Pyininpakka, Putti, Nutto, Pokku, Tirkka, Ruitintullo — neural networks imagine Marimekko patterns

I’ve often thought what would have happened if in the 1960s the founder of Marimekko, Armi Ratia had been interested in programming instead of fashion as way to create a “cultural phenomenon guiding the quality of living”.

But an equally interesting question might be how would computers see Marimekko today? Last weekend I got to judge a hackathon at Marimekko and since I’ve been itching to try out neural networks in practice, I decided to do a quick Sunday evening project to see if I could give an answer.

The name Marimekko famously came from Armi scribbling names on a piece of paper (Armi was taken already — Mari is an anagram!). Her husband listed garments — skirt, shirt, dress — and thus Marimekko, which literally translates as Mary’s Dress was born.

Armi was also very poetic about the language Marimekko uses, writing in an early business plan for a bank loan that “Marimekko is the forest path between the practical Venla and the ethereal Anna (characters from Aleksis Kivi’s immortal story ‘Seven Brothers’) projected into the highways and byways, homes and whole living environment of this changing world”

So what if we ask a computer to dream up new names of Marimekko patterns?

Turns out, they do pretty well. I had a dataset of 1400 Marimekko pattern names and found a librabry by Andrej Karpathy’s called char-rnn. It’s a Multi-layer Recurrent Neural Network for character-level language models (a mouthful!). Within a few hours I had a list of pretty whimsical names. Some of my favourite RNN created Mari patterns:

Pyininpakka, Putti, Nutto, Pokku, Tirkka and Ruitintullo.

Most of the original pattern names are real Finnish words, but with a playful swing (“Helmipitsi, Helmisvyö, Hemppa, Hennika”). The computer keeps this non-sensical whimsicality really well (”Pokkapeita, “Pukukka”) and I think there is an interesting Karelian beat to the names as well. The first round resembled Estonian: Kulveania, Valos, Inso, Päränlö, but the computer pretty quickly started to get a grasp of Finnish (short) words (Sukka”, “Sumu”). Some Japanese influence too (”Hisakiro, Fumokare”)!

I’ll keep trying a few more things with the data — it would be awesome to combine the patterns into this and see if computers could come up with Marimekko inspired names and patterns themselves.

Thanks Marimekko & Junction for the inspiration of getting this done. I woke up this morning at 4 AM to check my RNN progress and it felt like Christmas meets Tamagotchi.


1. Collection of data

The data was collected from Marimekko website and consists of 1400 individual pattern names. You can download it here. It’s pretty clean, but obviously there could be more examples — this is a very small dataset for a neural network to work with.

2. Training

RNN (Recurrent Neural Networks) is a class of artificial neural networks. Andrej’s library takes a text file as input learns to predict the next character in the sequence. The char-rnn library makes life very easy (I think the most painful part of the entire process was updating Xcode!). I used the standard setting with a higher dropout value (0.7) since the dataset was so small.

3. Next steps

  • I’ll keep fiddling with the temperature and do a few more rounds of experiments.

I like shiny things and software. Childrens book author at Co-founder of @Codecademy alumni.

I like shiny things and software. Childrens book author at Co-founder of @Codecademy alumni.