Training the Trap

Antonia Rowlinson

Transcript

1. Training
   
   the
   
   TraP
   Antonia Rowlinson
   
   University of Amsterdam
   

   View Slide

2. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
   Swinbank (2011)
   

   View Slide

3. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
   Swinbank (2011)
   

   View Slide

4. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
   Swinbank (2011)
   

   View Slide

5. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
   High Bound Noise Threshold:
   Radio Sky Monitor dataset
   
   average RMS noise ~ 24 mJy
   
   (11 min snapshot images)
   

   View Slide

6. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
   Quality Control:
   Script to create Quality Control plots is available to TraP
   commissioners here (new version in development):
   
   !
   !
   This script needs running on each new dataset to
   determine the optimum thresholds for the flux ratio high
   bound, ellipticity and minimum A-Team separation
   https://github.com/transientskp/scripts/tree/master/TraP_QC_plots
   

   View Slide

7. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
   Transient search parameters
   After each new data point is inserted into the database we
   calculate:
   
   Variability Index:
   
   !
   !
   !
   Weighted Chi-squared:
   
   !
   !
   !
   !
   

   View Slide

8. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
   Transient Search Parameters:
   All newly detected sources are inserted into transient list
   
   However, the list is dominated by sources hovering at/near
   the source extraction detection threshold
   
   Will be resolved in Release 2 of the TraP
   
   Script providing temporary (approximate) fix available here:
   https://github.com/transientskp/scripts/tree/master/TraP_sort_transients
   

   View Slide

9. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
   2.7 sigma
   threshold
   

   View Slide

10. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    

    View Slide

11. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Bright sources -
    statistical flux
    errors very small,
    systematic errors
    dominating
    

    View Slide

12. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Faint sources - likely
    imaging artefacts
    Bright sources -
    statistical flux
    errors very small,
    systematic errors
    dominating
    

    View Slide

13. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Dataset overview
    Script to create these plots is available to TraP commissioners
    here:
    
    !
    !
    Also, this script identifies transient candidates for a given
    sigma threshold:
    https://github.com/transientskp/scripts/tree/master/TraP_source_overview
    https://github.com/transientskp/scripts/tree/master/TraP_anomaly_detection
    

    View Slide

14. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Simulated Datasets:
    Simulation technique enhanced and standardised
    
    Gives a set of known stable sources and known transients
    
    Further information available on LOFAR wiki:
    Take existing
    measurement set
    Delete contents
    of data columns
    Use BBS to insert sources
    using a sky model
    Insert Gaussian noise
    Calibrate and image using
    standard pipeline
    http://www.lofar.org/operations/doku.php?id=tkp:simulateddata
    

    View Slide

15. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Simulated datasets
    

    View Slide

16. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    

    View Slide

17. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Transients which were undetected at quiescent flux but
    are stable after they turn on.
    
    !
    N.B. if maximum flux is comparable to detection
    threshold, they are undistinguishable from faint stable
    sources (this will be rectified in release 2 of TraP)
    
    !
    

    View Slide

18. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    

    View Slide

19. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Can we do better?
    Logistic Regression
    (a linear classifier)
    
    4 features for each source:
    
    !
    !
    Trained using simulated transients
    and ‘stable’ sources from the Radio
    Sky Monitor dataset
    

    View Slide

20. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Precision ~ 99%
    (probability that
    identified
    transients are
    real)
    
    Recall ~ 80%
    (probability that
    all transients are
    identified)
    

    View Slide

21. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    

    View Slide

22. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Conclusions:
    TraP works!
    
    However transient list dominated by fake “new” sources
    close to source detection threshold - to be resolved in TraP
    Release 2 (temporary fix available)
    
    Choose quality control parameters and transient thresholds
    very carefully to avoid spurious transients
    
    Tools to help available on GitHub (frequently improved)
    
    Transients can be identified in several ways such as:
    
    Identify anomalous transient parameters: ideally using a sigma
    threshold (~2.7 sigma for the Radio Sky Monitor dataset)
    
    Classify using multiple features and logistic regression (Preliminary)
    

    View Slide

23. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Min Separation from A-Team:
    Significant
    deviation starts
    from ~10 degrees
    

    View Slide

24. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Restoring Beam Ellipticity:
    

    View Slide

25. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    Current caveat:
    
    The sky is not well known
    at LOFAR frequencies
    
    Extrapolating fluxes to
    LOFAR bands using poorly
    constrained spectra
    

    View Slide

26. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    

    View Slide

27. Antonia Rowlinson University of Amsterdam b.a.rowlinson@uva.nl
    

    View Slide