Original Source Here
GAN — A Simple Introduction
Generative Adversarial Networks , a Tug of War approach were two models battle each other for better results.
Assume you are a CEO and two of your employees competing with each other to impress you what would your thoughts be ? If I were in your shoes my answer would be “ Well, I am the winner of their rivalry”. Generative Adversarial Networks work on the same principle , two Deep learning models compete with one other for better accuracy and the winner would always be You 🙂.
The two Employees (DL Models) are called Generatore and Discriminator. Assume thier relationship is similar to that of a Programmmer and Software Tester. A software tester is paid to test code or application programmed by the programmer and find mistakes in them but we Programmers are focused on evading getting caught rather than producing a perfect application. The Generator and Discriminator does the same thing ! Generator(Programmer) tries to generate an image and a Discriminator (Software Tester) tries to differentiate between Fake Images generated by Generator and Real Images. Both compete with the other to get better results.
Goal of GANs
GANs are used to generate images that doesnt exist in real life , generate missing portions of image , create a cartooned version of yourself and the list goes on and on. So from the applications could you figure out who should win the tug of war?? Its the Generator.
Deep Dive Into Its Architecture
Generator creates an image from random input and is trained to produces images similar to that of original image.A real image and the image generated from the generator are passed into the discriminator . Discrimator labels one of them as real image and another as fake. After the discriminator labels both images, discriminator loss and generator loss are calculated and lame old backpropogation takes place.
PRO TIP: If the discriminator is getting better and better during the training process it is a sign that your generator is not performing to your expectations. Try to improve your data or reconstruct your generator.
Interesting Articles To Look Into
Style GAN 2 : A research done by NVIDIA which generates high quality images of people who does not exist. Here is the link for its research paper and here is a website to look into to experience the prowess of Style Gan 2.
GP-GAN : A resarch which focuses of high quality blending of two images using a type of GAN called Wasserstein GAN . The Research Paper could be found here.
Toonify : A website which converts which converts your face into a catoon by leveraging the power of GANs. Here is the link (Try the Free Classic Version).
What Next ?
I would feel happy if my article has played a part to get you interested with GANs. I suggest you start learning more about GANs right away. I would create a series on detailed working and explaination of GANs and would post the link right here when its ready. All the best!
Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot