Hot items:

• NERSC ERCAP allocation:  got 790K hours (1.2M requested)

  • NERSC account request form -"Cross Cut HEP Tracking Algorithms". Repo name m2660.

  • Steve tried his NN using keras/tensorflow on cori II

• December Milestone: produce our first simulated tracks dataset

  • completed

• Conference contributions (poster, potentially paper):

  • Connecting the Dots/Intelligent Detector - LAL Mar 2017 (all plenary workshop)

    • Deadline Jan 9

      • Abstract not needed but useful. Volunteers to submit?

• Maria: potential collaboration with google-cloud

Outstanding Action Items:

• Pietro → introduce group activities

• Maria: I want to try to put all the agendas and the links and the minutes in a Basecamp : do you people use Basecamp? Pietro Shall we try? Others?

  • PC: Indico+google seems to work but I am open to evaluate basecamp

• Data Storage: JBK

• Computing platform:

  • CalTech GPU servers. Jean-Roch contact person if you need access.

New Action Items:

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

• Paolo → send samples to Dustin

Round the Table for Status Reports:

Caltech

• Dustin: detector as an image. Use Steve’s “2d data” as input. Extract slope/intercept.

  • Multi-track dealt with by processing n-times single track and feeding the outputs to a LSTM layer. Variable # of tracks dealt with Keras “sample_weight” layer. Still need to know a priori how many track there are in given event

    • Slides available on indico

  • Mayur: this paper on amortized inference and LSTM may be relevant https://arxiv.org/abs/1603.08575

  • Steve: intuition suggest convolutions may act as stub-finders; maps well to similar pattern recognition approaches like pattern-bank matching with AM

FNAL

• Jim: Aris interested in joining the project. Works in NOVA

LBL

• Mayur: LSTM 2d prediction (phi/z, fixed rho planes). Setup pykalman filter to use as baseline.

  • Using 9-hits trajectories to avoid double hits

• Steve:

  • Added hit classification metric, which has promising results (~97% accuracy)

  • Added track parameter (slope, intercept) prediction to hit finding NN. Too early to say if it helps stabilize predictions, but the model is clearly able to learn both outputs simultaneously.

    • Dustin: was accuracy for fitting NN calculated only after the last layer?

      • Yes (it actually uses all layers)

Next meeting Jan 9