data-driven BOSS model

The BOSS meeting got me fired up enough that on the airplane home I worked a bit more on specifying (and simplifying) Roweis and my data-driven model of the BOSS spectroscopic data (consisting of flats, arcs, and science frames). I realized that a lot of the problem comes down to sensible (fast, differentiable, and good) interpolation; this explains why such a big part of the SDSS spectro2d code (written by Schlegel and Burles) is it's B-spline core. I hope this all exists in Python in a useful form. If it doesn't, I might have to bring Burles out of retirement. I also realized that optimization might become an issue. Roweis was pushing stochastic gradient. I have to figure that out too.

1 comment:

  1. Hogg,

    Looks like I can stay out to pasture. There's a large set of b-spline libraries including 2-d bsplines in python. I don't know how much they've been optimized, as I haven't had the need to use them recently. But at least they're supported, and computers must be at least 10x faster than when sped2d was written.

    But thanks for the shoutout. I do think sped2d would have been rough without b-splines.

    Oh, and don't forget the #!