Machines draw and recognize your doodles
Yesterday, David Ha posted a new article on the Google Research Blog. He is explaining how their machine learning research is taking shape and beginning to show interesting results.
When teaching machines it’s not like you are teaching children. Machines have no baseline and no past experiences to draw their information from. For instance, you need to define how cat doodles look like, but also you need to explain to them why the doodle can’t be a pig. And the other way around.
Even if you let machines now reconstruct human input that is faulty, they will correct flaws, based on what they have learned to be right. David Ha writes, “when we feed in a sketch of a three-eyed cat, the model generates a similar looking cat that has two eyes instead, suggesting that our model has learned that cats usually only have two eyes.
To show that our model is not simply choosing the closest normal-looking cat from a large collection of memorized cat-sketches, we can try to input something totally different, like a sketch of a toothbrush. We see that the network generates a cat-like figure with long whiskers that mimics the features and orientation of the toothbrush. This suggests that the network has learned to encode an input sketch into a set of abstract cat-concepts embedded into the latent vector, and is also able to reconstruct an entirely new sketch based on this latent vector.”
Visual communication supports transporting a message. When we want to explain something to others, too complex for words, we often tend to draw sketches that go along with what we try to explain.
Now not everybody is an artist and in order to support the quality of visual communication, you can leverage the fast drawing system from Google. You don’t even need to download anything. Use your device’s browser and go to their AutoDraw website in order to use it.
We’ve seen the first of this from the tweet by Hantz Févry, Lead Innovation for EMEA region:
The quick draw game
The question that Google’s quick draw game is asking is whether or not a neural network can learn to recognize doodles. They say, “Help teach it by adding your drawings to the world’s largest doodle data set, which could be shared publicly to help with machine learning research in the future.”
The task is to draw something they tell you in less than 20 seconds and the goal is that the system recognizes it within that time limit. You can try the game out at the Quick Draw website and share your results on your social media networks if you want to.
More about Project Magenta
Project Magenta was made available via open source. There are various touching points here that go beyond machine learning and extend to subjects such as art and creativity. The underlying theme is teaching machines and systems to generate and produce music and other arts. You can find out more on the Project Magenta website.
YouTube: AutoDraw: Fast Drawing for Everyone