Machine-Learning Methods

Finding the rare objects that do not fit any known pattern — the outliers that may be the next new class of transient or a genuinely unexpected discovery. As surveys outgrow human inspection, anomaly detection is how the strangest objects get noticed at all.

Assigning objects to discrete categories from their measured features — star or galaxy, supernova type, real detection or artefact. It is the workhorse of survey astronomy, sorting the millions of objects a modern survey finds into kinds a human never could review one by one.

Grouping objects by similarity without being told the categories in advance — letting the structure in the data reveal itself. It is used to find natural classes of objects and to organise catalogues too large to label by hand.

Large models pre-trained on broad data that can then be adapted to many specific tasks with little extra training. In astronomy they are an emerging approach to building a single model that understands images, spectra, or light curves across many problems at once.

Predicting a continuous quantity from an object's features — a galaxy's redshift, a star's temperature or age — rather than a discrete label. Regression turns cheap, plentiful measurements into estimates of quantities that would otherwise need expensive follow-up.

Learning a compact, informative summary — an 'embedding' — of a complex object like an image or a light curve, so that similar objects sit near each other. Good representations make every downstream task, from classification to search, easier and more accurate.

Training a model on unlabelled data by inventing a task it can grade itself on — predicting a hidden part of an image, or telling two views of the same object apart. It is powerful in astronomy, where raw data is abundant but expert labels are scarce.