Every artist once started with something. Today we can apply this catch phrase in relation to machines. What does it take to create creative artificial intelligence? Sometimes it seems that this difference between machines and humans, machines will never catch up. However, AI is already showing a growing penchant for creativity, whether it be composing a heavy metal rock album or creating an original portrait strikingly reminiscent of the Rembrandt brush.
Applying AI to the art world may seem like overkill: there will always be people who create great work. Proponents of this approach, however, say that the real beauty of AI teaching creative skills lies not in the final product, but rather in the potential of technology to extend its own machine learning, to learn how to solve problems outside the box, faster and better than humans. For example, a creative AI could one day decide to save the lives of the passengers of a self-driving car if its sensors fail, or suggest an unconventional combination of chemical components that would lead to a drug that could treat previously incurable diseases.
AI with creativity will be essential for the design of highly automated systems that can respond appropriately to human life, says Mark Riedl, professor at Georgia Tech's School of Interactive Computing. “The fact is, we do something creative every day, many problems are solved creatively,” he says. "If my son's toy gets stuck under the chair, I'll have to take the tool out of the hanger and get it out."
Riedl notes that human creativity is also important for social interactions, even, for example, to tell a joke or recognize a pun. Computers cannot handle such subtleties. For example, an incomplete understanding of how humans construct metaphors led the AI to write a new chapter of Harry Potterra, filling it with meaningless sentences, such as, "The floor of the castle looked like a big heap of magic."
Still, getting the machines to accurately mimic the human style - Rembrandt or Rowling, it doesn't matter - is a good start to creative AI, Riedl said. After all, creators often start by imitating the skills and processes of established artists. The next step, for both humans and machines, is to use these skills as part of a strategy to create something original.
Modern AI programs are not advanced enough to spontaneously compose hit songs or works of art. For an AI to do this, a person needs to calibrate the program by feeding it a huge number of examples. German Mario Klingemann, for example, designed a neural network capable of composing strange, frightening images from existing photographs and other works. A neural network is made up of a series of interconnected processing nodes that resemble the neural structure of the brain. In a neural network, each electronic "neuron" takes an array of numbers, performs simple calculations based on that input, and then sends the result to the next layer of neurons, which in turn performs more complex calculations.
Klingemann's approach involves feeding source material, drawings, and photographs to generative adversarial networks (GANs) that combine the power of two neural networks. One network generates images united by a specific theme or set of conditions; the other evaluates the images based on its knowledge of these conditions. Thanks to the feedback from the second network, the first network gradually gets better and makes the images more and more relevant to the given topic. “These networks are now just tools to complement our own creativity,” says Klingemann. "We humans still need to recognize creativity or innovation." Its goal is to create an artistic neural network that can independently select and even publish its best work on a given topic.
GANs are now strictly used to create new content or images in the broader creative system, says Alex Champandard, founder of creative.ai, a startup that develops AI tools for creative people. GANs can produce a lot of material, but still rely on humans to determine their conditions.
Promotional video:
Content generation is a good start for developing AI that can solve real-world problems, according to Ian Goodfellow, a Google scientist working on the concept of GAN. Goodfellow is working on machine learning models that allow computers to write dynamic narratives that go beyond the limited scenarios (like planning chess moves) that computers have long excelled at.
Let's take a classic example of planning that people do all the time: when we go to the airport, we often sketch out a rough map - in our head - key travel points, traffic jams or transfers. GANs may plan such a trip, but they will do so in great detail and will offer many routes. We, in fact, need a layer of computing network that will skip all these options and intuitively choose the best one.
Another key component of human creative thinking is the ability to take knowledge from one context and apply it in another. George Harrison takes a sitar and plays it like a guitar. Shakespeare takes stories from Greek mythology and writes an English play based on these stories. An executive director uses knowledge of military strategy or even chess to plan a business deal.
For this reason, experiments are being carried out to help AI algorithms that can mix and match material. For example, scientists at the University of California, Berkeley are using the CycleGAN network to turn horse videos into zebra videos. The AI detects the basic shape of the horse in the first video and plays with the image over the video, instantly and imperceptibly replacing the horse's brown torso with a striped zebra as it moves. Such work will help the AI of a self-driving car to adapt to unfamiliar conditions and avoid accidents.
Artificial intelligence should not only learn the rules, but also break them, like a real artist.