High Performance Unsupervised Learning in Numerical Simulations and Satellite Imaging

Unsupervised machine-learning (UML) techniques play an increasingly crucial role in data analytics: clustering algorithms, in particular, are widely adopted in astronomy. We present our porting of a UML algorithm, Advanced Density Peak (ADP), which couples extreme sensitivity to small signal-to-noise and limited run- time, from the original Python code to an optimized OpenMP–parallelized C code. We discuss the preliminary results obtained by applying ADP to two astrophysical problems: (i) finding substructures in cosmological numerical simulations and (ii) deblending satellite images. Finally, we present the preliminary work in the devel- opment of a fully distributed memory parallelization for ADP using MPI+OpenMP, which is of general interest in this field because of the constant increase in the dataset sizes that makes it impossible to fit the data from cutting-edge experiments on a single node.