Draw a phone in less than 20 seconds. "If you do not have by profession or by hobby to scribble animals in record time, fulfilling this order may require an effort of abstraction that you do not do since your first years in school. you have more homework. "Draw some skates." "Draw a pear." "Draw a fan." "Draw a cow." "Draw a watermelon" …
The challenge is posed by Google in its experiment Quick, draw! -Quick, draw! – In which has put a neural network to identify the scribbles drawn by more than 15 million souls. "When you draw something, try to guess what you are drawing, obviously it does not always work, but the more you play, the more you will learn," they explain.
The rules are simple: six drawings and twenty seconds for each one. You do what you can and the neural network too. "Our goal is to show an example of how you can use machine learning in a fun way." After the sixty seconds needed to complete a round of the game, you will see how you have done it. And you can also go a little deeper into the failed attempts.
For example, the profoundly incomprehensible fact that the neural network has not been able to identify the most perfect phone that the History of Art has ever seen -hemhem- and if it has identified a dangerously abstract cow seems to greatly influence the fact that the majority of users choose to draw mobile phones and relegate the poor fixed to a second background.
In the cow, on the other hand, horns and painted spots appear to be decisive at full speed, although the neural network was also considering other options: the second most similar was a dog. The third, a rhinoceros.
- And what does it matter to humanity?
The result of the effort of the fifteen million cartoonists who have tried to portray some of the 346 categories included in the game – all the above, planes, ambulances, bananas, bridges, mustaches … – is a database of more than fifty million drawings completely labeled with their names, the date and time of creation, the country in which their creators were and the doodle in question.
Thanks to her, mere mortals can kill the time seeing how humanity faces the task of drawing a camel, with 115,554 examples. SPOILER: there are up to four humps.
For developers, it's a dataset unique for the training of new neural networks; for the scientific community, a place to look for patterns in the ways of drawing people from all over the planet; for artists, a source of inspiration and, more likely, horror.