Generative Adversarial Networks (GANs)

September 12, 2022 Shrutika
Reading Time: 10 minutes

A computer can be trained in a variety of ways to produce an output based on hidden data. Everyone is shocked by the technical development across several industries. We have reached a stage where deep learning and neural networks are so powerful that they can create a brand-new human face from nothing that looks real depending on some trained data. No other method than GAN is used (Generative Adversarial Network)

Introduction to GANs

Ian Goodfellow and his colleagues developed the Generative Adversarial Networks (GANs) in 2014. In essence, GAN is a generative modelling technique that creates a fresh collection of data from training data that resembles training data. To capture, duplicate, and evaluate the variations in a dataset, GANs have two major blocks (two neural networks) that compete with one another. We shall cover the two models in Components on GANs under the names Generator and Discriminator. Let’s break the term “GAN” into its three parts so that you can understand it.

  • Generative -to gain knowledge of generative models, which explain how data is generated using probabilistic models. It explains how data is generated visually in easy language.
  • Adversarial -The model is trained in an adversarial environment.
  • Networks -use deep neural networks for training purposes.

Let’s first understand the many real-world use cases that Generative Adversarial Networks (GANs) see in tech firms, highlighting their relevance today.

GANs are used by Adobe to create the next-generation of Photoshop. Google makes use of the capabilities of GANs to generate both text and graphics. IBM successfully augments data using GANs. They are used by Snap chat for effective image effects and by Disney for high resolutions. GANs have a wide range of applications and offer a number of benefits on the global market today, and their demand is expected to rise over the future years.

Why did GANs develop?

By introducing a small amount of noise to the data, machine learning algorithms and neural networks can be easily fooled into misclassifying objects. The likelihood of misclassifying the photos increases after adding some noise. Hence the modest increase in interest in whether it is practical to create something that would enable neural networks to begin recognizing novel patterns in sample train data. As a result, GANs were developed that generate fake results that are similar to the original.

Understanding Generative and Discriminative Models

In machine learning and deep learning, discriminative models act as classifiers. They are frequently applied to make a distinction between two classes or a group of labels. Examples of these kinds of activities include distinguishing a dog from a cat, identifying different dog breeds, or categorizing different fruits (like apple, grapes, oranges, etc.).

On the other hand, generative models operate in a different manner than their discriminative counterparts. In a generative model, random samples often called noise are taken into account and used to create fresh, lifelike visuals. As an illustration, a generative model may make artificial yet lifelike dog images by first learning from real-world dog images.

Shown below is an example of a GAN. There is a database that has real 50 rupee notes. The generator neural network generates fake 50 rupee notes. The discriminator network will help identify the real and fake notes.

What is a Generator?

A Generator in GANs is a neural network that creates fake data to be trained on the discriminator. It learns to generate plausible data. The generated examples/instances become negative training examples for the discriminator. It creates a sample by receiving as input a fixed-length random vector carrying noise.

The Generator’s main goal is to get the discriminator to classify its output as real. The GAN’s generator training component consists of the following:

  • input vector with noise
  • network generator that turns the random input into a data instance.
  • the data is classified by a discriminator network.
  • generator loss, which fines the Generator for not fooling the discriminator

By calculating how each weight will affect the output, the backpropagation method is used to change each weight in the desired direction. It is also used to obtain gradients, and these gradients can be used to modify the weights of the generator.

What is a Discriminator?

The Discriminator, a neural network, separates real data from fake data produced by the Generator. Data used to train the discriminator is gathered from two different sources:

  • The Discriminator uses real data instances, such as photographs of real birds, people, money, etc., as positive training examples.
  • During the training phase, the fake data instances generated by the Generator are used as negative examples.

The discriminator is connected to two loss functions during training. The discriminator only uses the discriminator loss during training, ignoring the generator loss.

The discriminator classifies both real data and fake data from the generator throughout the training phase. A discriminator who incorrectly labels a real data instance as fake or a fake data instance as real is penalized by the discriminator loss.

The discriminator updates its weights through backpropagation from the discriminator loss through the discriminator network.

How Do GANs Work?

GANs have two neural networks in them. There is a Discriminator and a Generator . They engage in competitive play. By creating data that are similar to those in the training set, the generator tries to fool the discriminator. By distinguishing between fake and real data, the discriminator will attempt to avoid being fooled. To learn and train complex data, such as audio, video, or image files, they both work simultaneously.

The Generator network creates a fake sample of data using a real sample as input. The Generator is trained to increase the probability of errors in the Discriminator network.

The GAN in the example below is attempting to determine whether the 100 rupee notes are real or fake. As a result, the Generator network is initially fed a noise vector or the input vector. Fake 100-rupee notes are produced by the generator. Along with the fake notes, the discriminator also receives the authentic photographs of 100 rupee notes that are kept in a database. The notes are then classified as real or fake by the Discriminator.

The model is trained, the loss function at the end of the discriminator network is calculated, and the loss is backpropagated into the discriminator and generator models.

Steps for GAN Training

  • Define the issue
  • Select the GAN’s architecture
  • Train a discriminator with real data.
  • Create fake inputs for the generator.
  • train a discriminator to find fake data
  • Train generator with the discriminator’s output

Types of GANs

Vanilla GANs: Vanilla GANs use sigmoid cross-entropy loss as part of their min-max optimization formulation, and the discriminator is a binary classifier. In Vanilla GANs, the Generator and the Discriminator are multi-layer perceptrons. The approach uses stochastic gradient descent to attempt to optimize the mathematical equation.

Deep Convolutional GANs (DCGANs): DCGANs support convolution neural networks at both the discriminator and generator in place of standard neural networks. They produce photos of higher clarity and are more steady. The Generator up-samples the input picture at each convolutional layer since it is a collection of convolution layers with transpose or fractional-strided convolutions. The input image is down-sampled at each convolution layer since the discriminator is a collection of convolution layers with strided convolutions.

Conditional GANs: To get superior results, vanilla GANs can be expanded into Conditional models by utilizing extra-label data. In CGAN, the Generator is given an extra parameter, “y,” to produce the necessary data. To help the Discriminator differentiate between the real data and the fake generated data, labels are provided as input.

Super Resolution GANs: SRGANs produce better resolution images by combining deep neural networks with an adversarial network. When given a low-resolution image, SRGANs produce a photo realistic high-resolution image.

Now that we have understood what are GANs, let’s look at some of the important applications of GANs.

Applications of Generative Adversarial Networks (GANs)

Reading about GANs is too fascinating, and when you read about their use, I’m hoping that excitement will rise. Studying how GANs operate then has a different effect on learning.

  1. Create new data from existing data- it means creating new samples from a sample that already exists but isn’t exactly like the real thing.
  2. Generate realistic pictures of people that have never existed.
  3. Gans may also produce text, articles, songs, poems, and other types of content in addition to images.
  4. Create music using some cloned voice. If you supply some voice, GANs can create a feature that is similar to that voice. Researchers from NIT in Tokyo presented a system in this study that can create melodies from lyrics using relationships between notes and subjects that have been learned.
  5. Generating images from text (Object GAN and Object Driven GAN).
  6. Creation of anime characters in Game Development and animation production.
  7. Image to Image Translation – Without changing the background of the original image, we can translate one image to another. Gans, for example, can switch out a dog for a cat.
  8. Low resolution to high resolution – If you send GAN a low-quality image or video, it will output a high-resolution image rendition of the same.
  9. Prediction of Next Frame in the video – By training a neural network on small frames of video, GANs are capable to generate or predict a small next frame of video
  10. Interactive Image Generation – If GANs are trained on the appropriate real dataset, they can produce artistically created images and video recordings.

In the following years, we will see GANs produce high-quality video, audio, and photos as research on them reaches a peak. Microsoft and OpenAI have already worked together to develop GPT and explore the potential of GAN at a higher level.

Issues with Generative Adversarial Networks (GANs)

  1. The issue of generator and discriminator stability. We want to be tolerant and do not want the discriminator to be too strict.
  2. Finding the positions of the objects is difficult. Imagine that there are 3 horses in the image, 6 eyes were generated, and 1 horse.
  3. Similar to the perspective issue, the problem with understanding global objects is that GANs do not comprehend the global structure or holistic structure. It indicates that occasionally, GAN creates an image that is implausible and surreal.
  4. It has trouble recognizing 3-D images, and if we train it on these kinds of images, it will be unable to produce 3-D images because GANs are currently only able to operate with 1-dimensional images.


Our main objective when writing this article was to develop a practical grasp of how Generative Adversarial Networks (GANs) work. GANs are a remarkable achievement of the modern deep learning era. They offer a distinctive method for producing data like photos and text and are also capable of a wide range of other tasks, including data augmentation and natural image synthesis.

Let’s quickly recap the numerous topics we discussed in this article. We had a quick introduction and learned about realistic expectations from GANs, including their usage in industry. We then proceeded to understand the different types of modeling, including discriminative models and generative models then we have seen Generative Adversarial Network (GAN) applications

, , , , ,
Reading Time: 10 minutes


Business Intelligence
Data Analytics
Product Development


#603, Aarush Arya Apartment
Channasandra Layout, Uttarahalli-Kengeri Main Road
Bangalore – 560098
+91 – 8050095950


#8, A Block, Gurudatta Complex
Station Road, Malmaddi,
Dharwad – 580001
+91 – 8050095950


© 2023 Aiwoox®

Privacy Policy.

© 2023 Aiwoox®