SkyKnit - AI inspired knitting for geeks

You may have heard of AI - artificial intelligence - and the futuristic promises of machine learning and automation in every realm of human endeavour. (Check out Google's AI program, for example, or IBM Watson).

So I guess it was really just a matter of time before someone decided to bring artificial intelligence to knitting... 

Connecting AI with creative arts is still in its infancy, although there are a few people out there using neural networks to experiment with drawings (David Ha) and with language and literature (Robin Sloan, Allison Parrish, Ross Goodwin), but all those efforts stay within the world of the computer itself. 

Janelle Shane, a research scientist in optics, has been playing with neural networks to create humour. A joint project with knitters from a group on knitting site Ravelry was her first project that took a creative AI project into the real world. 

In December 2017, Ravelry member JohannaB called for a "large collection of patterns for a knit “thing” - preferably simple, like dishcloths, or other rectangles - and folks who could scrape out JUST the knitting instructions into text and feed them to the ai."

Janelle Shane needed at least a thousand patterns of a similar type (like dish cloths) for the neural network to work. Ravelry knitters volunteered to copy and paste pdf patterns into a format that could be read by a machine, others were happy to knit up the end results. Have a look at the screenshot from the Ravelry page below!

After being fed 88,000 lines of knitting instructions, the neural network was set to create its own knitting patterns. It seems simple enough: Knitting patterns are a form of coding, with short commands such as k, p, yo, k2tog (=knit, purl, yarn over, knit 2 together) that form a knitting pattern, so it should be easy enough for a machine to learn and translate this into a unique and new pattern.

The patterns produced by the neural network certainly looked like knitting patterns, but they were only usable with substantial human interpretation or "debugging", as Janelle Shane described it in an article in The Atlantic. 

As most knitters would know, knitting patterns are different across the world and have changed substantially over time, so there is always some element of interpretation with any pattern. However, usually knitters have a good idea of what the end product is supposed to look like - whether it is a jumper or a scarf or a tea cozy. And a lot of times there is a photo or a diagram for clarification. This helps when interpreting a pattern even if the instructions may be a bit difficult to understand. However, the AI generated patterns were completely random without any intent or purpose, making this an interesting challenge for the knitters involved.

Knitters in the Ravelry group nevertheless had a lot of fun participating in this unusual "Mystery-Knit-A-Long". For more on this amazing project, check out the thread on Ravelry or go to the SkyKnit collection.