AI Notes

Notes from canSAS 2024 AI/ML in SAS Topical Presentation/Discussion

• Talk
  • AI is pervasive
    • The most impactful AI applications will be outside he facility wall, these mostly demand good, reduced data
  • Decision ready data
    • bespoke GUI reduction is unnecessarily user-unfriendly, AND prevents future data-intensive work
  • Accessible documented API
    • Magic cable story
    • We should more regularly support outside decision engines as the 'center of the universe' with the instrument as a subordinate worker

• Discussion
  • Andrew Jackson
    • Accessible documented API is good, public internet is not going to happen. All instruments should be able to be ‘plugged into’
  • Brian Pauw
    • Has solution for rolling nexus/NxCansas into scicat
  • Jan Ilavsky
    • Instrument control from open internet is not going to happen; decision is in wrong hands
    • Why not just use EPICS?
      • A: We’ve done that.
    • People hate tiled/bluesky
  • Unknown person
    • How do you make sure the AI doesn’t break your instrument or put the instrument in a bad state?
      • A: This is a problem with the instrument. If the instrument can be put in that state, it’s the instruments problem
  • Unknown Person
    • How to preserve privacy of data for use in ML/AI applications?
      • A: Good important question, no easy answers
  • Adrian Rennie
    • Comment about “center of the universe”. Should be broader. All instruments should be switchable between these two views
    • How to get the users integrated into the decision making?
      • A: Human-machine interfaces is important. Code and efforts exists (Tsuchinoko)
  • Brian Pauw
    • We should support and push for instruments that aren’t the center of the universe
  • Tanny Chavez
    • Uncertainty quantification?
      • A: Yes, very important