How good is a neural network at creating and naming paint colors?

Answer: not great

by / May 22, 2017

Turdly. Stanky Bean. Bank Butt. Maybe not the best names for paint colors — but certainly entertaining.

Janelle Shane is a research scientist interested in playing with neural networks and seeing what kinds of ideas they generate. According to her Tumblr, Shane input about 7,700 Sherwin-Williams paint colors into a neural network along with their RGB (red, green and blue) color values. The idea was to have the network invent new paint colors and give them appropriate and pleasing names.

Shane checked the network’s progress on various creativity settings and parameters. At first it played with colors that were on the tame side, mostly in the brown and gray family, and gave them less-than-descriptive names like “Caae Brae.” Later it learned to spell “green” and “gray,” but couldn’t necessarily identify those colors correctly. Some of the network’s “more creative” paint names — like “Bylfgoam Glosd” and “Dondarf” — were similarly unhelpful.

Over time, the neural network got smarter and came up with some more apt color and name combinations, such as “Burf Pink” (actually pink!) and “Homestar Brown” (actually brown!).

In the end Shane came to two conclusions based on the neural network’s output on the whole: "1. The neural network really likes, brown, beige and gray. 2. The neural network has really bad ideas for paint names."