Abstract
We present the application of Restricted Boltzmann Machines (RBMs) to thetask of astronomical image classification using a quantum annealer built byD-Wave Systems. Morphological analysis of galaxies provides criticalinformation for studying their formation and evolution across cosmic timescales. We compress the images using principal component analysis to fit arepresentation on the quantum hardware. Then, we train RBMs with discriminativeand generative algorithms, including contrastive divergence and hybridgenerative-discriminative approaches. We compare these methods to QuantumAnnealing (QA), Markov Chain Monte Carlo (MCMC) Gibbs Sampling, SimulatedAnnealing (SA) as well as machine learning algorithms like gradient boosteddecision trees. We find that RBMs implemented on D-wave hardware perform well,and that they show some classification performance advantages on smalldatasets, but they don't offer a broadly strategic advantage for this task.During this exploration, we analyzed the steps required for Boltzmann samplingwith the D-Wave 2000Q, including a study of temperature estimation, andexamined the impact of qubit noise by comparing and contrasting the originalD-Wave 2000Q to the lower-noise version recently made available. While theseanalyses ultimately had minimal impact on the performance of the RBMs, weinclude them for reference.