The rebuff of another readers to the comments:

http://science.slashdot.org/comments.pl?sid=353771&cid=21280237

I'll address the technical/methodological criticisms in the post. I already posted a rebuttal in the Slashdot thread, but a response is deserved here.

The first technical criticism is that we have not "carefully select[ed] a machine-learning approach to identify good ways of analyzing images". Image analysis comes first, machine learning comes second. Features are extracted from the raw image first, only then can machine learning tell us which features are predictive, and which are not. This is an iterative process: successful features are then refined and expanded, others rejected, and a new crop of features can then be extracted. Accuracy of the computer vision system is improved at each step.

The research leading up to this Grid project passed through several of these iterations. The previous iteration identified several promising families of image features. These have been refined, and expanded upon. We have also included other image analysis techniques shown in the literature to be effective on protein crystallization images. Altogether, these combined methods are those running in the HCC Grid agent. It is our fifth generation feature set. They are not arbitrary, or designed to waste CPU cycles.

The second technical criticism is that features derived from Fourier transforms are not included in our set of calculations. Fourier analysis was largely supplanted by wavelet analysis a decade ago. The Slashdot poster also claims we are doing in 5 hours what can be approximated to "~98.5%" in 5 minutes. I invite the Anonymous Coward (Slashdot lingo) to provide a reference supporting the claim.

Christian

Thanks for the response Christian. Interesting stuff.