models/braai
braai
Braai is a neural network that looks at images from a sky survey and decides which flickers of light are real cosmic events and which are just faults in the picture.
What it is
Telescopes that scan the whole sky each night work by comparing a new image with an older one of the same patch. Anything that has changed shows up as a residual spot, and each spot becomes an alert. Most alerts are not real. They come from bad pixels, cosmic rays striking the detector, satellite trails, poor subtraction around bright stars, or the blur of the atmosphere. Braai is a trained image classifier: a program that has been shown many small picture cutouts already labelled real or bogus by people, and has learned the visual patterns that separate the two. Given a fresh cutout, it returns a score between zero and one for how real the spot looks.
Scores are not verdicts. A survey picks a threshold, and alerts scoring above it pass through to the next stage. Raise the threshold and fewer junk alerts survive, but genuine faint events are lost too. Lower it and the reverse happens. Braai is usually one early step in a longer filter, with ordinary rules about brightness, position and repeat detections doing the rest.
Why scientists use it
A wide survey can produce more alerts in one night than any group of astronomers could ever look at, and the great majority are worthless. Something has to throw away the junk before a human, or a robotic telescope, spends attention on it. Braai stands in for the practised eye of a person flicking through cutouts and saying yes, no, no, maybe. It does that work in the instant an alert arrives, which is what makes automatic follow-up possible: a supernova can be caught while it is still brightening, rather than days later.
The same shape of problem turns up far from astronomy. Wherever an instrument generates far more candidate signals than can be checked, and where humans can label examples but not keep pace with the flow, a trained classifier is used as a first sieve. The science is in what survives the sieve.
In the record
Limits
Braai only judges how a spot looks. It cannot say what an event is, how far away it lies, or whether it is interesting. Having learned from past labelled images, it reflects the habits and mistakes of the people who labelled them, and it is weakest on the unusual cases that matter most. Its score is a confidence, not a fact, and any threshold trades lost real events against admitted junk. Change the camera, the sky conditions or the image processing, and performance can quietly drift.
Written by AIxSci: the background from general knowledge, the examples from checked records.