fields/astronomy/model families
Support vector machines in astronomy and astrophysics
How support vector machines, a method that separates labelled classes by finding the widest possible gap between them, are used in astronomy and astrophysics, and why researchers choose them.
How it is used
Astronomy studies objects that cannot be handled or repeated: stars, exploding stars, galaxies and the ripples in spacetime called gravitational waves. Almost everything is known through light or faint signals recorded by telescopes and detectors, so much of the work is sorting measurements into categories. A support vector machine is a sorting tool. Given examples already labelled by hand or by simulation, it looks for the dividing surface that leaves the largest margin, the widest clear gap, between one class and another. New measurements are then placed on one side or the other. The same idea can also predict a continuous quantity rather than a label.
In this field the inputs are usually numbers extracted from a recording: the brightness of a source in several colour bands, the shape of a light curve over time, the pixels of a small image, or the wiggle of a waveform in time. These are often compressed first, for instance by reducing many pixel values to a handful of summary components. A support vector machine can work on raw measurements too, and a trick called a kernel lets it draw curved boundaries without the researcher having to say in advance what shape the boundary should take. Training labels frequently come from physical simulations rather than from real observations.
Why researchers use it
Labelled astronomical data is scarce and expensive. Confirming what a source really is can take spectroscopy on a large telescope, or detailed simulation on a supercomputer, so a method that learns from a modest number of examples is attractive. Support vector machines care only about the examples nearest the boundary, which makes them stable when the training set is small and when the classes are unbalanced. They have few settings to tune, they train quickly, and they give reproducible answers.
In the record
Limits
A support vector machine learns a boundary, not a physical explanation, so it cannot say why a signal falls where it does. It needs features chosen by hand, and it handles raw images and long waveforms poorly compared with networks built for that purpose. Trained on simulations, it can fail on real observations whose noise and instrument quirks differ. Its output is a bare label, with no honest measure of confidence, and it slows down badly as datasets grow to survey scale.
Written by AIxSci: the background from general knowledge, the examples from checked records.