Astronomers search distant galaxies for giant star-forming clumps — cradles of stars that were furiously born billions of years ago. Manually sifting through hundreds of thousands of images is like a mushroom forager combing an endless forest. But machine learning gives the forager-program an atlas of examples.

First, thousands of volunteers labeled galaxies — this is the ABCs. Then the neural network examined every patch of sky in six bands — from ultraviolet to infrared. Ultraviolet, by the way, added discoveries: young stars shine there like tiny lanterns. To weed out mistakes, scientists sprinkled in computer-generated fakes — and the program spotted them.

The result: one and a half million candidates in 710,000 galaxies. By evaluating the reddening of light (galaxies are racing apart), heavy-element impurities, and interstellar gas, scientists will reconstruct galaxy growth. Meanwhile, splitting light will tell the stellar life story.