Using real images of galaxies from astronomical surveys of our Universe, a neural network can be trained to classify these with impressive accuracy. This visualization shows the output of the penultimate layer of a deep neural network during training as it is learning to classify two types of galaxies: spirals and ellipticals. Training is shown from epoch 0 to epoch 10 as the network learns a distinct feature representation for each class. In the final visualization, we project the higher-dimensional feature manifold, which the network has discovered, onto a 3D parameter space in which the galaxy classes form two distinct, orthogonal clusters.

Science: Asad Khan, Eliu Huerta, Sibo Wang, and Robert Gruendl, National Center for Supercomputing Applications, University of Illinois at Urbana-Champaign; Elise Jennings and Huihuo Zheng, Argonne National Laboratory

Visualization: Janet Knowles, Joseph A. Insley, and Silvio Rizzi, Argonne National Laboratory