“Rotation-invariant convolutional neural networks for galaxy morphology prediction,” by Sander Dieleman, Kyle Willett and Joni Dambre, was submitted to arXiv in March 2015 and accepted by Monthly Notices of the Royal Astronomical Society. It applied a convolutional neural network to a problem astronomers had long handled by eye: classifying the shapes of galaxies as spirals, ellipticals, mergers and so on.
The training labels came from Galaxy Zoo, a citizen-science project in which volunteers classified images of galaxies online. Even crowdsourcing does not scale to the hundreds of millions of galaxies that modern surveys image, so the authors built a network to learn the volunteers’ collective judgments. Galaxies have no preferred orientation in an image, so the model was designed to be rotation-invariant - it processed several rotated and flipped copies of each image so that its answer did not depend on which way the galaxy happened to be turned. On images where Galaxy Zoo volunteers strongly agreed, the model reproduced their consensus with greater than 99 percent accuracy for most questions, and the design won the Galaxy Challenge competition on Kaggle.
The paper became an influential template for using deep learning to triage astronomical surveys: let the network handle the confident, easy cases and forward only the ambiguous ones to human experts. The authors explicitly noted this approach would be needed for the flood of data expected from future surveys such as the LSST.