Deep learning relies on finding meaningful representations of data. These representations are particularly important for galaxy morphology tasks, where complex images are difficult to interpret directly. We argue that the recently-created Galaxy Zoo DECaLS model, trained to accurately answer every Galaxy Zoo question simultaneously, has learned a meaningful representation of morphology that is useful for new tasks. We exploit this to provide several open-source tools for investigating large galaxy samples. These are aimed at researchers hoping to exploit deep learning approaches for their own challenges but without the capacity for citizen-science-scale labeling.

These tools are; a similarity search web interface, to identify galaxies of similar morphology to a query galaxy; an active learning anomaly detection algorithm (extending astronomaly), to identify the most interesting anomalies to a particular researcher; and a morphology transfer learning Python package, to build classifiers from only a few hundred labelled examples.

We develop and demonstrate the performance of our tools using 911,442 galaxies imaged by DECaLS. This includes producing the first large-scale catalogue of ring galaxies, identified using transfer learning and 212 examples tagged by volunteers on the Galaxy Zoo forum.