LONDON — A new artificial intelligence system can turn simple sketches into paintings reminiscent of works by great artists of the 19th and 20th centuries, researchers say.
The artificial intelligence (AI) system, dubbed Vincent, learned to paint by “studying” 8,000 works of art from the Renaissance up to the 20th century. According to the system’s creators — engineers from the United Kingdom-based research and innovation company Cambridge Consultants — Vincent is unique not only in its ability to make art that is actually enjoyable but also in its capability to respond promptly to human input.
“Vincent allows you to draw edges with a pen, edges of a picture you can imagine in your mind, and from those pictures, it produces a possible painting based on its training,” said Monty Barlow, director of machine learning at Cambridge Consultants, who led the project. “There is this concern that artificial intelligence will start replacing people doing things for them, but Vincent allows humans to take part in the decisions of the creativity of artificial intelligence.” [Super-Intelligent Machines: 7 Robotic Futures]
Some previous attempts to produce AI-generated art delivered rather scary results, such as the human portraits drawn by the Pix2Pix tool that was introduced earlier this year by Dutch Public Broadcaster NPO. Pix2Pix used sketches drawn by humans as a starting point and attempted to turn them into what is meant to resemble an oil painting of a female face. The creations, however, looked more like they were pulled from a horror movie.
While Vincent’s art doesn’t look entirely realistic, it could pass for some of the more abstract creations of masters of the impressionist or expressionist era, such as Vincent van Gogh or Edvard Munch.
“It has learned contrast and color and brushstrokes,” Barlow told Live Science here at the Re.Work Deep Learning Summit on Sept. 22, where Vincent was first presented. “It can bring all of that to play when you draw a picture, giving you access to all that artistic content.”
Barlow said that using only 8,000 works of art to train Vincent is by itself a major achievement. Previously, a similar system would have needed millions, or even billions, of samples to learn to paint.
“Most machine learning deployed today has been about classifying and feeding lots and lots of examples into a system,” Barlow said. “It’s called supervised learning. You show a million photos of a face, for example, and a million photos of not a face, and it learns to detect faces.”
Vincent uses a more sophisticated technique that allows the machine to teach itself automatically, without constant human input. The system behind Vincent’s abilities is based on the so-called generative adversarial network, which was first described in 2014. The technique uses two neural networks that compete with each other. At the beginning, both networks are trained, for example, on images of birds. Subsequently, one network is tasked with producing more images of birds that would persuade the other network that they are real. Gradually, the first network gets better at producing realistic images, while the second one gets better at spotting fakes, according to the researchers.
“For Vincent, we had to combine several of those networks into a fairly complicated circuit,” Barlow said. “If you asked us five years ago how much art we would need to train this system, we would have guessed that maybe a million times more.”
To speed up the learning, the researchers occasionally continued providing the machine with feedback on the quality of its creations. [Gallery: Hidden Gems in Renaissance Art]
The need for extremely large data sets to produce reliable results is a major hindrance to the use of AI systems in practical applications. Therefore, researchers are trying to design new techniques that would allow machines to learn faster in different ways.
Barlow said a system such as the one behind Vincent could, for example, help teach self-driving cars how to do a better job of spotting pedestrians.
“If you want an autonomous car to reliably detect pedestrians, you can’t just have a face detector, because you can have faces on billboards, on the side of buses, and equally, some pedestrians might be wearing a hood or walk in a shadow; you wouldn’t even see their face,” Barlow said. “To even train a system that would reliably decide that something dangerous is happening on the road — that somebody has walked out — you need a ridiculous number of examples in different weather and lighting, with different people and heights.”
To collect such an enormous amount of data is, according to Barlow, nearly impossible. Systems such as those behind Vincent could use their creative abilities to generate more images from a limited data set. The system would, with a little bit of human help, learn to synthesize realistic images and subsequently teach itself to reliably evaluate all sorts of real-life scenarios.
“It’s a virtual circle where not only can machine learning do some amazing things, but it is in itself helping to drive forward the progress of machine learning,” Barlow said.