Abstract: The advent of machine and deep learning algorithms on petascale supercomputers is accelerating the pace of discovery in astrophysics and poses significant challenges to the interpretability of these deep neural networks. We present a novel visualization of a deep neural network output during training as it is learning to classify galaxy images as either spiral or elliptical. The network is trained using labeled datasets from the citizen science campaign, Galaxy Zoo, adopted by the Sloan Digital Sky Survey. These trained neural network models can then be used to classify galaxies in the Dark Energy Survey that overlap the footprint of both surveys. Visualizing a reduced representation of the network output, projected into 3-D parameter space, reveals how the network has discovered two distinct clusters of features which allows it to classify galaxies into two groups. These visualizations of the neural network during training aid in the interpretability of the black box of deep learning and reveal how the network responds to the input images at various stages of training. Finally, it allows a wider net to be cast to a general audience, thereby generating interest in and visibility to an otherwise highly specialized field.
Video Summary: pdf
Video Link: video
Back to the Visualization & Data Analytics Showcase Archive Listing