Modern radio telescopes generate millions of images of distant galaxies every night — giant star systems bound by gravity. Manually sorting such a volume of data is impossible. In the first half of the 20th century, Edwin Hubble laid the foundations of morphological classification by dividing galaxies according to their visual shape, but today machine learning algorithms are used for high-speed processing. Their main challenge is the need for huge training sets labeled by experts.

The STRADAViT method takes a fundamentally different approach, reminiscent of how a child learns to recognize a cat by seeing it in different poses and conditions. The algorithm is shown the same galaxy images subjected to random rotations and noise, so-called augmentations. Instead of memorizing specific numerical parameters such as brightness (the flux of radiation in certain spectral bands), the system is forced to capture invariant structural features: the bends of spiral arms, relativistic jets, and the distribution of cosmic dust clouds — solid microparticles that scatter and absorb light.

STRADAViT has demonstrated higher accuracy compared to many existing alternatives while requiring less computational effort. The method is invariant to instrumental characteristics, allowing it to be applied to data from different radio telescopes. This paves the way for automated monitoring of the evolution of billions of galaxies over cosmological timescales. One of the most impressive objects discovered with the help of such algorithms is the giant radio galaxy J1420-0545: its extent reaches 16 million light-years. If it were visible to the naked eye, it would cover an area on the celestial sphere comparable to five lunar disks. Thus, modern methods continue the work that Karl Jansky accidentally began in 1931 when he first detected cosmic radio emission.