Present: Dustin, Jim, Aris, Steve, Mayur, Paolo

Hot items:

• Conference contributions:

  • Steve will present @ Connecting the Dots/Intelligent Detector - LAL Mar 6-9 https://ctdwit2017.lal.in2p3.fr/

    • Abstract: include three main prototypes Dustin/Steve, Mayur.

    • We will try to participate to trackML RAMP. Lots of details @ https://indico.cern.ch/event/607400/   (access code 31415)

    • David R is putting together a starter kit (track generator, fitter)

    • There is also a call for help in these areas

      • Different starter kits

      • Scoring function (big thing)

      • Estimation of required training/testing statistics

      • Bits and pieces

  • ICLR Feb 17 arkiv 2-3 pages. Mayur working on it, will push it this week to our git.

  • Confirmed talk @ IML workshop @ CERN Mar 20-22 https://indico.cern.ch/event/595059/overview

  • ACAT Seattle Aug 21: https://indico.cern.ch/event/567550/

    • Submission deadline?

• Maria: potential collaboration with google-cloud

Outstanding Action Items:

• Paolo → add folks to m2660 nersc repo [sfarrell, jbk,... ]. DONE

• Computing platform:

  • CalTech GPU servers. Jean-Roch contact person if you need access. Contact JR on skype

New Action Items:

• Paolo to discuss Ramp participation with David R.

Round the Table for Status Reports:

Caltech

• Dustin: identify which pixels belong to a track, similar to what Steve does, but forcing it to point to ON pixels. This would allow to subtract track pixels from image, and recur.

  • Steve: perhaps one could subtract the probability for each given pixel that it belong to a track.

  • Paolo: are you assuming you have three layers as seeding?

    • Dustin, started by giving the model the intercept of the track.

FNAL

• Aris slides on pull calculations. Issue is how to calculate standard deviation on slope and intercept. Dustin discussed with Stefan the possibility to use variational inference to have the network calculate its errors.

LBL

• Mayur, working on ICLR abstract

• Steve, playing with 3D CNN for hit classification. Getting results as good or better than LSTMs. Accuracy ~97%. Will look into finding out/visualizing what CNN filters are doing.

Next meeting Feb 6