Attendees
| Attendees | Regrets | Notified |
|---|---|---|
Mark Lacy |
Agenda/Notes
I discussed this paper: https://ui.adsabs.harvard.edu/abs/2023MNRAS.518.3407D/abstract that uses several ML techniques to find and classify objects in (simulated) ALMA cubes.
Topics going forwards:
Identifiable problems with VLASS - can we speed speed up SE imaging? What would be needed to do it?
- would need to identify what we wanted in terms of output images (e.g. point source compts vs extended flux) and benchmarking metrics.
- might be able to train model with the psfs alone.
- unsupervised learning could help with more complicated source shapes.
- Analysis realm (e.g. targeting point source lists) might be more approachable (especially initially).
- Targeting a specific post-processing stage might be good, gridding sped up by GPUs, next high nail is the deconvolution.
- Brian Kirk is trying to revive the POLISH code, may not be very useable though.
- Going from visibiities to images is hard (why most algorithms start with dirty images), but machine learning could be used.
Ryan Loomis and Brian M. thinking about calibration pipelines.
How do we move algorithms from development to production?