What would you say to one of the latest discoveries in artificial intelligence (AI) that makes it possible to design machines that possess imaginative power?
It is difficult to imagine machines truly exhibiting creative behaviour because we have always viewed them as tools that do not possess cognitive abilities. We are humans, and they are simply machines.
However, recent AI developments—including self-driving cars, virtual assistants, medical and investigative robots, computers running global finances and more—may have brought us closer to learning machines. But how does this affect the way engineers design and build innovative machinery that depends on AI?
I believe that modern AI and machine learning applications are what Alan Turing, a computer scientist and (AI) theorist, had in mind when he proposed a test to settle machine intelligence in a seminal paper titled “Computing Machinery and Intelligence.” This paper was his bid to answer the question “Can machines think?”
I suppose Turing would have been very excited when IBM’s Deep Blue supercomputer beat Gary Kasparov—the world’s preeminent chess grandmaster—in a chess game. He also would have probably danced for joy at a more recent machine versus human competition. Google’s AI AlphaGo program beat the world’s best in one of the most complicated board game, Go.
Coming from the Turing test and contemporary AI innovations integrated with years of research, development, and evolution, computer engineers have presented the fact that machines can truly be taught to mimic human intelligence.
Technology That Gives Machines Ability to Imagine?
Ian Goodfellow, the man behind the innovation that gives machines imaginative capability, brought us into an era that he himself thought impossible at first. He had met with a group of friends who happened to ask for his help on a project they were working on—developing a computer program that could create images on its own.
His friends were proposing to apply complex statistical analysis techniques to help the computer generate decent images on its own. However, their proposal meant a massive amount of power consumption that Goodfellow thought was impossible and wouldn’t work.
However, he kept pondering on the idea even after parting with his friends, wondering what would happen if someone pitted two different neural networks against one another. At the time, researchers had been using neural networks to create generate new data of their own, but the outputs have always been very disappointing. They often resulted in blurry and erroneous images with missing components.
Goodfellow didn’t stop on the idea he had in mind. He implemented it and it worked. By pitting neural networks against one another, he actualized the generative adversarial network (GAN), a machine learning technique that has sparked lots of excitement in the field of AI and machine learning ever since.
Generative Adversarial Networks (GANs)
In the recent past, AI researchers and engineers have made remarkable progress with machine learning capabilities. By supplying a computer with numerous images, deep-learning enables it to learn and recognize objects. For instance, it can recognize a pedestrian who’s about to cross a road or traffic lights, gearing innovations like autonomous cars and voice technology. However, deep-learning systems cannot generate objects and only go as far as recognizing them. GANs technology is the new innovation that comes as close to giving machines a power that is somewhat similar to human imagination.
The concept behind GANs is two competing neural networks trained with the same data set. The first neural network is referred to as the generator. It receives noise as input and is therefore charged with generating samples, say photos or writings that can be as realistic as possible. The second one, the discriminator, receives samples from the generator and compares them with original photos from the training data to determine which ones are real and which ones are not.
Like a competition, the two models play this game back and forth, with the generator producing more and more convincing samples. Meanwhile, the discriminator learning gets better and better at differentiating real from generated data. Here’s the picture, it’s like a person forging a picture on one hand and another playing detective to see who will outwit the other. The generator model continuously adjusts its parameters for creating new images until the discriminator can no longer tell the difference between fake and real.
From the time GANs were defined, they have been primarily applied in natural image modelling and are now producing outstanding results—sharper images delivering higher quality compared to models trained with other generative techniques.
This is not a juncture where we say we don’t know what the future will bring. AI and machine learning systems are poised to revolutionize our age and transform industries across all fields. By deriving data from large data sets, these systems have the capability to display true brilliance and originality.
Moving on, I foresee a future where we will be working and interacting with machines that truly imagine and it’s not far off. Where GANs stand presently is just the beginning.
AI researchers and engineers are continuously aiming to develop more flexible approaches whereby software systems can continually learn from new data representations, using it to make modifications that exhibit creative behaviour.
At this point, the emphasis is now leaning towards creating intelligence that is unique for the machine and intelligence that can be harnessed to amplify AI, progressing machine learning capabilities in the long run.