Bug Hunt

C-Day plus 34

For over a week now, the name of the game has been bug hunting. I had a finished first draft since the last blog post, so I’ve been trying to get the output of my custom SSS filter to match the proprietary version with sample data. One issue that took a couple days to track down was a simple but erroneous switch of two angles in a complex spherical coordinate gradient to Cartesian coordinate gradient transformation matrix. I can’t say that this is a new class of obstacles – spherical coordinates have thrown wrench after wrench into my code since different mathematicians regularly define these coordinates in different ways. (Is it just me, or is having seven separately accepted conventions for the spherical coordinate system a bit absurd?) My project crosses a couple domains of mathematics, so wrestling with these different conventions has helped me deeply appreciate the other mathematical concepts that do have a single accepted formulation.

Regardless, weeding out the spherical coordinate issue and a menagerie of other bugs has left me with a filter that produces filtered data that is similar to (but not exactly matching) the proprietary code (see some example output below). Luckily, I do have several checkpoints in the filter’s processing chain and I know the problem is between the last checkpoint and the final output. My mentors have been fantastic so far, and we have a potential bead on the last issue; the weak magnetic signals produced by the brain are measured with two major classes of MEG pickup coils: magnetometers and gradiometers. In a very simple sense, one measures the magnetic field while the other measures the spatial derivative of the magnetic field, and (because of this difference) they provide readings on very different scales that I have yet to normalize. Given some luck, this last patch could fix the issue and yield a working solution to the first half of my GSoC project! (Knock on wood.)

Exemplar data showing raw unfiltered MEG signal and the same data after the benchmark SSS filter and my own custom filtering (top). Difference between benchmark and my custom implementation (bottom). The filter in progress is close, but not quite the same as the benchmark implying there remains some bugs to fix.