This story was originally published in the Data Discovery digital magazine, created in coordination with NASA Datanauts, engaging with each other and subject matter experts to solve data challenges. For this story and more like it, stay tuned to the OpenNASA blog or view the magazine here. This blog is part of an 11-part series.
Today’s blog author is Nadia Chilmonik, a Brooklyn-based artist and engineer (and former ballerina) who is interested in space programs, data science, and predictive models. She has expertise in optimization algorithms, and works at Thicket Labs on collaborative intelligence.
Giving the World Order Through Self-Organizing Maps
By Nadia Chilmonik
A self-organizing map (SOM) is a type of machine learning algorithm, more popularly known as artificial intelligence, neural network. SOMs are trained using unsupervised learning. It is useful for reducing a high-dimensional space to two-dimensional or low-dimensional representation. Essentially you can take many objects with many traits with non-discrete or discrete values and watch as they organize themselves according to those traits.
As an example, let’s imagine the traits of planets. Planets have qualities including size, size of atmosphere, distance from the sun, amount of water, and primary color. So if we used the planets in a self-organizing map, they would rearrange in order to be closest to the other planets whose traits they have in common. This is an iterative process, so the map might start with Mars and Earth classified far apart in the two-dimensional space, because Mars may seem most similar to Jupiter on a first pass, being a primarily orange/red tone. But over iterations, they may reorganize to align based on other similarities, for instance similarities in size to Earth or distance to the Sun. As a result, Mars would end up somewhere in between. That component is dictated by the competitive learning that SOMs use instead of the typical error-correcting or back-propagation.