• refactor the way the min and maximum for the B and S parameters is estimated.
• simplify the plotting functions:: they do too many things

MVPA pipeline: problems left to fix

• run subject 06 that has one run with some missing conditions. This messes up the partition of the data for one fold.
• figure out why there is abnormally high classification accuracy even when running on noise data

Originally posted by @Remi-Gau in #10 (comment)

For the contrast between 2 conditions C1 and C2 (in a given ROI), here are the steps:

a) take median value across vertices for C1 and C2 at a given layer L (for each run R and each subject S)
b) compute the difference D between C1 and C2 (for each L, R, S)
b) take mean value of D across runs (for each L, S)
c) take mean value across subject (for each L)

Wouldn't it be better when contrasting conditions to do b) before a): in other words do within vertex contrast before averaging across contrasts?

I think that if we were using mean across vertices this would not make a difference, but because we take the median this could in principle change things (because you are in fact then contrasting the value between different vertices).

I honestly don't think the difference will be huge (we have many vertices and the median is fairly robust estimator) but just to make sure we do things in clean fashion, it would be better, no?

2 X 2 anova

For example in audio regions:

• V,T
• ipsi, contra

Compare three families

• S(A,V,T)
• I(A,V,T)
• I & S one family for the auditory re. visual area

Question is whether

1. we equate priors across families (i.e. then models from a family with many members have a lower prior) or
2. equate priors over models (i.e. then models have all the same prior likelihood, but different families have different priors)

In spm_compare_family, 1 is the default

Broken functions

Litteraly will not run:

mvpa/vol/MVPA_vol_grp_avg.m

Remi-Gau/eCobidas#206

• for BOLD profiles
• for MVPA

• mvpa/vol/MVPA_vol_grp_avg.m

AVT_analysis/pcm/surf/Extract_wht_data_for_PCM_profile_stim_surf.m

can extract data that have gone through the RSA toolboxMVNN or not (raw) but when extracting the raw data for stimuli and targe, a different number of vertex was extracted for stimuli and targets making it impossible to concatenate the 2 datasets (stimuli and targets).

This will likely require to reextract the data directly from the VTK files.

• easy: run with matlab using the code from AV-att
• better: run with R

• implement a step wise approach

• could highlight consistency in posterior PT

with python ?

• https://www.google.com/search?channel=fs&client=ubuntu&q=pyuthon+package+vizualize+vtk
• https://docs.pyvista.org/examples/index.html
• https://brainvisa.info/web/download.html
• https://brainder.org/research/brain-for-blender/

failure: 'tabulate' undefined near line 14 column 76
  IdentifyRunsToRemove:14 (/tmp/oct-V8pL7j/subfun/utils/IdentifyRunsToRemove.m)
  test_IdentifyRunsToRemove>test_IdentifyRunsToRemoveBasic:15 (/github/workspace/tests/test_IdentifyRunsToRemove.m)

From here:
https://github.com/Remi-Gau/AVT_analysis/pull/10/checks?check_run_id=1438933751

• BOLD profiles
• Rasters
  • generate figures
• PCM
• MVPA
• surface plots of S parameters

make it add only some of the subfolders of the dependency (for example CPP_SPM)