Agenda And Minutes for 11/25 10am CST¶
1. Dial Toll-Free Number: 866-740-1260 (U.S. & Canada)
2. International participants dial:
Toll Number: 303-248-0285
Or International Toll-Free Number: http://www.readytalk.com/intl
3. Enter your 7-digit access code, followed by “#”
Your access code: 7691719
- Star/Galaxy Separation Extravaganza!
- Simple star/gal from Eli A Modest Proposal for Preliminary Star/Galaxy Separation
- Maayane's neural network
- Ryan Keisler's Random Forest
- The S-G separation Challenge?
- Translating Mangle Masks Into Data Units (Eli)
- MAG_MODEL tests via Chris Miller
Attending: Anne, Chris L., Gary, Nacho, Eli, Dave B., Keith B., Brian Y, Josh F., Angelo, John Fisher, Rachel, David Finley, Emmanuel, Diego, Maayane, Ryan, Mike J., Alex D-W, Felipe M., others...
1. We discussed the Star/Galaxy Extravaganza! William commented on Eli's Modest Proposal that using multiple bands (is this a star or galaxy in 2 or 3 of 4 bands?) also helps. Maayane discussed her Multi-Class code which is now running. It is simple to play around with different bands and parameters for a given training set. Right now she's using COSMOS + HST. The code works in 2 stages: first a PCA finds the most relevant super-parameters, then the neural network runs. The output is a value between 0-1 which can be interpreted as a probability.
A question arose whether this formalism could/should incorporate quasars (3 class) and perhaps also image artifacts (4 class). Quasars in DES full depth imaging are a particularly interesting lot because the host galaxy should be visible out to z~1. Mike Jarvis asked if more information could be gotten from the single epoch (SE) images. Eli was a bit skeptical because the SE images aren't as deep and the real challenges are at the deep end; however, SE images should be very useful in rooting out artifacts. Nevertheless, there may be additional info to be gleaned from (esp.) the best seeing SE images.
Felipe M. asks if Maayane's code will be integrated into DESDM production. Certainly we should have a star/galaxy classifier built into the production as a baseline for comparison.
Ryan discussed his Random Forest classifier. Much of the discussion revolved around definitional issues of purity/completeness/etc, which gets back to one of Maayane's and Nacho's main comments:
2. The Star/Galaxy separation challenge. Nacho will lead this. (Thanks, Nacho!) One of the first orders of business is to agree on a set of metrics. Maayane has added her preferred metric to the The S-G separation Challenge page.
We also discussed systematics checks: do the numbers of stars selected by the various s/g classifiers correlated with Galactic latitude (expected), seeing (expected, but is this problematic), or depth?
A brief interlude on using SPREAD_MODEL or MAG_PSF - MAG_MODEL as a S/G classifier. According to Emmanuel, there are a few problems with MAG_PSF - MAG_MODEL. First, MAG_MODEL isn't a fixed target and has many more free parameters. Second, MAG_MODEL can be confused in deblended regions (due to the inadequate deblender) such that the way it gets computed over all the pixels in the region may create biases. SPREAD_MODEL was developed with a fixed center and constrained set of fit parameters to avoid these biases.
3. Eli discussed translating mangle masks into depth maps for various magnitudes. Diego asked and Eli confirmed that the mangle mask does not know about SLR: only the GCM zeropoints that were input. There was a discussion of MAG_AUTO and how it handles varying seeing. Biases should be small (I think Emmanuel said a few percent?) but this is something we should test.
4. Yuanyuan briefly mentioned Chris Miller's recent tests on the disk-only MODEL magnitudes for cluster galaxies. And so far the colors look good in spite of using the wrong model (Since these galaxies are mostly bulges). This is good news!