Google's Artificial Intelligence Research Division, which developed AlphaGo, developed "AI to draw letters while considering writing order by machine learning"

As seen in calligraphy and caricatures, when we see works of paintings, books, and three-dimensional works, we not only impressed the design, but also to reproduce the work after understanding the brush strokes and structure etc. I can.Artificial intelligence (AI) to reproduce a given imageHas been announced so far, but the strongest Go AI ·AlphaGoDevelopedGoogle Deep Mind"Synthesizing Programs for Images using Reinforced Adversarial Learning" (drawing program synthesis using hostile reinforcement learning), abbreviated as "SPIRAL"Is that it is a framework that updates its drawing program by himself while learning while considering the writing order and structure of the target character and image.

Learning to write programs that generate images | DeepMind
https://deepmind.com/blog/learning-to-generate-images/

"SPIRAL" is built with two neural networks. The first is to learn the change of brush size, pressure, color from computer paint software, and synthesize a program to actually draw a pictureAgentis. The other is a neural network called "distinguished person" that judges whether the outputted image is a copy of the original image or output by the program. The more you can not judge whether the image is output, the more the agent receives a lot of reward.


In order to deceive the eyes of the discriminator,Program synthesisTo improve the accuracy of the drawing program. The following image is the result of letting the agent of "SPIRAL" learn handwritten Arabic numerals and output it. The first left to reproduce at the very beginning is a number that can not be said to be a number, but when I look at the result on the left most than 25 million trials repeatedly, I can write letters that can be read with little trouble I understand that.


In this framework, in order to generate the brush control motion of the paint software, it is possible to link the simulated drawing program with the robot arm and program and actually write the letters on the campus. The program generated by "SPIRAL" and how to actually write letters using the robot arm can be seen in the following movie.

Synthesizing Programs for Images using Reinforced Adversarial Learning - YouTube


Also, if you learn not only figures such as numbers and letters, but caricatures of celebrities, you still can not describe the detail, but as if a portrait artist writes it like a face shape, color tone, hair style etc I draw it with capturing a big feature. Clicking on the following thumbnail will display the gif animation (about 8.6 MB) that the program actually generated by "SPIRAL" draws a portrait.


Google Deep Mind, who developed "SPIRAL", reproduces the ability often used by humans, "Reproduce expression structured by observation", so that we can access the tools we usually use It shows that AI can reproduce the same expression as human being.This "SPIRAL" is a small step aimed at flexible program synthesis, but similar technology is used for AI to human It is thought that they will be able to have similar cognitive abilities and communication abilities. "