Date & Time

Date 03 Jun 2019 , NCSA-3000, 9:00 - 17:00 PT

This meeting precedes the DMLT F2F meeting

Location

Attendees

• Leanne Guy
• Zeljko Ivezic  (attending remotely)
• Melissa Graham (attending remotely)
• Simon Krughoff
• Colin Slater
• Eric Bellm
• Michael Wood-Vasey
• Jim Bosch
• Robert Gruendl
• John Swinbank
• Wil O'Mullane
• Gregory Dubois-Felsmann
• Robert Lupton  (partially and remotely from Finland)

Regrets

Pre-meeting planning

Please add ideas to the table of suggested topics

Questions to be addressed in each topic: 

• Where are we now and what do we need to do to get where we want to be?
• How are we going to verify and validate the algorithmic performance (e.g requirements and anything missing for requirements)?
• What are the "success criteria" for these algorithms?
• What, if any studies do we need to do to (e.g SST study, engage science collaborations, interact with other outside groups)?

Agenda:  room = NCSA 3000

Session Notes

(Please correct if I have misrepresented your statements)

Image differencing  & DCR

• paraphrase: Main thing is to evaluate whether we hit reqs based on AL, if not we can go onto Zogy.
  • KSK: Preconvolution isn't yet working, needed to make deconvolution work
  • Would be nice to do a bakeoff, but we should focus on getting AL successful as higher priority.
  • Eric: "Main focus is to get in a position to hit requirements as soon as possible, with just one algorithm"
• LPG: Could we do some sort of monthly-roll up, i.e. what Meredith is generating manually? Could load the monthly run? 
  • KSK: There is a templated notebook system, meant for periodic reports. Would that be useful for AP?
• Eric shows 3 Decam fields of detections, show false positive features. Patch-sized problems have been fixed recently.
  • ZI: Are these false positives a strong function of SNR? EB: Good question, don't know
  • EB: Edge flag needs to be used to throw away those sources, but just hasn't been added to the query yet.
  • DMTN-21 getting ~50 det/ccd, but current hits seems to be ~on average ~hundreds per ccd.
• RHL: Brighter Fatter? Not being corrected right now. Means the kernel is being built from bright stars and then applied to the faint stars.

DCR

• RHL: What happens if you run Ian's code on a telescope with an ADC [as a null test]? Good question, haven't tried. 
• EB: Current opsims often producing pessimal image sets with the same airmasses, but opposite sign of the parallactic angle. Gives no leverage for fitting.
• Discussion: Current DCR code doesn't handle PSFs that vary between input images into the coadd. Because it's pixel-based, the pixels don't know about their objects or their PSF.
• Haven't gotten to RB yet. "What we need to train a RB model is the same as what we need to show we're meeting our requirements, i.e. a training set". Will need a larger dataset.
• RHL: have to understand the full system before we go on sky, can't pick off bits and pieces. Need to understand non-dcr data before we understand dcr data. 
• MWV: Focus on good templates with ~12 images?
  • RHL: That's the best case, we can do that before doing the harder part of small N. Learn different things from the two cases.
• LPG: DESC is interested in joining the image differencing effort, how can we help this?
  • MWV: DESC definitely has energy on this, but are they going to inundate the AP team with questions/problems.
  • JB: "We're already under a lot of known-issues, just finding more issues doesn't necessarily help".
  • JB: DESC has a lot of decam experts, could they help getting obs_decam and calibrations in better shape, or is that too indirect for DESC to care. MWV: Software dev people are more from the French/CFHT side. 
  • Robert G: Still working on Decam issues from the DES side. Could reprocess the HITS fields with the DESDM, see if the same issues remain/is it something with the CP calibration side.
• MWV: Could produce a coherent report on the DC2 image differencing

Deblending

• SDSS deblender: Spatial symmetry constraint.
• Scarlet: Assumes each component in a source has the same morphology across bands, but different fluxes.  Includes spatial symmetry and radial flux monotonicity constraints.
• KSK: What about objects with multiple physical components, bulge+disk.
  • JB: Need a way to tell it that that is ok, but theoretically it could.
• Without constraints, failure mode is the big objects eat the flux from all their neighbors. Some concerns about what monotonicity means, implementation-wise. Still working on that.
• Currently identifying peaks and then deblending with fixed set of peaks, but doesn't necessarily have to be that way (but other ways are probably harder).
• Tried implementing some algorithms that don't work on the deblended pixels, only the direct pixels in the very central region of the peak. Good for colors for photo-z, but no hope at getting the total flux right.
• LPG: slide says not the best for all cases, examples?
  • JB: Works best for weakly blended galaxies, wrong one for stellar fields and galaxy clusters.
• Biggest cost probably comes in the per-step convolution process. Need to reduce the number of steps by having better initial guesses.
• JB: Would like to have the computational performance optimization first, too hard to do scientific optimization now when the runtime is bad.
• KSK: Mentioned earlier that if we don't need multifit then we can give that back to deblending
  • JB: And scarlet will happily eat all of that compute.
• LPG: Need to get metrics to know whether we actually fit in the compute budget, especially w.r.t. multifit.
• LPG: Has there been work on GPU?s
  • JB: Not so far, some of the other projects have been using frameworks like tensorflow to enable easy switching to GPUs. Would be super interesting for someone to try a GPU implementation.
• KSK: Could we make this pluggable, or is the interface not definable enough?
  • JB: Could be, but worried that there might still be incompatibilities.

List of SST tasks (Confluence)

Overdue or Undated DM Science Team tickets

LIT tickets of interest to DM Science