FIRST IDEAS TO CONNECT ASTRONOMICAL DATA, DEEP LEARNING AND IMAGE ANALYSIS

Transcription

1 FIRST IDEAS TO CONNECT ASTRONOMICAL DATA, DEEP LEARNING AND IMAGE ANALYSIS Germán Gómez-Vargas FONDECYT Postdoctoral Fellow at Pontificia Universidad Católica de Chile Accelerating the search of dark matter with machine learning Lorentz Center Leiden January 2018

2 Roadmap Big data in astronomy What is the problem? Astronomical data:the Fermi Large Area telescope Dark matter-induced gamma radiation The Fermi-LAT catalogs of point (and small extended) sources dark matter clumps vs pulsars Semi-supervised learning Future: data-driven classification of objects

3 Big data in astronomy #WhatIsDarkMatter? New technologies have allowed the development of large-scale astronomical surveys Large Synoptic Survey Telescope (LSST) Start operations in 2022 Cerro Pachón, Antofagasta region, Chile 8.4 m telescope 3.2 Gpixel camera Hundreds of PetaBytes in images Tens of PetaBytes in catalogs LSST science goals include: What is dark matter, how is it distributed, and how do its properties affect the formation of stars, galaxies, and larger structures? Cherenkov Telescope Array (CTA) Start operations in ~2020 Paranal, Antofagasta region, Chile / Canary Island, Spain. 99 telescopes in Chile / 19 telescopes in Spain Expected ~100 PetaBytes by 2030 CTA study themes include: What is the nature of dark matter? How is it distributed?

4 What is the problem? We need/want to automatically: Discriminate relevant phenomena from noise and artifacts Classify objects into known categories Detect newly or unobserved phenomena Do all these in a efficient and scalable way

5 ASTRONOMICAL DATA: THE FERMI LAT Onboard the Fermi Gamma-ray Space Telescope. Launched June 11, 2008 The Fermi LAT collects high energy gamma rays (~20 MeV to > 300 GeV) with a large effective area (~6200 cm2) and a large field of view (2.4 sr)

6 GAMMA RAYS FROM DARK MATTER Annihilation

7 PREDICTION OF ALL-SKY DM- INDUCED GAMMA-RAY EMISSION Dark matter clumps Simulation by Pieri et.al. Phys.Rev. D83 (2011)

8 THE GAMMA-RAY SKY Fermi-LAT, P8data, 9 years, energies > 1 GeV

9 3FGL catalog of point sources aaaaaaaaaa Fermi-LAT 4 years Source Catalog extracted from arxiv:

10 3FGL catalog of point sources Dark matter aaaaaaaaaa clumps? Fermi-LAT 4 years Source Catalog extracted from arxiv:

11 Unidentify sources Pulsar Unidentify Active Galaxies Others 3FGL (4 years) 182 3% 6% FL8Y (8 years) 245 3% 4% 57% % % %

12 Dark matter vs. pulsars Mirabal et.al. Astrophys.J. 825 (2016) no.1, 69 Millisecond pulsar Dark matter Spectral shape from dark matter annihilation (purple line) of a 30 GeV particle into bottom quark pairs (Fornengo et al. 2004). Other Standard Model annihilation channels are expected to produce similar spectra. Also shown is a representative millisecond pulsar spectrum (black dashed line) from the second Fermi pulsar catalog (Abdo et al. 2013).

13 Methods to match DM clumps with UND sources I N-body simulation Fermi-LAT performance Predict detectability/distribution of DM clumps Compare to Fermi-LAT catalogs How many unidentify sources match my DM predictions? Some recent papers using this method: Schoonenberg et al JCAP05(2016)028, Hooper and Witte JCAP04(2017)018, Calore et.al. Phys. Rev. D 96, (2017)

14 Methods to match DM clumps with UND sources II To use supervised machine learning methods to label unidentify sources in the catalog. Some recent papers: Mirabal et.al. Astrophys.J. 825 (2016) no.1, 69, Salveti et.al, MNRAS 470 (2017) no.2, , SazParkison et.al. Astrophys.J. 820 (2016) no.1, 8, Chiaro et.al. MNRAS 462 (2016) no.3, The only one dealing with dark matter vs pulsars is Mirabal et.al The other works focus on classify the unidentify sources. They use classical methods including Decision Trees, Support Vector Machines, a Logistic Regression (LR) model, various modified versions of LR ( e.g., Boosted LR, logistic decision trees), RF, as well as some combination of methods ( e.g., a two-step method involving decision trees followed by LR). In multi-wavelength observations of tagged UNS sources as pulsars some have been confirmed Salveti et.al. MNRAS 470 (2017) no.1, Orange: source classification provided by Chiaro et al. (2016) and Saz Parkinson et al. (2016). Red: Salveti et.al. (2017) classification of unassociated sources classified as likely AGN. Here UCS are unassociated sources that are not classified as PSR or AGN candidates. 3FGLzoo from Salveti et.al. 2017

15 However Textbook machine learning

16 However Textbook machine learning Variability Variability AGN AGN UND PSR Index PSR Index

17 However Most machine learning algorithms do not suitably handle noise or measurement errors Realistic astronomical data often quite noisy, and with errors.

18 However Unknown data is often statistically distinct from training data (e.g. fainter, serendipitous) Realistic astronomical data Pre-labeled objects are often biased toward easy to observe (bright and/or nearby) objects

19 Human bias, for instance the pulsar spectra Pulsars detected in gamma rays, counterpart in radio -> timing information. Light curves and spectral shapes of a bunch of pulsars Pulsar mission models tweaked to reproduce data Some UND sources may be pulsars without energy cutoff Scheme of a neutron star magnetosphere with the internal emission regions highlighted: polar cap model is in green, outer gap model in dark blue, and slot gap model in red. Figure from Caraveo, 2014

20 Key insight Predict data from labels to create a data-driven generative model & treat label prediction as a least squares inference problem. Model labeled: train & test Label x30 30x30 15 = (1-f src ) + f src + 15x x True label Inferred f src Analyzing γ-rays of the Galactic Center with Deep Learning Sascha Caron, GAG-V, Luc Hendriks, Roberto Ruiz de Austri arxiv: Predicted label 20

21 Semi-supervised Learning

22 Semi-supervised Learning The goal is to use both labelled and unlabelled data to build better learners, than using each one alone. Is motivated by real world scenarios: Abundant unlabelled data High labelling costs Xiaojin Zhu Semi-Supervised Learning Tutorial ICML 2007 Unsupervised Semi-Supervised Data learning w/o labelled learning w labelled

23 Semi-supervised Learning The goal is to use both labelled and unlabelled data to build better learners, than using each one alone. Is motivated by real world scenarios: Abundant unlabelled data High labelling costs Supervised Semi-Supervised Labelled data O and + Unlabelled data Xiaojin Zhu Semi-Supervised Learning Tutorial ICML 2007 learning w/o unlabelled learning w unlabelled

24 Semi-supervised Learning Different algorithms based on different assumptions on the data Xiaojin Zhu Semi-Supervised Learning Tutorial ICML 2007

25 Graph-based semi-supervised learning Assumption: A graph is given on the labeled and unlabelled data. Instances connected by heavy edge tend to have the same label. Core steeps: Construct a connectivity graph using all the data (nodes labelled and unlabelled data) Propagate label information through the graph (edges: similarity weights computed from features) A good graph should reflect our prior knowledge about the domain. (Zhu 2005) Xiaojin Zhu Semi-Supervised Learning Tutorial ICML 2007

26 A random-walk interpretation Xiaojin Zhu Semi-Supervised Learning Tutorial ICML 2007

27 Semi-supervised learning on the 3FGL and FL8Y Collaborators: Roberto Muñoz MetricArts Prof. Andreas Reisenegger Astrophysics Institute PUC Chile Prof. Karim Pichara Computer Science Department PUC Chile IACS Harvard

28 Semi-supervised learning on the 3FGL and FL8Y AGN FL8Y AGN PSR PSR UND UND SNR+ SNR+ SFG+ SFG+ We apply a method called Label propagation to the Fermi Catalog using many of the columns (features) to propagate PSR and AGN labels to the UND sources. For 3FGL same features as in Saz Parkinson For FL8Y all features in the catalog.

29 Semi-supervised learning on the 3FGL and FL8Y AGN FL8Y AGN PSR PSR Pulsar as pulsar Overall accuracy 3FGL 95% 95.0% FL8Y 99.5% 97.7% 3FGL Saz Parkinson % (RF) 98% (LR) 96.7% (RF) 94.7% (LR)

30 Future The way we put labels to objects is based on modelization of nature that can be wrong. Let s think a machine taking the data and creating categories, then we scientist working on making sense of the different categories.

31 Summary Astronomical data is noisy, with errors and unlabelled data is different from labelled. A way to go around these issues is to predict data from labels to create a data-driven generative model & treat label prediction as a least squares inference problem. Another way is to use hybrid models, as semi-supervised methods. Other hybrid methods to discuss this week: GAN Generative Adversarial Networks Transfer learning

32 For further discussion Are we too naïve in the modelization of dark matter? Maybe it is time to let the data tell us how the dark side looks like, to use machine learning methods would be the way Simulation of dark matter The Hungry Lion Throws Itself on the Antelope Simulation of dark matter Pieri et.al. Phys.Rev. D83 (2011) Henry Rousseau (1905), maximum representative of Naïve art

33 Backup slides

34 Pulsar interlude ( A pulsar is a highly magnetised, rotating neutron star or white dwarf, that emit electromagnetic radiation Dame Susan Jocelyn Bell Burnell, (born 15 July 1943). As a postgraduate student, she discovered the first radio pulsars while studying and advised by her thesis supervisor Antony Hewish, for which Hewish shared the Nobel Prize in Physics with astronomer Martin Ryle.

35 The Crab Gamma rays Nebula Pulsar

36 Crab pulsation, 30/sec.

37 Fermi-LAT Collaboration Astrophysical Journal Supplement 208, 17 (2013) Fermi pulsar catalog Crab-like pulsars Millisecond pulsars are extremely hard to detect in gamma rays, specially if there is no timing information (radio loud)

38 Fermi-LAT Collaboration Astrophysical Journal Supplement 208, 17 (2013) Artistic top-side view of the Milky Way Crab pulsar All pulsars detected seem to be in the nearby

39 It is expected to have a large population of pulsars in the Galactic bulge. ) Actual MW image Collective emission of pulsars in the Galactic Center can be confused with a posible Galactic halo dark matter-induced gammaray emission. O. Macias+arXiv: , R. Bartels+arXiv: