Sparse Feature Networks for Interpretable Galaxy-Property Estimation from Images
A neural-network architecture, the Sparse Feature Network (SFNet), that learns a small set of sparse, human-interpretable features from galaxy image cutouts which can then be linearly combined to estimate physical properties such as optical emission-line ratios and gas-phase metallicity. SFNets match the accuracy of leading black-box models on astronomical machine-learning tasks while keeping the learned image features inspectable, helping astronomers find physical patterns in large imaging datasets.
Sparse Feature Networks for Interpretable Galaxy-Property Estimation from Images
Galaxy appearances reveal the physics of how they formed and evolved. Machine learning models can now exploit galaxies' information-rich morphologies to predict physical properties directly from imag…