Self-expression through painting is a worthy goal, but most of us are not far from the children's drawing technique "stick, stick, cucumber, here comes the little man." In the fast-paced age of artificial intelligence development, there is no time and desire to hone painting skills, so the generous company Google announced a new service that takes all the hard work on itself.
The AutoDraw system grew out of a handwriting recognition project - when the developers realized that they were skilled at working with hieroglyphs and text in pictures, they decided to go further. In a simplified form, this is a new mechanism for auto-correction of input data, which analyzes what the user has scribbled with his finger or cursor, and then does "how to do it right". Replaces curved lines with smooth lines and fuzzy images with specific objects.
The service is based on a neural network, the pride of Google, since the company without false modesty believes that they were the first in the world not only to assess the potential of machine learning systems, but also to deploy their mass use. To learn to distinguish between scribbles, the neural network "read" millions of entries in a variety of handwriting. And now her heiress is being "fed" with thousands of hand drawings so that she learns to understand - how do people themselves depict what they mean?
In parallel with the recognition of abstract patterns, the system learns to draw up patterns with how people call and describe them. In order to be able to draw a cat by request "cat" or call a creature with triangle ears and a tail a cat, not a mouse. To test the accumulated experience, AutoDraw was once asked to draw a three-eyed cat, but the system gave a negative answer - it has already firmly learned that a "normal cat" has only two eyes.
The task of creating, creating new drawings, at this stage, is not given to the neural network at all. Most likely, the first commercial implementation will be in the format of an add-on for a search engine or messengers, which converts an arbitrary picture into a set of tags for working with them according to standard algorithms. And at the same time gives it a beautiful, finished look - to the delight of those who cannot draw.