Experiment with me: Draw together with a neural network

How AI can improve our understanding of human creative thought processes

3 years ago   •   3 min read

By Angelique Tzanakakis

Welcome back to my series where I review AI experiments - you can find the last blog here. Today I am delving into the world of machines that can draw. Sounds a bit strange, if you ask me, but stay with me here - it gets interesting.

Humans have used visual abstractions to communicate with one another for centuries. These images not only reflect the world but also how humans interact with the world around them. Think about it - from a young age, we are encouraged to draw pictures for many reasons which I will not get into. Training machines to draw has applications beyond a fun experiment. One such possibility is that these models can be used by scientists to better understand human creativity.

There are quite a few models available online all of which behave a little differently. In this post, I’m going to present three of them - Magic Sketchpad, Multi-Prediction Demo, and Quick, Draw! (you can click each heading to try the demos for yourself).

Magic Sketchpad

This experiment works by you choosing an initial prompt - I chose a bicycle - and then you draw an initial shape. To draw the shape, you click and hold your mouse and then draw like normal. Be warned - it’s a single continuous drawing as the machine learning algorithm responds as soon as you stop drawing.

Here’s my attempt (p.s. I have outlined what I initially drew in red to show you what I drew since the machine uses the same colour):

It’s obviously not perfect but they’re fairly recognisable. There are many shapes to choose from including some unusual ones like the Mona Lisa and a rabbit turtle.

Multi-Prediction Demo

Similar to the first experiment, this demo also predicts the rest of your drawing. It also has a wide range of prompts for you to choose from. Unlike the first demo, this one allows you to carry on with your drawing to help it along. The model is trained on a database of other people’s doodles.

The outcome isn’t perfect and I had to help it along, but it’s still very interesting to see the outcome.

Quick, Draw!

This one is my favourite and for a good reason - you get 6 prompts and 20 seconds to draw the doodles. The model has been trained by millions of people’s drawings which provides valuable insight for developers, researchers and artists.

You can browse the collection of drawings by clicking on the images on the front page or you can play the game. There is a computer generated voice that guesses as you doodle, and when it guesses correctly or time runs out, it moves on.

Here’s one of my attempts:

Clearly I am not great at it, but I still had fun and it’s very interesting to see how these models work.

Final thoughts

These experiments have a purpose that goes beyond learning how AI works and I think that’s pretty neat. The only downside is that these demos work best on desktop rather than on mobile devices. Would I play with them again? Maybe… Look out for future instalments and feel free to send me your suggestions and favourite experiments on Twitter, LinkedIn or Facebook. If you’re still here, here are the links again: Quick, Draw!, Multi-Prediction Demo, and Magic Sketchpad.

Spread the word

Keep reading