2023-12-11 Meeting notes

Agenda

Discussion of some ways AI/ML can help us

• artifact recognition for QA (supervised/unsupervised learning etc - Sergio Garza has already made some progress in this area for VLASS)
  • Brian Kent took us through his VLASS demo. ADASS discussing a coding/ML prize. 
  • Sergio used CNN for VLASS "primary beam hole" detection. Worked well for that, but other artifacts more difficult (also would use actual images, not PNGs next time). Need to teach the net what a "normal" image looks like. 
  • Ignatio's experience at JAO - working with a 3rd party company that can supply advice. Setup of databases is different - data lakes needed. Data quality and consistency important, many algorithms exist ready made. Initial applications include predictive maintenance and help with proposal reviewer allocation.
• RFI detection and excision
  • Difficulty with visibilities and cal table solutions might be that they are complex - Sergio - really need to map the data to positive reals and normalize it. Matthew Schiller - in defense systems, it depended, sometimes rejection on phase was better, sometimes on amplitude depending on what you were trying to find. Patrick - can also flatten arrays, AI is looking for patterns. 
• machine learning approaches to the VLASS imaging problem (and to interferometric imaging in general)
  • POLISH - Steve Myers may give it a try, Brian Kirk getting it set up at NRAO. Urvashi - unclear how it will do e.g. on wideband data. Ryan Loomis - ALMA emission morphologies different from VLASS, ESO dev study looked into some of this. Regularization techniques getting a boost from similar technologies. 
• Using AI bots (e.g. ChatGPT) to help with scripting/coding, data cleaning etc.
  • Brian Kirk made a quick help agent by training a model on the CASA manual, something similar could be part of an NRAO science platform to answer straightforward questions without needing to submit a Helpdesk ticket. Ignatio - JAO also working on ALMA technical handbook
• AI help with pipelines (Brian Kirk/Urvashi working on this)
  • Trying to predict from the data how the pipeline should run (with/without flagging etc). Marcel Neeleman is working on consequential flagging studies, may be relevant. 
• Using language models to obtain user science goals from proposal abstracts+titles to inform SRDP pipeline choices (e.g. spectral line/continuum/detection expt/imaging exit) (Adele working on related project for ALMA).
  • Ryan - may also be useful for guiding users to interesting regions of large data cubes when combined with feature identification (ngADMIT). 

Available software resources to try out (scikit-learn, tensor flow, PyTorch etc)

• • Discussion in chat about needed resources - would need some investment in GPUs, ChatGPT+ licenses etc. Ignatio - on prem. vs cloud provided, provided may be better for initial experiments. 
  • Tony R. Should focus on some specific areas for future meetings, what can help most with VLASS? Sergio and the DAs will make a list of some of the highest priority things that they think AI/ML can help with. 

AI institute status

• Brian Mason, Ryan Loomis, Eric Murphy and others worked with UT Austin and UVa on a pre-proposal led by Stella Offner, waiting on whether this will go to a full proposal, focus on ngVLA & ALMA WSU.

Next meeting in ~ 1 month