Using Neural Networks for Styling Data Visualisations

Machine learning is everywhere. It's used for self driving cars, face recognition, automated text generation and many other tasks. At webkid we create a lot of data visualisations. Unfortunately there are almost no tasks that allow the use of complex machine learning techniques like neural networks. Until now!

Semantic Style Transfer FTW

EVERYBODY loves christmas and winter stuff - so why not use neural networks to make data visualisations all christmas-y?! All charts in this post were created with Datawrapper. Unfortunately you can't choose fancy stylings out of the box but I am sure they will implement something like neural style transfer after seeing these awesome charts powered by AI!

Style transfer techniques using neural networks got very popular with photo manipulation apps like Prisma. The basic idea is as follows: You choose one picture that defines the style and another one with the content and feed them into a neural network to get a new image. With semantic style transfers the results get even better because you can define the different areas (like ground, sky, etc) in the style image and in the content image with help of annotations.
For the examples in this blog post we used neural-doodle.

Crypto Currencies

These days we hear a lot about bitcoins. They reach new record highs almost every day and consume so much electricity that power providers should consider using their energy to mine crypto currencies instead of feeding it into power grids. If you check out the progress of the bitcoin prices it already looks a lot like a mountain range. Here we present you our first piece “Crypto Mountain Landscape with Stars”:


Bitcoin Prices


Mountain
Source: https://commons.wikimedia.org/wiki/File:Sohlberg-vinternatt_1901.jpg


Crypto Mountain Landscape with Stars

German Elections

Germany just voted and we were very happy to finally be able to present the results to you. But hang on! The next election might follow shortly if the parties can't find a way to form a proper government. That would mean we'd have to create all the data visualisations all over again. As we don't want to bore our readers, we are considering coming up with some different styles this time. We searched for christmas images to theme our charts with and were not disappointed:


Election Results


XMAS Night
Source: https://pixabay.com/en/xmas-painting-xmas-background-xmas-2018061/


German Election Winter Nights

Christmas Markets

We *all* love christmas markets. It's so much fun to be there. The people are so kind and some of them even apologize as they stampede over you on the way to the glühwein booth. But where are all these magical places, people ask. We found the Berlin data on the open data portal daten.berlin.de. But a standard map is too dull and boring, especially for the younger readers. It may sound obvious but we searched for images of christmas trees to style our christmas market map. Sometimes the simple ideas are the best:


Berlin Christmas Market Map


Christmas Tree
Source: http://www.publicdomainpictures.net/view-image.php?image=61111&picture=background-snow-1


O Christmas Tree, O Christmas Map


On my MacBook the calculations took several hours so I rented a GPU. You can find a list of providers on the NVIDIA GPU Cloud Computing overview. As said above I used neural-doodle for these experiments but there are lots of other promising scripts that you can play around with:

Stay tuned and check our next post where we will discuss how to use blockchain powered Excel macros to create immersive 3D visualisations for VR/AR!


< Back to the tree



Views 12

About

Moritz Klack

Moritz is an interactive developer, journalist and neural net fanboy. While doing projects with the interactive team of the Berliner Morgenpost he co-founded webkid (a data visualisation agency with a focus on the news sector) and ResiApp (a personal news assistent for younger people).

Next year he would like to find more time to contribute to Open Source software.

snow flake
© 2017 Journocode