The Cannon is a really, really big model

Andy Casey and I got our continuum-fitting working on all the individual sub-exposures that make up the APOGEE data today. We also built our own stacking code that combines these into a high signal-to-noise spectrum for each star (not that that is ever necessary for any scientific experiment—Don't Stack Your Data tm). We were able to show that our precision improves with signal-to-noise exactly as expected, which is good and expected (and not true for many physics-model-based pipelines). We launched our large suite (a grid search) of validations to set the hyper-parameters.

Over the last few days we have realized that our model has 1.5 million parameters and more than twice that many hyper-parameters. Uh oh. So we set many of them equal and got the whole thing down to just 2. Hyper-parameters, that is. All this just shows that we ought to be able to do way better with more cleverness.

Unfortunately Casey leaves NYC this weekend, and we don't have a finished product! But we got close. We are just days away from having a full set of figures, and we have a paper half-written.

In the afternoon, a few of the SCDA members recapped things they learned at NIPS 2015 this past week. The most relevant to me right now was work by Chen and Candès on solving many quadratic equations. This is equivalent to—or contains as a special case—phase retrieval. What they have done sounds like magic, so I will have to read the paper.

No comments:

Post a Comment