Bayesian deep learning addresses two common challenges for machine learning in astronomy: learning from uncertain data and making predictions with well-calibrated error bars. I present a novel Bayesian approach used for Galaxy Zoo DECaLS, a new catalog of detailed visual morphologies for galaxies imaged by the Dark Energy Camera Legacy Survey.
DECaLS images reveal spiral arms, weak bars, and tidal features not previously visible in shallower SDSS images. Such features are scientifically valuable but challenging to classify automatically. I trained an ensemble of Bayesian convolutional neural networks to predict posteriors for the detailed morphology of 314,000 galaxies. The Bayesian networks efficiently learn from uncertain (heteroskedastic) volunteer responses without the need for “clean” training samples. They can also flexibly express their own uncertainty when making predictions. Each model predicts volunteer responses for every question, learning a shared representation across questions.
The networks were trained using 7.5 million individual classifications by Galaxy Zoo volunteers. When measured against confident volunteer classifications, the trained networks are approximately 99% accurate on every question.
Below: galaxies automatically classified as most likely to be spirals with loose, medium, or tight winding, respectively by row.