Deep data: discovery and visualization Application to hyperspectral ALMA imagery. Issue Volume 12:Issue S325(2016) (30th May 2017)

• Record Type: Journal Article
• Authors: Merényi, Erzsébet; Taylor, Joshua; Isella, Andrea
• Editors: Brescia, M.; Djorgovski, S.G.; Feigelson, E.; Longo, G.; Cavuoti, S.
• Is Part Of: Proceedings of the International Astronomical Union. Volume 12:Issue S325(2016)
• Page Start: 281
• Page End: 290
• Publication Date: 2017-05-30
• Subjects: neural machine learning, -- clustering, -- ALMA, -- protoplanetary disks; Astronomy -- Congresses; Astronomy -- Periodicals; 520
• DOI: 10.1017/S1743921317000175
• Languages: English
• ISSNs: 1743-9213

Abstract

Abstract: Leading-edge telescopes such as the Atacama Large Millimeter and sub-millimeter Array (ALMA), and near-future ones, are capable of imaging the same sky area at hundreds-to-thousands of frequencies with both high spectral and spatial resolution. This provides unprecedented opportunities for discovery about the spatial, kinematical and compositional structure of sources such as molecular clouds or protoplanetary disks, and more. However, in addition to enormous volume, the data also exhibit unprecedented complexity, mandating new approaches for extracting and summarizing relevant information. Traditional techniques such as examining images at selected frequencies become intractable while tools that integrate data across frequencies or pixels (like moment maps) can no longer fully exploit and visualize the rich information. We present a neural map-based machine learning approach that can handle all spectral channels simultaneously, utilizing the full depth of these data for discovery and visualization of spectrally homogeneous spatial regions (spectral clusters) that characterize distinct kinematic behaviors. We demonstrate the effectiveness on an ALMA image cube of the protoplanetary disk HD142527. The tools we collectively name "NeuroScope" are efficient for "Big Data" due to intelligent data summarization that results in significant sparsity and noise reduction. We also demonstrate a new approach to automate our clustering for fast distillation of large data cubes.