{"id":444072,"date":"2017-12-06T12:30:13","date_gmt":"2017-12-06T20:30:13","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-project&p=444072"},"modified":"2019-05-22T10:12:45","modified_gmt":"2019-05-22T17:12:45","slug":"skies-epitomized","status":"publish","type":"msr-project","link":"https:\/\/www.microsoft.com\/en-us\/research\/project\/skies-epitomized\/","title":{"rendered":"The skies epitomized"},"content":{"rendered":"\t\t\t
\n\t\t\t
\n\t\t\t\t\t
\n\t\t

Art + Machine Learning showing how we see the sky.<\/em><\/h3>

The Skies Epitomized<\/em> is a series of artworks exploring the essence of the sky from the perspective of humans gazing at it. The works were created in collaboration between Maja Petri\u0107<\/a>, an artist, and Neboj\u0161a Joji\u0107<\/a>, a machine learning researcher. The artwork is derived from big data<\/em> through a machine learning algorithm<\/em> that is used to create visual summarizations (epitomes) of sky images people posted on the Internet.<\/p>

Exhibited: December 2015 \u2013 March 2016<\/em><\/p>

Maja Petri\u0107 explains the motivation behind the art: I find it fascinating to see a visualization of how we see the world. Neboj\u0161a Joji\u0107\u2019s epitome algorithm actually paints the picture of the sky in the eyes of the Internet community, which includes most of us. I believe that the image of \u201cthe sky in our eyes\u201d can unlock understanding about human relationship towards our origin, and our surroundings.<\/em><\/p>

Neboj\u0161a Joji\u0107\u2019s epitome model is a generative model of images that assumes that all images of interest come from a large panorama in which individual pixels can change slightly. The individuals are small cutouts with small local perturbations. The panorama is much larger than the images in the set, but much smaller than a tiling of all images would be. The panorama is thus referred to as an epitome of the set, as it captures the variation seen in the data in a miniature form, but also expands on individual images by placing them in the context of other images. The learning algorithm iteratively maps each of the images from the set to the best matching locations in the epitome and then re-estimates the panorama and the variation parameters until it achieves convergence.<\/p>

<\/div>