GANs : How machines train their brains
One way, humans learn is through playing and competing, the same can be said of machines. In 2014, Ian Goodfellow came up with the idea of Generative Adversarial Networks.
The basic principle behind the idea is that instead of a human feeding information to a single network, one connects two separate neural networks. They will have two different tasks across which they will duel each other. It is important to distinguish between the functions of the two networks. One is the generative network, its job is to generate new images. Its goal is to make the second network believe that the images are real and not artificial creations. This second network is called adversarial or discriminative, its purpose is not to be fooled. After setting up the GANs, they will be left without human supervision until the training has been completed.
So basically, the Discriminator is either fed a real image out of a pre-existing pool or an image created by the Generator. It will assess the image and decide if it thinks it’s a fake or a real image. It will do by giving a value between 0 and 1, indicating the probability of the image being authentic. Both networks will constantly improve on the execution of their task. The training ends when the networks achieve a status quo. Meaning the Generator will create “perfect” images and the Discriminator will give them 0.5 probability of being authentic.
This new model of training a machine is an important step in the research surrounding Artificial Intelligence. The AI can now learn faster with less human interaction. Furthermore, GANs have allowed Computers to become vastly better in creating, which was an area where the deep learning model didn’t excel at.
AI pioneers are ready and eager to use this new technology in the film and videogame industry. The artificial creation of images has plenty of potential useful applications across plenty industries. Researchers already envision GANs used in the field of medicine and even pharmaceuticals. Where medical research faces a sturdy wall in form of privacy concerns for patients’ files, we could very soon have a pathway. GANs could create “fake” patient records that would be able to be shared and used to further the reach of modern medicine.
At a point in time where Fake News and misinformation are omnipresent, perfecting image forgery could become a big problem. Ian Goodfellow and his colleagues already work on ways to face the problems GANs could bring. Essentially security systems are in danger of being tested by GANs which are getting smarter until the attacker finds a way to bypass the security.