Information decomposition in complex systems via machine learning

Significance A defining characteristic of complex systems is an abundance of variation at one scale of observation that contains, hidden within, information about organization at another scale. To see the forest through the trees is a challenge faced whether studying society or a sandpile, climate, or a brain. We present a fully general and practical methodology, rigorously grounded in information theory, that surfaces important information out of a sea of variation in a comprehensible manner. At its core is the concept of lossy compression: Some information in a measurement is preserved, and the rest is discarded. We use machine learning to lossily compress tens to hundreds of measurements simultaneously, providing a route to insight about complex systems through information decomposition.