There is a lot of buzz about AI and the prospect that computers will soon be doing something hugely different than what they’re doing now. It’s apprehension of what Ray Kurzweil calls the singularity, except that people don’t call it that much anymore.
Under the headline An AI invented a bunch of new paint colors that are hilariously wrong, Annalee Newtiz discusses the result of training neural networks on Sherwin-Williams decorator colour names. The original work was done by Janelle Shane, who recounted it in a Tumblr post.
Some whinging below the fold.
UC San Diego was a hotbed of neural nets when I started grad school there in the mid 90s. Connectionism, the view that neural nets are the key to understanding minds and cognition, was in full swing. Yet neural nets were already a well understood tool, and the work was just in applying them to all sorts of things.
Yet the fires of connectionism have cooled. Philosophers have accepted the limitations of neural nets as a model of actual cognition, and they are pretty well black-boxed as algorithms for pattern recognition. Yet, in the popular press, neural networks count now count as AI. As far as I can tell, recent deep learning work that pops up on my news feeds is just bigger and faster than work that people were doing in the 90s. It’s the inevitable result of bigger drives and faster processors rather than any harbinger of our robot overlords.
Someone actually naming paint colours imposes a mapping from the paint as a colour to the name as a bit of language. Word associations matter. I really wanted to lead with a joke about painting the kitchen Sink, but damn that would be an ugly kitchen. And it’s funny to imagine the customer who decides to paint a room in Grade Bat with accents in Grass Bat because the names suggest they belong together.
The algorithm isn’t naming colours so much as reflecting the Sherwin-Williams catalog in a funhouse mirror. The swatches juxtapose a colour and a name, but the network is just working on strings.
Shane encoded the Sherwin-Williams catalog as strings of names followed by decimal RGB values. The network ends up with a lot of greys and browns, because most of RGB colour space is grey or brown. Since paints are mixed pigments, CMYK would arguably be more appropriate. My hunch is that HSV would be a better choice for yielding fewer muddy outcomes, but it might have been darker overall.
The choice of representation, made in setting up the algorithm, is doing a lot of work. And understanding that is more important than framing it as the brave new age of AI.
 Since Kurzweil coined the term a bit more than a decade ago, I guess “singularity” is too old a name for it to sound cutting edge.
 To be fair: Neither Newitz nor Shane call this “AI” in the body text, and Shane is fully aware she’s just using an off-the-shelf tool to generate some funny results.