Artificial intelligence enthusiast Janelle Shane had just finished painting her living room when she began a paint job of a different shade: an AI experiment capable of creating new color names merely from existing paint colors.
In this exercise, Shane gave a recurrent neural network—which is just a kind of artificial intelligence that can generate new data (in this case, color names) from an original dataset—a list of about 7,700 Sherwin-Williams paint colors, along with their color values. The neural network was then asked to produce sequences of letters to form new paint color names, while coming up with sequences of numbers that map to RGB values.
It’s not unusual for paint colors to have flowery, even fantastical names. There are people out there who get paid to create them. Shane wanted to see how well a neural network could manage the same task.
Not very well at all, as it turns out.
At first, the output resembled a menagerie of little phonetic nightmares. The Atlantic even pointed out that “Caae Brae” sounds like it could be a Beowulf character. Shane gave the neural network more time to process. And the result? “Stanky Bean,” “Bank Butt,” “Dorkwood,” to name a few.
Ars Technica joked that this experiment would quell any fears of “the robot uprising.”
But a scientist’s job is never done. Last week, Shane issued an update.
At the request of multiple readers of her blog, she tried swapping out RGB values for other color representation values in hopes of better results. She tested HSV values (Hue, Saturation, and Value), but these changes didn’t make a noticeable improvement. She lowered the “temperature” so that the neural network would act more conservatively in choosing the next characters in a sequence. That helped a bit.
What really tipped the scale was a much larger dataset, compiled by one of Shane’s readers. It contains paint colors from additional companies such as Behr and Benjamin Moore, as well as user-submitted colors from a XKCD survey.
This time, the AI did much better. It generated paint color names that were actually descriptive of the colors themselves, rather than just flowery assortments of gobbledygook.
“The answer seems to be, as it often is for neural networks: More data,” says Shane.
“I was pretty happy with this outcome. It was impressive how it would come up with the color that gray candy would be. And it is kind of gray and pink and frosty still.”
Shane says she finds new little gems each time she reviews the list, like a “Copper Panty” or a “Shivable Peach.”
“I think one of my favorite ones that the new neural network came up with was ‘peacake bring,’” she says. “What does that mean? I don’t know and that’s why I love it.”