Sanity checks with Uri Hasson's fMRI data

We could use fMRI data (where we have the full-brain patterns) as a sanity check for SuperEEG by pretending that some of the fMRI data were missing.

Here's what I'm thinking of for [Uri's data](https://drive.google.com/drive/folders/0Bz7Bb5yh66TndmpFcmh5V3hZWTg) (we could pick out one or two datasets to apply this to):

**Analysis 1**
1) For each subject s, use every other subject's data to estimate the full-brain correlation matrix
2) Hold out (randomly) *x*% of the voxels from subject s
3) Using the remaining (100 - *x*)% of the voxels and the correlation matrix estimated from everyone else's data, fill in data from the missing voxels using SuperEEG
4) Compare the reconstructions to the observed data (e.g. correlations between the reconstructed vs. observed activity, by voxel)

We can then make a brain map (reconstruction accuracy by voxel, averaged across subjects) for different values of *x*.

Key questions:
- What is the minimum value of x that still leads to decent reconstruction accuracy?
- Are there particular brain areas that are reconstructed especially well (or poorly)?

**Analysis 2**

Since everyone in each dataset experienced the same story, we could also use ISFC (in sliding windows, or using a new technique my lab is developing that estimates moment-by-moment ISFC matrices) to take into account dynamic correlation matrices. This would entail an analysis similar to the one above, except that:

1) We'd estimate (using everyone else's data) the moment-by-moment ISFC matrices, using our ["timecorr" method](https://github.com/ContextLab/timecorr)
2) Separately for each timepoint, we'd then reconstructing the activity for the held-out voxels using that timepoint's ISFC matrix (for different proportions, *x*, of held-out voxels)

Key questions:
- Do reconstruction accuracies improve when we take into account the dynamics of the stimulus-driven correlations?
- Does this change the minimum value of x that is needed to reliably reconstruct held-out data?

Start by reviewing Uri Hasson's linked fMRI datasets and the linked timecorr method. Define which one or two datasets and held-out voxel proportions to use, then implement both static and dynamic reconstruction analyses and compare reconstructed with observed activity. Done means reporting reconstruction accuracy, brain-area patterns, and whether dynamic ISFC improves the minimum usable voxel proportion.

Written by the indexing model from the issue text.