Zoobot

What it is

For years, members of the public have looked at pictures of galaxies and answered simple questions about them: is this one smooth or patchy, does it have spiral arms, is there a bar across the middle. Those answers, gathered in the Galaxy Zoo project, form a very large labelled picture collection. Zoobot is the name given to neural networks, programs that learn patterns in images by adjusting millions of internal numbers until their guesses match the labels, trained on that collection. The result is a model that has already learned what astronomical images look like: faint smudges, bright cores, noisy backgrounds, overlapping objects.

Once trained, the model is rarely used as it stands. Researchers take the trained numbers as a starting point and continue training on their own, usually much smaller, set of pictures. This is called fine-tuning. The early layers of such a network, which detect edges, blobs and textures, tend to carry over well, so the new task needs far fewer labelled examples than training from nothing would.

Why scientists use it

Labelling astronomical images by hand is slow, and modern telescopes produce far more pictures than anyone can inspect. A model that starts with some sense of what galaxies look like reaches useful accuracy on a new question with a small fraction of the labels, and tends to be steadier when the new set is small or uneven. It also saves the cost of training a large network from scratch each time.

Pre-training on related pictures matters because the alternative, starting from models trained on everyday photographs of cats, cars and faces, teaches a network about a very different kind of image. Astronomical pictures are faint, grey and dominated by noise. Starting closer to the target domain is a general tactic in machine learning, used wherever labels are scarce and raw data is plentiful, from medical scans to recordings of animal calls.

In the record

Limits

The model knows only the kinds of pictures and questions it was trained on. Point it at images with a different resolution, field of view, wavelength or noise character and its advantage can shrink or vanish. It also inherits whatever the volunteers' answers contained, including their disagreements and their tendency to miss faint or unusual objects, so its confident outputs can quietly repeat those gaps rather than correct them.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Model family

• Convolutional neural network

Every article