Quote of the Week, May 29, 2007

Marty Weinberg , January 26, 2006, at the opening day of the Source and Feature Detection Working Group of the SAMSI 2006 Special Semester on Astrostatistics :

You can’t think about source detection and feature detection
without thinking of what you are going to use them for. The
ultimate inference problem and source/feature detection need to
go together.

  1. vlk:

    That is indeed very good advice, but I think it is too idealistic. Very often detected sources and their properties are simply listed out in catalogs which are then made use of by other researchers who never see the data on which the catalogs are based. How they use these summarized properties is beyond the control of, and also the imaginations of, the people who did the analysis in the first place; i.e., in practice it is quite impossible to carry out source detection with any real inkling of how the detected sources will be used. This is why source detection algorithms like wavdetect try to be as general as possible.

    05-29-2007, 6:51 pm
  2. Jeff Scargle:

    I would go further than VLK and claim that the Weinberg quote contains great advice. I remember many times when I have gone into a professional statistican (let’s call him D), say at Stanford, and posed a problem that I thought was reasonable well defined, and D would ask simply “What are you trying to do?” This question almost always provided a much needed reality check. At the very least it forced me to confront the practical aspects of the problem, and sometimes made me realize I was really chasing an abstract chimera that was in fact neither well defined nor useful.

    From another point of view, VLK’s example actually reinforces this view. Constructing a catalog is an example of Weinberg’s “what you are going to use them for.” I think there are two broad classes of common uses: (a) generic, where as VLK emphasizes the creator of the method has limited knowledge of what the “user” is after, and (b) special purpose applications, where this is not so much of a problem. A Bayesian might use a non-informative prior in (a) but a specific prior based on actual knowledge and previous experience in (b).

    Of course, it is still immensely fun and useful to sometimes indulge in flights of abstract fancy, without even necessarily filing a flight plan ahead of time.

    06-29-2007, 11:35 am
Leave a comment