Upgrade to Pro — share decks privately, control downloads, hide ads and more …

データサイエンス勉強会 20190425/meeting on datascience at ...

データサイエンス勉強会 20190425/meeting on datascience at RIKEN 20190425

データサイエンスの勉強会@理研 2019-04-27 の発表資料.

Avatar for Ryou Ohsawa

Ryou Ohsawa

April 25, 2019
Tweet

More Decks by Ryou Ohsawa

Other Decks in Science

Transcript

  1. Outline Brief self Introduction Topics 1. 2. 3. Finding needles

    in haystacks. Low-rank and sparse matrices decomposition Lightcurve and Gaussian process
  2. Tomo-e Gozen (トモエゴゼン) A 105-cm Schmidt Telescope & a wide-field

    mosaic CMOS camera Monitoring 20 sq-deg sky (up to) at 2 Hz Wide-area survey for transients and moving objects Kojima, et al. 2018. 10709:107091T; Sako, et al. 2018. 10702:107020J 2K×1K CMOS image sensor by Canon moon
  3. Data in optical and infrared astronomy (ground-based) Data are dominated

    by noise (low signal-to-noise ratio). Noise statistics is well approximated by a normal distribution. Signals are highly localized.
  4. Tomo-e Gozen Supernova Survey Tomo-e Gozen is originally developed to

    detect explosions in the Universe. Take a 6-second video, move the telescope, take a 6-second video, and .... Patrol the entire visible sky at least twice in a night. ⇉ this data are useful for a blind survey of moving objects For the details of NEO detection system, refer to the master thesis by Yuto Kojima Image by Stellarium
  5. Tomo-e Gozen Supernova Survey Tomo-e Gozen is originally developed to

    detect explosions in the Universe. Take a 6-second video, move the telescope, take a 6-second video, and .... Patrol the entire visible sky at least twice in a night. ⇉ this data are useful for a blind survey of moving objects For the details of NEO detection system, refer to the master thesis by Yuto Kojima Image by Stellarium Near-Earth Object
  6. Tomo-e Gozen Supernova Survey 1st frame 12th frame Binary Mask

    Image Moving Object detected Raw Data 0.5sec x 12frames SExtractor 1.5σ Frame Extraction Dark Subtraction Flat Fielding WCS Mapping AND NOT Masked 1st frame kd-tree Masked 12th frame X X SExtractor SExtractor tracklets For the details of NEO detection system, refer to the master thesis by Yuto Kojima Near-Earth Object
  7. Tomo-e Gozen Supernova Survey 1st frame 12th frame Binary Mask

    Image Raw Data 0.5sec x 12frames SExtractor 1.5σ Frame Extraction Dark Subtraction Flat Fielding WCS Mapping AND NOT Masked 1st frame kd-tree Masked 12th frame X X SExtractor SExtractor tracklets NEOs GEO objects LEO objects Garbage Cosmic Rays For the details of NEO detection system, refer to the master thesis by Yuto Kojima Near-Earth Object
  8. Tomo-e Gozen Supernova Survey 1st frame 12th frame Binary Mask

    Image Raw Data 0.5sec x 12frames SExtractor 1.5σ Frame Extraction Dark Subtraction Flat Fielding WCS Mapping AND NOT Masked 1st frame kd-tree Masked 12th frame X X SExtractor SExtractor tracklets NEOs GEO objects LEO objects Garbage Cosmic Rays 1000 : 200000 moving objects bogus detections For the details of NEO detection system, refer to the master thesis by Yuto Kojima Near-Earth Object
  9. Tomo-e Gozen Supernova Survey 1st frame 12th frame Binary Mask

    Image Raw Data 0.5sec x 12frames SExtractor 1.5σ Frame Extraction Dark Subtraction Flat Fielding WCS Mapping AND NOT Masked 1st frame kd-tree Masked 12th frame X X SExtractor SExtractor tracklets NEOs GEO objects LEO objects Garbage Cosmic Rays 1000 : 200000 moving objects bogus detections 1000 : 1000 moving objects bogus detections put labels observed videos detected candidates inspect complie the characteristics into a table Characteristics Table Random Forest Discriminator Training Validate Candidates Bogus detections For the details of NEO detection system, refer to the master thesis by Yuto Kojima Near-Earth Object
  10. Low-rank and sparse matrices decomposition Movie decomposition with GoDec (low-rank,

    sparse, residual and original) Demonstrated in https:/ /kastnerkyle.github.io/posts/robust-matrix-decomposition/
  11. Application to the Tomo-e Gozen data Extract flash-like transient signals

    from 2 Hz video data. Morii, Mikio, Shiro Ikeda, Shigeyuki Sako, and Ryou Ohsawa. 2017, ApJ 835 1. flash meteor
  12. Ground-based MIR observations An important tool to investigate Orion KL

    region in MIR planet formation process; evolution and formation of solid state materials; star formation activity heavily obscured. There are couples of technical difficulties. astronomical objects are intrinsically faint; telescope and atmosphare are quite bright in the MIR; atmospheric emission is highly variable.
  13. Chopping observation pixel count signal atmosphere telescope bias noise flat

    : astronomical signal is weak : poor detector flattness Difference A-B Position A Position B
  14. Characteristics of signals Much redundancy in non-scientific signals Scientific Not

    scientific An: rapidly fluctuating / almost constant in space Tn: almost constant in time Bn: almost constant in time
  15. Scanning observation (on-the-fly) Much redundancy in non-scientific signals Scientific Not

    scientific An: rapidly fluctuating / almost constant in space Tn: almost constant in time Bn: almost constant in time Object Field-of-View V elocity V Time t 1 =0 t 2 t 3 t M =T scan observation Sn: compact and moving sources Less redundancy in scientific signals
  16. Gaussian Process GP = A generator of smooth functions. Observations

    constrain a set of functions: 持橋大地 & 大羽成征, 機械学習プロフェッショナルシリーズ『ガウス過程と機械学習』 GPy(Pythonのガウス過程用ライブラリ)の使い方 http:/ /statmodeling.hatenablog.com/entry/how-to-use-GPy obtaining a smoothed continous function inter- or extrapolate data with uncertantines Responsitivity is fainite. Time evolution should be continuous/smooth. Kernel optimization ⇨ characteristics of data.
  17. GoDec + Gaussian Process Decompose a lightcurve into components of

    different timescales Optical lightcurve of MAXI J1820+070, Ryou Ohsawa+, in prep. a smooth baseline + stochastic pulses