Humans aren't nearly as creative as we think. Craft brewers, for example, have run out of fun names and are sending each other cease and desist letters for coming up with the same ideas. So, what if we let computers come up with new names for us?
That's the problem optics industry research scientist Janelle Shane has been trying to solve using neural networks, but with paint colours. The initial results are downright ridiculous. Like "Stanky Bean" and "Sindis Poop".
Tag yourself, I'm stanky bean (Image: Screenshot)
By problem, I actually mean she was just trying to have a good time online. "What inspired me was I found a post online from someone who'd done neural network cookbook recipes," she told Gizmodo. "I thought they were hilarious and I wanted more, but there weren't any more. The only way to get more was to make more."
Neural networks are essentially computer systems that can be trained on large datasets to solve problems like speech or pattern recognition. Shane analysed a list of 7700 paint colours from Sherwin Williams with a neural network called char-rnn, including both the paint names and their red, blue and green values.
Once a neural network is trained, it can learn to find the next logical thing based on an input, which is how we ended up with those strange dog pictures last year. In this case, the neural network starts with a letter, then picks the next most logical letter (or a letter further down on the list, depending on the creativity setting) to create pronounceable words. It's like a child learning to speak if its parents only spoke about paint colours.
Shane had the network spit out names during checkpoints as it was learning, at varying levels of creativity. Naturally, its most creative setting eventually started spitting out gibberish:
(I am having trouble breathing from how hard I am laughing right now.)
But eventually, it learned to make some really wacky paint names.
"I'm sure 'Hurky' wasn't in the original dataset," said Shane. "But somehow it's come up with that."
Shane previously trained a neural network to come up with new recipe names, creating some of the funniest combinations of words imaginable.
Image: Screenshot/Postcards from the Frontiers of Science
"It's tempting to correct the spelling if it almost spells a word, but somehow that takes the fun out of it," said Shane. "This is as it comes out of the computer, I'm not changing a thing."
Shane's just doing this for fun, but here's the link to char-rnn if you have your own ideas.