Hunt Club Red, Plunge Pool, Stanky Bean, Gray Pock, and Burf Pink.
These aren't the names of underground punk bands, they're unique new colors you might soon see on kitchen walls. The first two were created by designers at PPG, a leading industrial chemicals firm and Fortune 500 company. The others were generated by a neural algorithm trained by Janelle Shane, an optics research scientist, who taught it to analyze patterns in colors and dream up new hues in the style of Sherwin-Williams.
Below, find one output of her algorithm:
Shane stumbled on the idea of teaching an algorithm to learn to create and name colors after reading a conversation on Twitter about whether colors could be invented by code. She learned Sherwin-Williams had published a list of its 7,700 color names in a format computers can easily understand.
With that data, Shane was able to teach the algorithm to look for relationships in data and produce and name new colors. An algorithm trained on Benjamin Moore or Pantone paints would find entirely different relationships and generate colors and names fitting a pattern unique to those companies.
Shane's algorithm did not learn to generate Sherwin-Williams paints right away, however. The data was learned in stages. Below shows an early phase, when the algorithm was still learning understand the difference between "blue," "green," and "red."
At the second checkpoint, the algorithm can properly spell green and gray. However, it still does not know the relationship between the name and the color.
As the algorithm reads and sorts more data, it gets better at recognizing colors. By this point, it is fully trained and is able to grasp and generate white, red, and gray.
Likewise, in this fully-trained example, grays, whites, and blues are output.
Shane posted her full results in a Tumblr post that has gone viral.
Shane used an algorithm called a neural network, which finds patterns in data similar to how the brain's neurons share information in a web structure.
On a technical level, maybe the most remarkable part of Shane's algorithm is that it has no concept of what a word is. Instead, it learns to see the relationships between characters — how they are spaced, which characters tend to be next to each other — and simply replicates that pattern.
That might be the biggest difference between a computer color designer and a professional human.
"The designers are pulling from a larger sense of what these words mean and their associations," Shane says. "So because of that the algorithm will come up with words like stanky or horrible and it has no idea that it has these negative connotations. They're just words that resemble other words in the dataset."
For now, her machine's lack of restraint also amounts to its greatest weakness, Shane says. "Human designers have an edge over the neural network in that they can recognize bad ideas when they see it."