Teaching a machine to see: unsupervised image segmentation and categorisation using growing neural gas and hierarchical clustering
We present a novel unsupervised learning approach to automatically segment and label images in astronomical surveys. Automation of this procedure will be essential as next-generation surveys enter the petabyte scale: data volumes will exceed the capability of even large crowd-sourced analyses. We demonstrate how a growing neural gas (GNG) can be used to encode the feature space of imaging data. When coupled with a technique called hierarchical clustering, imaging data can be automatically segmented and labelled by organising nodes in the GNG. The key distinction of unsupervised learning is that these labels need not be known prior to training, rather they are determined by the algorithm itself. Importantly, after training a network can be be presented with images it has never 'seen' before and provide consistent categorisation of features. As a proof-of-concept we demonstrate application on data from the Hubble Space Telescope Frontier Fields: images of clusters of galaxies containing a mixture of galaxy types that would easily be recognised and classified by a human inspector. By training the algorithm using one field (Abell 2744) and applying the result to another (MACS0416.1-2403), we show how the algorithm can cleanly separate image features that a human would associate with early and late type galaxies. We suggest that the algorithm has potential as a tool in the automatic analysis and data mining of next-generation imaging and spectral surveys, and could also find application beyond astronomy.