At the Deep Learning Summit conference here in London, DeepMind Research Scientist Ali Eslami presented a very interesting project, entitled “Artificial Intelligence and Creativity”.

Eslami’s team at DeepMind set an AI agent (a neural network which takes some kind of action) the challenge of drawing a human face, given a computer drawing program (Mypaint) and allowing it to experiment with variables including brush selection and placement, and line pressure and colour. The agent was fed an unlabelled training dataset of photographs of faces.

The results were startlingly realistic.

 

Faces drawn by AI

Drawings of faces created by an AI agent trained on unlabelled photos of people (Image: DeepMind)

Eslami’s group used a technique called reinforcement learning.

With most AI systems today, we use supervised learning, where we feed in data that is labelled in some way, so the system can compare its results with the right answer. The opposite, unsupervised learning, is where we feed in unlabelled training data and let the system try to identify features on its own.

Reinforcement learning is a form of unsupervised learning. During training, a second AI agent called the discriminator provides feedback on the results to allow the creating agent to learn. In an image generation context, the discriminator might compare the generated image to the training data and give feedback on whether it could tell the difference. This feedback might be a score which quantifies how difficult it was to tell the difference between the generated image and the training dataset.

 

Reinforcement learning diagram

Reinforcement learning uses two AI agents – one creates images and the other tries to tell if they are real or not (Image: DeepMind)

Before DeepMind taught their system how to draw, they taught it to write. The AI was originally trained last year on images of handwritten letters and characters from various alphabets (MNIST and Omniglot datasets), which it recreated, very successfully. The team was surprised to see that restricting the numbers of strokes that were allowed produced results similar to what a person would do when writing in a hurry; dots and smaller features joined together. DeepMind even connected the algorithm to a robot arm with a paint brush to produce calligraphy.

Once the AI worked for handwriting, Eslami’s team scaled the system up to use bigger networks and trained it on more CPUs. Using the photos of faces as a training dataset, the images it drew started to become more realistic. The image below shows the stages of the drawing process; note that the AI is not given a target image, it is just creating an image it thinks looks like a face. Remember, the computer has never seen a person drawing, it discovered everything about how to draw by trial and error through reinforcement learning.

Eslami said that there are actually two complex tasks going on here — one is controlling the brush with a high level of precision, the other is managing its time, trading off how real it makes the image look during the process compared to how real it wants the end image to look.

Stages of AI drawing

A sample drawing of a face created by AI showing the stages of the drawing process (Image: DeepMind)

The next thing the team did was to ask: what happens if we make it harder? They reduced the number of strokes allowed from 1,000 to just 20.  To their surprise, the agent was still capable of producing images that suggested faces, albeit rather more abstract. Eslami said that the most striking thing about the abstract images was that the agent had clearly identified the features that make up the face — eyes, nose and mouth — and that these abstractions they previously thought could only be taught through imitation, or with supervised learning, are indeed possible with reinforcement learning.

Abstract faces drawn by AI

Samples of abstract “faces” produced by different AI agents with different hyperparameters (Image: DeepMind)

So AI has learned to draw successfully, but is this creative or is it just random? And is it art?

You could argue that the AI agents have used creativity in trying to represent the face in a variety of ways. The drawings in the image above certainly have more differences than similarities, even if they all suggest faces. But the fact remains that the AI’s intention wasn’t to abstract the face into its most basic elements in its drawing, or to produce a picture that evoked an emotional response, its aim was realism, and it judged its success on how real the drawings looked.

It could also be argued that the images are drawn with a high level of skill, but does that make AI an accomplished artist? It definitely and produced better images by the end of the training process, even learning to start with blurry strokes and add sharper strokes towards the end.

Unfortunately, there is no strict definition for what art is today. I expect that means it’s up to the viewer to decide.