Skip to content

2019_10_11

Jump to bottom
Gabriel Perdue edited this page Oct 11, 2019 · 1 revision

pre-meeting agenda

  • Abstract for "3rd ICFA Mini-Workshop on Machine Learning for Particle Accelerators"
  • Data products status
  • Data infrastructure status
  • Algorithm development
  • Board deployment status
  • ML algs on FPGAs status
  • Bonus: GitHub projects...

meeting

  • used computer found
    • purchase request for memory
    • Brian will run off this rig
    • lots of memory required to run the software (64 GB)
    • to build the firmware for the dev kit
  • ICFA workshop
    • PNNL is willing to put one in for RL
    • We should send someone (Jason looking for funding)
  • Timelines going out and linac is awake
    • Will be up and running in a week or two
    • May not start delivering beam as soon as able
    • ping Brian about zero gain runs
  • Datasets
    • I need to follow up with Shreyas about FERRY onboarding
    • one prebuilt set from Rachael
      • mysterious lag?
      • Rachael lags one of the variables intentionally
    • looking at correlations across variables
    • ran Rachael's MLP
    • when doing PCA can get 97% capturing with 5 variables
      • some are very correlated, and should be removed
    • some strong bifurcations?
    • are we using the timeline events
    • training a surrogate model to mimic the data
    • TLG in the surrogate model?
  • Algorithm development
    • LSTMs
    • surrogate model development
    • are actions discretized or continuous?
  • ML algs on FPGAs
    • Christian has connected with CERN student for HLS4ML
    • Runs with reasonable latency on Intel, but not quite as fast as Xilinx (50 clocks instead of 10)
    • maybe look at LSTMs (talk to Malachi)
    • another intermediate step - scale up one model (as many parameters as possible)
  • R&D work is needed to store weights in a register
  • Re-flashing is a few seconds, building takes much longer
  • Multiple cards
    • Ride shotgun and learn what works
Clone this wiki locally