Generative Adversarial Networks Promise to Change Computing, and Us
You have heard of machine learning, in which a computer gains “knowledge” and expertise as it tries and fails at a human-designed task until it learns the correct approach. That type of machine learning is still being used, but it could soon become very old-school. Some day, generative adversarial networks will take over, as two computers test each other, with little human intervention, to come up with a solution.
Of course, humans still must program the computers before they start on their task, but with generative adversarial networks, or GANs, the computers then test themselves to develop solutions that have minimal human input.
GANs are confined at the moment mostly to creating paintings and simulated photographs, but scientists say the technology could blaze the path to computers that think like humans – or come so close most of us won’t be able to tell the difference.
Positive or negative?
If you dread the singularity, when we become one with technology, this might be the time to start hoarding the freeze-dried food pellets at your mountain hideaway.
If you are more optimistic, you might see GANs as moving us closer to a world in which information technology is maximized to help us solve more of humanity’s problems.
In generative adversarial networks two computers teach each other to solve problems. With conventional machine learning, a human feeds labeled information to a computer until the machine learns a task.
How it works
GANs pairs two networks to share and compare information. One network generates data sets. The companion network “discriminates” to determine which data are real and which are fake, and thus should not be part of the set. As the discriminator network learns what is accurate and what is not, the generative network develops new information that includes both accurate and inaccurate data. Within this loop, the discriminator network learns how to identify what works. It uses its knowledge to, for example, write a speech, compose a song or create an image that looks like a painting.
GANs produces data that is not something humans already labeled, or taught the computers to label – the type of information that today’s computers are so good at tracking and analyzing.
GANs images, songs and writing are not copies of something humans already created, but something original – and possibly artistic. In fact, the networks teach themselves unsupervised, much as humans learn through trial and error.
Imagining a brave new future
Potential uses for GANs output could be in creating more life-like real-time options in video games or to help jet fighter pilots recognize visual threats more quickly. In the longer term, scientists are imagining a computer network that can teach itself how to solve complex, unique problems it has never before encountered and to interact as people do with other people.
After that is the territory of true science fiction, in which robots and computers express believable (or possibly real) emotions.
We’re not there yet, or even close, partly because our own brains, even while working within human networks, haven’t figured out the missing pieces that will allow computers to gain such skills.
In the meantime, existing machine learning will continue to help the human race, but not as equals.
