Back to the main page.

Bug 414 - beamformer tutorial needs a little restructiring

Reported 2011-01-18 20:21:00 +0100
Modified 2012-08-22 15:52:17 +0200
Product: FieldTrip
Component: documentation
Version: unspecified
Hardware: All
Operating System: All
Importance: P1 critical
Assigned to: Johanna
Depends on:
See also:

Robert Oostenveld - 2011-01-18 20:21:52 +0100

at the moment there is a projectnoise=yes without explanation (that follows 15 minutes further in the session) and there is a lambda introduced at a moment which is not appropriate.

Johanna - 2011-04-19 15:28:32 +0200

For now I just added 2 extra sentences where projectnoise and lambda are introduced, to point to where exactly they are discussed later on. I have some other questions/comments though for this tutorial which will wait until after this toolkit (as well as any further changes regarding this initial bug).

Johanna - 2012-01-11 14:41:49 +0100

1) Is Robert happy with the resolution of the initial purpose of the bug (projectnoise and lambda)? 2) One other questions/comments I had pertained to the use of 'common filter'. I think someone else has since added the note/warning at an appropriate place, linking to which shows some nice example scripts. I think this is sufficient for a general user, but for the purpose of a toolkit, is this annoying? Should the main tutorial still be remade to include 'common filters'? 3) Another idea I had to add would be on how to use beamforming with timedomain LCMV. Should this be included in this tutorial, or as a separate 'example'?

Robert Oostenveld - 2012-01-12 10:41:58 +0100

(In reply to comment #2) Re 1: yes, that is fine. Re 2: it is no required, I think. It should be explained that we always want to work with contrasts (i.e. ratio's). One way is a) activity divided by estimated noise b) activity divided by baseline c) activity1 divided by activity2 In case b and c you can make the H0 that the data in both conditions is the same, and that therefore the best spatial filter is constructed is on the two conditions copmbined (common). That common filter is then applied separately to each condition. Note that the alternative for "a" is to use normalized leadfields. That is something to consider to put in the tutorial. Re 3: would be nice to have it in this tutorial, but don't know whether it works well on this data. The stimuli are auditory presented words. Perhaps other data would demonstrate it better.

Johanna - 2012-01-13 14:16:13 +0100

I have added an extra sentence in the Background, Robert's summary in the section, and the option/reason to use a normalized lead field in Still todo: (1) create picture/example using normalized leadfield (2) Find/use separate dataset for example/tutorial using LCMV (and with common filter)

Jan-Mathijs Schoffelen - 2012-03-21 13:40:58 +0100

Team up with JM to get it done.

Jan-Mathijs Schoffelen - 2012-03-21 13:42:36 +0100

start with updating the test_tutorial_beamformer (which then will automatically update the figures) and fill in the tutorial text around it. move the old test file to test_tutorial_beamformer20120321

Jan-Mathijs Schoffelen - 2012-03-21 15:53:18 +0100

copied the test_tutorial_beamformer to test_tutorial_beamformer20120321

Johanna - 2012-03-27 14:47:43 +0200

Completed: 1) Restructured existing content in a more sensible order. 2) Contrast of post vs pre now uses the common filter. 3) Regularization of 5% is used for this common filter, although an exercise now shows difference (not that much actually) of using 0% or 10%. 4) Corresponding figures for post-vs-pre are updated. 5) Code updated in test_tutorial_beamformer.m (svn commit 5545). Note this code now also contains 'answers' to some of the exercises, even though that code is not present on the wiki of course. Maybe useful to print out to have when presenting a tutorial at a toolkit. Note: The normalized leadfield option is better than 'center of head' but not really a viable option (still quite a bit in central head area; much worse than the NAI result) for this particular dataset. Still to do: 1) JM said that he would use volume_reslice to re-make figures so the heads were not upside down. 2) Create new 'bug' for creating time-domain beamformer tutorial (compare LCMV and SAM, and discuss use for ERP data). Consider using different dataset.

Johanna - 2012-03-27 15:18:12 +0200

re-assigning to JM for him to do the volume_reslice part. I've created and assigned to myself bug 1393.

Johanna - 2012-03-27 16:12:13 +0200

reminder to figure out how the source output should be saved/renamed e.g. save sourceNAI sourcePost; and later on: save source_commonfilter sourcePost sourcePre

Robert Oostenveld - 2012-03-27 20:56:23 +0200

(In reply to comment #8) > Note: > The normalized leadfield option is better than 'center of head' but not really > a viable option (still quite a bit in central head area; much worse than the > NAI result) for this particular dataset. When looking at this with Cristiano last week, it seemed more that it was a relative deep blob in occipital cortex, i.e. deep but quite shifted from the apparent center-of-head in the original depth bias picture. So it might be more physiological than only depth bias. But nai indeed appears better.

Jan-Mathijs Schoffelen - 2012-03-29 11:25:15 +0200

figures updated. assigned back to Johanna to close it ;-)

Johanna - 2012-03-29 11:46:34 +0200

I updated test_tutorial_beamformer.m to show that line of code for volumereslice. Still wondering: Does it matter what the filename of the saved source output is? Especially as there are now two (NAI and common filter output)?

Johanna - 2012-04-18 16:51:33 +0200

file names for saved source output is now consistent with the test_tutorial_beamformer and the wiki, and the files have been given to Robert to copy to the ftp/data directory.