Difference: Minues20170518 (1 vs. 4)

Attendees: Ashish Mahabal, Umaa Rebbapragada, Brian Bue, Frank Masci, Matthew Graham, Tiara Hung, David Shupe, Eric Bellm, Rahul Biswas, Scott Daniel, Yutaro Tachibana, Nadia Blagdanova, Lin Yan, ??

We had an overview of many topics and we will get to the nitty-gritty of many over the next weeks (through short presentations and updates).

Highlights included understanding some areas where the flow needs to be well defined, or perhaps modified (we need a block diagram soon):

There are still a few uncertainties that have a bearing on the ML flow. In particular, after real-bogus separation is done (an ML step), the real targets will be passed on to UW for filtering, and on to science teams. In the current assumed flow, there is no way for the real candidates that turn out to be bogus to be fed back into the system. We need a mechanism for that. Right at the start, the stream from IPAC to UW may not be pure. Will need requirements for False Positive Rate (FPR) and evaluation criteria for classifiers before start.

Similarly, there is no way to determine if any of the candidates that were called bogus were actually real. That feedback may have to come from archival work later.

Owing to the much larger scale of events, there will be no human scanners - or, there will not be enough human scanners, though individual science projects may employ some (Matthew). That also implies that it will be difficult to catch wrong classifications early through follow-up and help improve the system. We may need to incorporate existing labels (from other surveys) early.

Tentative dates include: 1 July 2017 (access to IPAC products pipeline) 22 August 2017 (first light) ~mid Sep - Dec (Science Verification) 1 Jan 2018 (Science) 15 April 2018 (Alerts out)

Some simulated data exists (equivalent to ~30 nights, ZOGY image subtracted). These could be used for early tests and keeping the software engineering part ready (Frank). Engineering data can also be used once available. Access to ZOGY parameters, and postage stamps will be needed.

A GalaxyZoo like interface could be set up for initial vetting. A UW undergrad can set things up (Eric).

Students at Caltech are available to help with some of the machine learning. A good way to utilize that help would be through the GalaxyZOO like interface, or by improving classifications on existing data/simulations that can then be used as models (Uli, and others in Stockholm/Munich doing some work on those lines - Nadia/Lin).

Filtering discussion is related but separate (Eric). Link to Maria’s IVOA talk: http://wiki.ivoa.net/internal/IVOA/InterOpMay2017-TD/patterson-alert-filter-ivoa.pdf

===========================

Next meeting: Thu 1 Jun 2017 Cahill #370 2 PM Ashish will not be present in person. Will try to join remotely

https://global.gotomeeting.com/join/489371021

Dia in: United States: +1 (571) 317-3112

Access Code: 489-371-021

===========================

-- AshishMahabal - 19 May 2017