LS4GAN LDRD

1 GAN LDRD history

  • Idea born from EDG discussions to go after some kind of ML LDRD.
  • Initially pitched for PO LDRD, declined.
    • Title: “Deep Augmentation”
  • Offered to EIC ML working group, picked up sPHENIX interest.
    • Jin Huang joins effort, target: jet reco / nuclear form factors, and sPHENIX TPC response.
  • Brought to CSI LDRD, accepted. (LDRD 21-029 CC/CSI)
    • Title: “Bridging the Gap between Scientific Simulations and Experiments with Cycle-Consistent Generative Models”
    • Yihui (Ray) Ren and Meifeng Lin (CSI PIs) + Haiwang, Jin, me.
  • 3 years, “type A” starting mid FY 2021.
  • One post doc in PO, one in CSI, small fraction of scientist time.
    • PO post-doc: first 2 years EDG/LArTPC, a few options for how to cover later sPHENIX-related applications.
  • initial organization: github, journal club discussion.

Final name: LS4GAN “Large Scale Scientific Simulation Systematics GAN”

2 LS4GAN idea in broad strokes

2.1 Translate simulation to real data via GAN

  • a resulting dataset becomes generally “more like real”
  • but retains specific features of each event
  • after training: \(d_i \to d'_i = \mathrm{ls4gan}(d_i)\)

2.2 Phase 1: simulation improvements

  • examine event-by-event difference \(d_i - d'_i\) for hints
  • use hints to determine reason for GAN changes, fix simulation
  • repeat until diminishing returns or needing prohibitive fixes

2.3 Phase 2: propagation of residual systematics

  • send both \(\{d_i\}\) and \(\{d'_i\}\) through downstream ML/reco/ana
  • difference of final output assigned it as overall simulation systematic

3 Existence proof: CycleGAN

3.1 :Bcolumns:

3.1.1 :BMCOL:

orig.jpg

Photograph

3.1.2 :BMCOL:

monet.jpg

Monet transfer

4 Challenges for LS4GAN

\footnotesize

4.1 Much larger images in limited GPU RAM

  • Photos: 1k × 1k → LArTPC: 10k × 10k
  • Image tiling/fragmenting possible
  • Sparse sigproc output may be exploitable.

4.2 Scientific data requires precision and accuracy

  • Photo-styled GAN output enjoys human bias to “look right”.
  • LArTPC simulated / real differences expected to be subtle
  • Can not rely on “artistic interpretation” for success.

4.3 Must work with large data sets

  • Will need to process a lot of data on large networks, but multi-PB from PDSP!
  • Conventional simulation code speed could be limitation, improvements already coming from H.Yu+CSI HEP-CCE-PPS work.
  • Likely need production HPC for training and inference. Leverage NPPS experience.

5

\Huge

\(\mathcal{FIN}\)

Author: Brett Viren

Created: 2021-01-07 Thu 12:56

Validate