With an incredible speed, ability and accuracy, artificial intelligence becomes proficient at identifying things while discerning through billions of pictures for instance identifying a squirrel in the woods.
However, it cannot create a new squirrel from its current library of files. We can translate this to people – the AI can identify a person entering a building, but it cannot create that person entirely from what it has. If AI was able to create people it would show numerous “realistic” but made-up pictures of people that do not exist or they cannot exist, for example the person with three legs. So, how an AI can get to create a brand new person?
Through the Generative Adversarial Network approach, or GAN which was first developed by Ian Goodfellow of the University of Montreal in 2014 while in the PhD program, he takes two neural networks while set against each other in a digital cat-and-mouse game. Modern machine learning follows the simplified model of human brain including this mathematical pattern of the neural networks.
Now both AIs are put to test to create something from “imagination”. This is the first time this is done and the way this is accomplished is explained below:
- AI use GANs to generate sounds speech from real sounding patterns and also creates realistic looking photos albeit fake. Nvidia company conducted a GAN research that stored celebrity photographs and resulted in generation of hundreds of credible faces of people who don’t exist. Another research showed how both AI networks using GANs created fake pictures of famous Renaissance painters.
- Both networks are trained on the same data set. One AI has already seen images and it is trained to create variations on these images for example a picture of a dog with five legs. The discriminator the “opposing” partner of the generator is asked to identify whether the example it sees is likely to be fake or real. In other words, could that five-legged dog be real?
- While exercising and developing images between the generator and the discriminator, it may happen that the discriminator fails to recognize the fake. The generator becomes better creating real images since it has been taught first to recognize such images of people, dogs, etc, and then create realistic photos of them.
A few companies play with these ground-breaking AI technologies: Google Brain, Nvidia, DeepMind.
Although some results are not perfect, these machines develop some sense of imagination, and later could turn them less dependable on humans and through the incorporated tech, they can create such a credible fake reality. With a few downsides that AI evolution could work on, this technology appears as one of most promising advances in AI in the recent past.