Difference between revisions of "ML Tracking at JLab Introductory Ad hoc Meeting 2019-02-07"

From GlueXWiki
Jump to: navigation, search
(Agenda)
 
Line 87: Line 87:
 
# [[Media:ML_TRACKING.pdf|Hall-B Status]] (Gagik)
 
# [[Media:ML_TRACKING.pdf|Hall-B Status]] (Gagik)
 
# [[Media:20190205_ML_projects.pdf|Hall-D Status]] (David)
 
# [[Media:20190205_ML_projects.pdf|Hall-D Status]] (David)
# JLEIC Status (Markus/Dmitry)
+
# [[Media:20190207_JLEIC_ML.pdf|JLEIC Status]] (Markus/Dmitry)

Latest revision as of 13:47, 7 February 2019

Time: Feb. 7, 13:00

Location: CC F326-327


Chip's Introductory e-mail

Colleagues,

As you are all aware, there is a strong and growing interest in machine learning at Jefferson Lab. IT Division (with your help) is organizing a mini-workshop on February 12 as part of the Computing Round Table series of events, and a major topic for that one-day event will be a session on ML at the lab (this is part of a more extended multi-day activity encompassing other computing activities). What we hope to come out of this session, including the preparation and follow-up, is a clear description of the opportunities and benefits of ML for the lab, and a coherent set of steps to realize those benefits. This multi-year plan will be presented to lab management for prioritization and funding (as needed).

This email is intended to reach everyone at the lab who is now involved in ML or who has already become convinced that it is a technology with strong advantages in solving a known computing problem. If you see that I've left someone off of this list, please let me know. The mini-workshop itself is open, but for now I need to engage the people driving / needing ML.

I will be organizing this ML session, and I am requesting your assistance, especially since the time is short. I would like to organize of order half a dozen short presentations (to later evolve into short whitepapers) each covering

   * the opportunity (problem definition)
   * why ML looks beneficial
   * time scale of the development and potential deployment
   * resources needed (including the proponent's time, hardware, additional staff desired)
   * synergy with other ML activities at the lab (if known)
   * impact if successful (science, cost savings, labor savings...)

As I have been coming up to speed on the scope of activities at the lab, I am aware of the following topics under development or consideration (for some my exposure is only in passing)

   1. accelerator status monitoring and control
   2. detector status / data quality monitoring and control
   3. physics event classification / track reconstruction
   4. detector simulation (GEANT)
   5. medical imaging

Some of these obviously span multiple people and groups, such as detector data analysis. For those, some self organization ahead of Feb 12 and a common presentation and later whitepaper would be beneficial.

We will have an hour and 45 minutes for the session, and I would like 80 minutes of presentation (of length ranging from (talk+discussion) 5+3 minutes to 10+5 minutes. For the remaining 25 minutes I would like to have an interactive discussion on building a common path forward to accelerate ML at the lab. I will be suggesting a few leading questions for that discussion in a subsequent email.

Please reply (preferably not reply/all) to let me know (1) if you are interested in participating in shaping ML at JLab in this way and are willing to help in preparing presentations and whitepapers; (2) if you are available to participate in the early afternoon on Feb 12; and if yes to either of these then (3) to which topic you would like to contribute (above list or other).

We won't have time for a dozen individual presentations, so if you are able to self organize by topics, that would be wonderful. I'll contact common-topic sets of people by email and perhaps we can have a short meeting or an email exchange within each subgroup before Feb 12 to assist in organizing for this event.

I look forward to learning more from all of you, and to the future impacts of ML at the lab.

thanks,

Chip

Chip's e-mail On ML Tracking

David, Gagik, Dmitry and Yulia,

I would like to have track reconstruction be one of the top focus areas for this ML activity. It has the potential to impact the two largest halls, plus EIC, touch the largest number of people, and thus make a big impact overall for the laboratory.

If the four of you could work some this week, that would help for the session on the 12th, but don't feel you have to get all the way there. We'll have a couple of weeks beyond the 12th to polish things up.

Possible presentation, with as many presenters as you like for a total of

   15-20 minutes of presentation
   10 minutes of Q/A and discussion in the room on just this sub-topic
   overall problem statement, for a generic JLab medium energy detector (5-6 minutes)
       what data goes into training (from a generic point of view)
       how will inferences be used
   hall specifics (up to 3?  maybe 4 if CLAS is separate from CLAS-12?)  (5-8 minutes total)
       current and future data -- how to coerce into what is needed (if you know)
       specific impacts on your halls science (data rates, CPU requirements, ...)
       any other specializations you might want to cover
   common approach (intent only, work in progress)  (5-8 minutes)
       how to divide up the work, or attack multiple ideas at once
       commitment to down select if multiple explorations
       time frame to get to the point of trial training and use
       how can IT division help as part of this collaborative approach
       what role can JLab collaborators play?  what role for non-Physics university collaborators?

Feel free to tackle this ad-hoc. If you organize a meeting, please feel free to invite me -- I'll come if I can. If you need any help from me, please ask! Reply/all to this email to hold an asynchronous discussion as you like.

regards,

Chip

Agenda

  1. Hall-B Status (Gagik)
  2. Hall-D Status (David)
  3. JLEIC Status (Markus/Dmitry)