highly parameterized models of data

As my loyal reader knows, I have no fear when it comes to models with huge numbers of parameters; indeed the ubercalibration project is effectively a fit with hundreds of millions of parameters, and we can prove that we got the global optimum (in the sense that we made sure the problem is guaranteed to be convex). Today Roweis pitched a generalization of all this, in which one creates a very flexible linear model space, where parameters are tied to or in a hierarchy of meta-data, such that some parameters are tied to, say, the date, some to the airmass, some to the seeing, some to the camera column, and so on. Then the model discovers which parameters are necessary for accurate modeling of the data and thereby discovers important meta-data, dependencies of the data on artificial issues, and bad data. We tentatively agreed to run this on the new BOSS spectra from SDSS-III.