The Galaxy Zoo catalogues have been used to train many different automatic classifiers, inspired by the need to classify the large datasets provided by missions such as Euclid and facilities such as LSST. The most successful have made use of modern deep learning methods, in particular convolutional neural nets. However, such methods only provide a single scalar classification for each galaxy, which fails to account for the varying difficulty of each image. I will present the result of a novel application of a Bayesian convolutional neural network to provide predictions with uncertainty. Measuring uncertainty facilitates understanding the performance of the network and how it might be improved, but more importantly allows us to select the most informative images for citizen scientist classification. This in turn allows the development and deployment of a hybrid system which classifies new surveys more accurately than either humans or machines alone. We will present the performance of such a system with images from DeCALS, and explore how it allows us to scale citizen science classification while still retaining the potential for serendipitous discoveries of unexpected objects such as the Galaxy Zoo 'Peas'. ...
|Published - 1 Jan 2019