Why does any of this matter? Is it that important? Yes, it’s very important. This is the next step in AI, taking old data to make new data. AI can use this to create synthetic data, such as fake videos, fake images, fake text, and make more accurate predictions. GANs are already being used today. Here’s how:
Make high-resolution images from low-resolution images:
Have you ever found an image on google, copied it, pasted it and realized the resolution was too low? It can make finding images a nightmare. Well, not with GANs. A type of GAN called an SRGAN (Super-Resolution GAN) can take these low-quality images and make them into high-quality images. You can test it and see for yourself right here: https://github.com/tensorlayer/srgan
Create images from the text:
If I gave you the words flowers in a field and told you to draw this in one minute, it wouldn’t look very realistic. Well, GANs can do that. The generator gets a list of images similar to the text, takes these images, and uses it to generate an image that it thinks is similar to the prompt. Then, the discriminator takes the generated image, compares it to other images, and try to guess if it’s real or fake. It keeps going until the GAN makes a good image.
Predict stock price movements:
This person took GANs and started using it to predict stock prices. They take trends from the following years, make accurate assumptions and try to guess what the stocks will be for a certain market. Here’s a link to the article if you want to read it. It’s rather long, but it’s rather interesting: https://towardsdatascience.com/aifortrading-2edd6fac689d
Deepfakes:
These are the most common GANs out there. They are videos where they swap the faces in one video for another or edit a video to look like someone is saying something when they aren’t. As of right now, they are somewhat controversial. They can create funny videos but can also ruin someone’s life and status.
— GANs work by using a discriminatory network and a generative network to develop more realistic data.
— Generative networks are networks that take training data to help create realistic data. Its goal is to fool the discriminator.
— Discriminatory networks work by taking the data made by the generator and the training data to try and determine whether it is real or not. The discriminatory network’s goal is to catch which one is made by the generator.
— GAN networks are used in many ways such as making high-resolution images from low-resolution images, create images from text, predict stock price movements, and create deepfakes.
Credit: BecomingHuman By: Assiatou Hann
