Convolutional neural networks in astronomy and astrophysics

How it is used

Astronomy produces data that already sits on a grid: images of the sky, maps of mass or temperature, and spectra, which are measurements of brightness at each wavelength of light. A convolutional neural network is a program that learns by sliding small filters across such a grid, picking up local patterns like edges and blobs, then combining them into larger ones. Nothing about the position on the grid is assumed in advance, so the same learned filter works anywhere in the frame.

In the work described here, these networks sort alerts from survey telescopes into real exploding stars and other kinds of variable source, read stellar temperature and chemical makeup from low-quality spectra, turn maps of distorted galaxy shapes into maps of cluster mass, label every cell of a simulated cosmic volume as void, wall, filament or halo, sort the shapes of fast radio bursts, and predict how a galaxy cluster would look in X-rays from a simulation containing only dark matter.

Why researchers use it

Surveys now record far more candidate objects and far more sky than people can inspect. The networks stand in for two things: a human eye that was formerly asked to look at each image, and a hand-written physical model that was formerly asked to invert the measurement. In both roles the attraction is speed, and consistency of judgement across the whole dataset.

They are also reached for when the quantity wanted is not directly measured. Cluster mass is inferred from the slight stretching of background galaxy shapes; structural labels for the cosmic web are computed from a gravitational potential; observable X-ray and pressure maps follow from gas physics absent in a dark-matter-only simulation. In each case a target can be computed cheaply in a simulation, and the network is trained to supply it from the data actually available.

In the record

Limits

These networks learn whatever is in their training material. Much of that material is simulation, or labels from one survey transferred to another, so a network can inherit the assumptions of the simulation or the limits of the reference catalogue and give confident answers outside the range it saw. Selection cuts on brightness, signal quality or object mass bound where the result holds. The networks produce a number but not, by themselves, an honest uncertainty, and they offer no account of why a given answer follows, which makes a surprising result hard to distinguish from a failure.

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

Related

• Convolutional neural network: the idea, across every field
• AI in astronomy and astrophysics

Every article