LHO

The Ledger · Entry 146 · Science

Classifying a galaxy by its shape

Ceded to machinesSince 2015Criterion practiceLast reviewed 2026-08-17

Event

The winning entry of the Galaxy Challenge, published in March 2015 by Dieleman, Willett and Dambre, reproduced the Galaxy Zoo volunteer classifications with better than 99 percent accuracy on the images the volunteers agreed about.

Hubble sorted galaxies by eye in the 1920s, and the classes he drew are still the ones in use. When the surveys grew past what astronomers could look at, the work was handed not to machines but to the public: Galaxy Zoo collected more than forty million classifications from about a hundred thousand people from 2007. That crowd is what trained the networks that then replaced it, which is the whole shape of this transfer. What was ceded is the routine sorting of ordinary galaxies at survey scale. What was not is the object that fits no class, the thing Galaxy Zoo volunteers kept finding by accident, and which a network trained on agreed labels is built to place in a box rather than to notice.