Multilayer perceptrons in astronomy and astrophysics

How it is used

Astronomy and astrophysics study objects far beyond reach: stars, galaxies, exploding stars and the matter between them. Almost everything is learnt indirectly, from light collected over time and split into colours. A multilayer perceptron is the simplest neural network: a stack of layers in which every number in one layer feeds into every number in the next, with no built-in assumption that the input is an image, a sequence or anything else. It takes a fixed list of numbers in and returns a fixed list out. In this field those inputs are usually measurements already reduced to a tidy list: brightnesses in several filters, properties of a detection, or the settings of a physical model.

Two uses dominate. In the first, the network maps observations onto a label or a physical quantity: is this new point of light an exploding star or a flickering galaxy centre, and what mass or distance best explains the measurements. In the second, it is trained on the output of a slow physical calculation so that it can reproduce that output quickly, a stand-in sometimes called an emulator. Training data are often simulated rather than observed, because the true answers are known by construction. Perceptrons also appear as small components inside larger models, reading out a summary that some other part of the network has produced.

Why researchers use it

The pull is practical. Much astronomical analysis involves running a physical model many times, adjusting its settings until the predicted light matches the measured light. Each run can be slow, and surveys now produce far more objects than that approach can handle one by one. A perceptron trained once on a grid of model runs can then answer in a fraction of the time, which makes the fitting feasible. Where the task is classification, the volume of alerts from modern sky surveys exceeds what people can inspect by eye, and a network trained on past human decisions can triage the stream and flag what deserves a telescope's attention.

In the record

astronomyNeural network picks bright exploding stars from sky survey alerts and books telescope timeA labelled training set of survey alerts was assembled across four source classes, from confirmed bright transients to candidates human scanners had declined to save.astronomyHubble and Webb images reveal a candidate massive touching binary star in WLMSynthetic light curves were generated per filter for two model grids, one contact binary and one detached, by randomly sampling binary parameters, then normalised alongside the parameters.astronomyNeural networks stand in for slow galaxy clustering calculations in cosmology analysesGalaxy power spectrum components were computed for samples of five cosmological parameters drawn by Latin hypercube, with inputs and targets rescaled for training.astronomyNeural network reads simulated quasar light curves to estimate black hole propertiesAn MLP estimated physical parameters at the end of a latent stochastic differential equation model trained on simulated six-band quasar light curves.

Limits

A multilayer perceptron assumes nothing about its input, so it learns any structure, such as the ordering of a light curve or the layout of an image, only from examples, and often less well than a model built for that shape. Trained on simulations, it inherits their errors: where the simulated sky differs from the real one, the answers are confidently wrong. It gives little sense of its own uncertainty, and little account of why it decided as it did. Pushed beyond the range of parameters it was trained on, it extrapolates without warning.

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