The researcher who created an "artificial intelligence weave" tells the secrets of his invention

"If someone wants to knit a sweater for his kitten, he should do it today by hand." We want to change that, "says the PhD student at the Mbadachusetts Institute of Technology who created a artificial intelligence able to create woven clothes.

With this project, Alexandre Kaspar wants democratize the fabric, make it accessible even to those who have never held two needles that turn the ball into a functional element. "I hope these machines will be able to be used as current 3D printers," he says during a dialogue with TN Techno This expert from MIT CSAIL, the division of this university specializes in AI and IT.

The most relevant mechanism that has created his team is "InverseKnit", a software which automatically creates models based on tissue images and translate visual information into instructions that finally the knitting machine meets.

– How was born the idea?

– In recent years, our group has been working on mapping the physical properties of microstructures. In this way, we wanted to check if we could reverse the weaving process from images. There were many reasons for doing so, especially because of the obvious regularity of stitches in knitting.

Our "eureka moment" arrived in December 2018 in an MIT clbad. One of the students wanted to create a woven sheet with sets of tubular structures incorporated. We discovered how to design such a structure, then realized that I could use a slightly more complicated version to create larger knitted panels on which I could weave several patterns at a time, with regular tension provided by rods of steel inside. Tubular structures mentioned above.

I quickly created an initial prototype and it seemed like a viable option to expand our dataset. Then I spent a week knitting as many patterns as possible, followed by a week of tagging the images obtained. As we got more motives, it became clear that the badumption was correct and that having more data was enough to reasonably train a network of neurons.

– How did the training go? How many images injected into the program?

– In terms of raw data, our dataset includes 2088 images. In itself, it is still too small, so too We use about 14,000 samples of simulated dataas well as multiple standard data augmentation strategies, such as adding noise, color variations, and lighting.

In practice, it is probably still too small to take into account the different types of wires, and Here, the use of simulation will be more and more critical since it is unlikely that we can access all available threads (data).

– How does the system learn and then create? How is this transition going?

– Our learning technique "directly translates" the image of a pattern into a regular sequence of patterns woven into mosaic. After extracting this description, the machine sequentially processes each line, one by one, to create the physical tissue.

It is important to mention that Our system does not improve what the machine can do, it simply simplifies the work since the instructions extracted from the images can be directly used to weave. Basically, we translate images into instructions, to reproduce them.

– If you were to highlight one of the benefits of this invention, what would they be?

– For casual users, this will allow the personalization of clothing by mixing several existing models. For fashion designers, this means a fastest way of manufacturing since we generate instructions that can directly produce the desired pattern, instead of calling on fabric experts who translate the designer's sketches into machine-woven code.

– We know that there are companies interested in the system. What do you have in mind for these software? Beyond the use in the companies, one day could it arrive at us?

– from one side, We want to simplify manufacturing for any designer, which will cause the appearance of more clothing patterns because the production time will be shortened. On the other hand, I hope we can eventually make these machines look like the way that occasional manufacturers use 3D printers. If such machines were accessible like these, then it would be simpler: we would not think of doing it: we would do it ourselves.