I've been crossing the streams recently and letting my nerdy computer stuff overlap my drawing stuff, the outcome is that I'm currently teaching one of my robots to draw. I'm using the Google Tensorflow tools, and in particular the project magenta sketch-rnn.
I've written about this in more nerdy detail elsewhere, but this is my drawing blog, so I'll try and keep it plain. Basically the robot gets given a bunch of images and told what they are, this is the training. Then you test it with things it hasn't seen. Then you tweak it to make it better at making the right decisions.
Google did this with the "quickdraw" project and the results are really quite interesting...
The variational autoencoder takes a cue, then produces a number of sketches based on your prompt.
The downside is that the neural network has been trained with data from people who can/can't draw, from sketches dashed off in 20 seconds with a mouse.
My aim is to see what happens when I train it with a set of "good" data.
So I'm drawing a bunch of cats, and have the robot out at Grand parade collecting cat drawings from folk to help train his neural net.
Before the exhibition opened I had 42 submissions, so I'm quite chuffed... I'll add them to the growing pile of cats I've drawn....
