Difference: 2018-03-14 ( vs. 1)

ZTF ML 2018-03-14

Minutes

Attendees: Emir, Ragnhild, Frank, Matthew Graham, Richard Walters, Ashish, Brian Bue, Vaidehi Varma, Nadia, Charlotte Ward

Format of the ML day (Thu) following ZTF meeting: Hameetman, 10:30 - 5 Morning: lightening talks (10+5) Afternoon: tutorial from AMPEL, hackathon - smaller groups (ML techniques)

ZTF update: not on sky much of last week wrong filter (g) in on one day, and r was assumed. There were 2 million alerts In operations it will be automatic Wrong references used during subtraction. The data will be excised.

Survey start on Friday

Ragnhild: SNe questions: what time-scales RB scores will be included Umaa: classifier not updated since Jan because recent data made it worse (due to mix of filters) Still needing more bogus labels. Cron job set up for auto pickup by ML. Only two nights of data that came through Get Ragnhild in the loop to ensure that we are getting all labels we can

Emir: Can help with other (real-real) classifications too

Ragnhild: Status of SG classification? Adam redelivered new PanStarrs sources with correct SG scores 10 days ago Should be ready in 3-4 days. Frank to convey when done

Update to schema: 3 closest associations mentioned with Pan-STARRS classes everything in 30 seconds will be too much But we could add 4th and 5th closest etc. Ragnhild: Ariel looked at this from extra-galactic cases. In 90% cases clsoest in 30-arcsec is the best fit.

zooniverse: 1000 sources, 1 week old added. Includes real bogus (RB < 0.1), real real (RB>0.9), and intermediate (0.4<RB<0.6). As always, help with classifications needed.

Charlotte’s slides: what features contributed to false classification in Random Forest comparing interpolated values with real, and values wrt the entire set.

Brian: noticed negative values near some edges Frank: artifacts seen near edges that are PSF like - could be ghosting or glints. Matthew: Talk to QA team Fremling had done a bright star experiemnt. Matthew to chase that. Nadia: Brad was working on ghosting Umaa/Charlotte will check if there are real sources in that region Frank: for other such near-edge sources we have looked at, there were no (real) sources there. Some artifacts are elongated along x (charge-bleed?)

Roadmap for RB: Are classes separable after the stringent filtering? Not enough training data yet to decide on that. Q about data access alert packets - can they be queried Nadia: Kowalski will have packets, and can be queried along with Pan-STARRS etc.

Brian: CNNs evaluating RB on ZTF data can be the next step (extrapolating from work on iPTF data)

-- AshishMahabal - 17 Mar 2018