Neural networks – aka artificial intelligence programs designed
to mimic the way the human brain learns – are changing the
world. They’ve been trained to recognise faces or handwriting;
keep driverless cars on the road; translate words between different
languages; and detect fraud. Now, Colorado research scientist
Janelle Shane is showing they can also be put to more frivolous
uses. Like, for instance, generating bizarre knitting patterns.
“I’ve always had an interest in machine learning, especially this kind
of programming where you set the computer up with an environment
in which it can learn, then just see what happens,” Janelle says.
To ‘teach’ a neural network a particular task, loads of sample data is
fed in for the system to analyse – it uses this to identify ‘rules’ about
the data, which can then be applied to create more of its own. Janelle’s
first foray into AI experiments came after watching a fellow researcher
use neural networks to create recipes from scratch. She began
running her own trials, feeding “bare bones”, open-source algorithms
unlikely data sets. A list of paint colour names wielded comical results
(sample shades generated included Navel Tan, Shy Bather and Clardic
Fug), and a search for Halloween costume ideas was similarly silly
(fancy dressing up as a Shark Cow or Panda Clam?).
“What I love about it is all the surprises; it solves problems in
ways you don’t expect,” she says. “It’s a lot of fun seeing what the
program manages to get and what it doesn’t. It’s inimitable – there’ssomething about the text it spits out that’s too hard to just make
up.” Massive commercial algorithms used for systems like facial
recognition will spend months ‘learning’ on powerful computers,
but Janelle works with rather more limited resources. “Mine is
definitely a baby neural network – it has about as much brain power
as a worm. So, if you imagine a worm devoting its entire being to
recipes instead of doing wormy things, this is about what it would do.”
Janelle was intrigued when a knitter suggested using neural
networks to create random knitting patterns – tapping into the
collective power of knitting forum Ravelry to help. “I looked at a
knitting pattern and thought, ‘Whoa, I can’t understandany of this.
There’s no way I’ll be able to tell if it’s doing the job right or not,’” she
says. “Being introduced to the knitters was wonderful; they have so
much depth of knowledge. They’d laugh at jokes the neural networks
had written, and I just didn’t get it. The knitters would say, ‘That’s the
funniest thing I’ve ever read,’ and I felt really out of the loop.”
Janelle called the task Project Hilarious Disaster; the Ravelry
community kindly changed the name to SkyKnit. Set in motion, the
neural networks generated more than 500 patterns with names like
‘Fishcock’ and ‘Tiny Baby Whale Soto’, giving little clue as to their
eventual forms. Then the knitters jumped in to give them a go, with
varying results. “Nobody knew what they were going to be creating,”
Janelle says. “They didn’t know if it was going to be a tangled mess
or have some kind of structure. In many cases, we’d get two different
knitters doing the same pattern, and they’d end up with completely
different results. One person said, ‘I’m treating this like it was written
by a very elderly relative who doesn’t speak much English. The
intention is there, but I have to figure out what they mean.’”
Janelle’s not convinced the neural networks ever learned to count
stitches properly – “Whatever they get right, I’m pretty sure is by luck
or probability” – but she enjoyed the project so much that there’s
another craft-based experiment in the wings. “When the crochet
community saw what I was doing with knitting, they said I should try
crochet. I’m now collecting the dataset for a crochet project I’m calling
‘Hat3000’. I want to see what will happen if we just focus on hats.”rise of the machines
JANELLE SHANE USES A.I. TO
PRODUCE A WHOLE LOT OF WACKY
KNITTING PATTERNS.Wor d s Lucy CorryIllustrationAshley Ronningmy project