Use Cases

Examples from the original paper by Ian Goodfellow


GAN Use cases

These monet-style paintings were GANs trained to generate objects with only the
generated from real images using a edges given as I/P.
GAN
Sample Outputs generated using
GANs
ThisPersonDoesNotExist
• https://thispersondoesnotexist.com/ is a
website that shows synthetic images of people
generated using a GAN. Below are a few
images from the website.
Applications
• GANs that produce photorealistic images can be used to
visualize interior/industrial design, or items for computer
games’ scenes.
• GANs can model patterns of motion in videos, reconstruct 3D
models of objects from images and improve astronomical
images.
• GANs can be used to age face photographs to generate an
individual’s appearance changing with age.
• GANs can be used to upscale low resolution 2D textures in old
video games by recreating them in 4k or higher resolution.
• Generate content like multimedia, text and speech.
• GANs can create data from scratch.
Conclusion
• In Conclusion, GAN’s, despite taking up a lot of
computing power, will have a wide array of
applications.
References
• Goodfellow, Ian; Pouget-Abadie, Jean; Mirza, Mehdi; Xu, Bing;
Warde-Farley, David; Ozair, Sherjil; Courville, Aaron; Bengio,
Yoshua (2014). Generative Adversarial Networks(PDF).
Proceedings of the International Conference on Neural Information
Processing Systems (NIPS 2014). pp. 2672–2680.
• Salimans, Tim; Goodfellow, Ian; Zaremba, Wojciech; Cheung,
Vicki; Radford, Alec; Chen, Xi (2016). "Improved Techniques for
Training GANs”
• https://freecontent.manning.com/what-are-gans/
• https://www.analyticsvidhya.com/blog/2017/06/introductory-
generative-adversarial-networks-gans/
• https://towardsdatascience.com/generative-adversarial-networks-
gans-a-beginners-guide-5b38eceece24?gi=d61f5b83e74a
• https://arxiv.org/pdf/1806.11302.pdf
Thank You

You might also like