Image credits: Venturebeat
A much-anticipated technology for this last decade, GANs or Generative Adversarial Networks are truly revolutionizing artificial intelligence technology today, and with it, the world of multimedia. GANs are quite impressive technologies, which generations before us only believed possible in science fiction movies. As Yann LeCun, director of Facebook A.I. Research said: "Generating adversary networks is the most interesting idea in machine learning in the last ten years".
Developed by Ian Goodfellow and his team in 2014, machine learning frameworks have already been described as a promising technology at that time.
Stimulated by advertising and social media production, GAN reached its zenith in the popular age of selfies and memes (you know exactly what I mean). However, just after the release of videos portraying Deepfakes of high state authorities, everyone began to wonder about the extent of the discoveries, which also gave clues about the potential weapons this technology could provide in malicious hands.
Having said that, let us set all this aside to focus on a few interesting applications of GAN and the mechanism behind them because there is more to be said about it.
- Concept and Architecture
- Use cases
Concept, Architecture et Algorithm
Based on Neural Networks, they allow creating faces, characters, objects, music, and even works of art that do not exist or made by a human hand, simply by training a neural network to create new ones from similar objects or the same concept taking life.
As a technique, they gather material from reading their description for example, by starting from scratch, "drawing inspiration" or imitating existing people and objects to create something new or alter the reality, hence their name "generative".
Speaking about GAN, many paradoxical concepts come to mind, concepts that are still a bit opaque although the development of GAN models is starting to blossom. The GAN models share similar principles and follow a common pattern for construction that can be categorized into these 6 well-defined GAN Architecture to cite :
They have all composed of layers of neural networks the role of those is divided into a generator and a discriminator. As the name suggests these neural networks are adversarial and to simplify, one of them performs a kind of Turing test on the other as this latter submits its interpretation of a subject. The submitter is the generator, whereas the other is the discriminator. The discriminator challenges the generator as long as the produced result is still detectable.
The process goes on until the discriminator validates that there is no more clue to distinguish what is real from what is not. This is how we obtain entirely new faces of persons that do not even exist.
This article details the different kinds of GANs as well as their architectures: Some cool applications of GAN.
-For a thorough explanation, this article provides more details on the subject What is Generative Adversarial Networks GAN?
-For more details on the machine learning approach see this Towards Data science article
- A hint on the mathematics behind it, check this Medium article.
-The GAN series: From beginning to the end (Jonathan Hui).
If you're not yet aware, many applications on your smartphone use GAN. Many social media platforms namely Instagram and Facebook Messenger, and a great number of beauty applications have already incorporated it and have since boosted their sales.
Machine Learning Mastery has identified the following applications of GAN as their most popular use cases:
- Generate Examples for Image Datasets
- Generate Photographs of Human Faces
- Generate Realistic Photographs
- Generate Cartoon Characters
- Image-to-Image Translation
- Text-to-Image Translation
- Semantic-Image-to-Photo Translation
- Face Frontal View Generation
- Generate New Human Poses
- Photos to Emojis
- Photograph Editing
- Face Aging
- Photo Blending
- Super Resolution
- Photo Inpainting
- Clothing Translation
- Video Prediction
- 3D Object Generation
Let us add to the list the generation of music with GAN.
Nevertheless, TechRepublic warns us of the potential dangers of an AI that is becoming more and more accessible to the public. The benefits are just as numerous as the potential dangers, to mention only the threats to digital and physical political security. As the article points out, while the knowledge of fake videos was limited to a few audiences previously, anyone can use them now, adding an arrow to the bow of cyber-criminals .
Researchers and thinkers gather thoughts on the topic of attacks done with the help of AI and the means of preventing them, in the report survey results: The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation.
This should prepare the juridic field of investigations to level up their wit and tools just in case...
As with the Generative Adversarial Network or GAN, the technological advances in Artificial intelligence are breathtaking. Most of us greet each discovery with frantic enthusiasm, and rightly so: AI is truly amazing.
Today, we are already at a time when an intricate wonderland, not to say the fantastic GAN simulated world is among us.
Now that what we thought only possible in the movies has been embodied, and has evolved beyond recognition, we are about to enter a cyber-world where artificial intelligence and humans mingle without us distinguishing what is "human" from what is "artificial". Artificial intelligence is getting closer and closer to passing the Turing test.
However, we must agree that, as with any technological discovery, we must be aware of its possible misuse, but also prepare for any of this latter to mitigate the risks.
All of this can only remind me of a thrilling episode of Black Mirror, as others have surely already noticed before me (nice series by the way). So we want technology to progress perpetually, but with it also, rhythm after rhythm, our moral dimension on alert, as a vigilant safeguard.
The links provided here deal with a more detailed explanation of GAN
As part of the GAN series, Jonathan Hui has covered a comprehensive study of the various aspects of GAN in these Medium articles including use cases, problems and solutions in the link provided. ici .
-This person does not exist provides links for training GAN.