jeremymanning / supereeg Goto Github PK

switch out the numerator/denominator approach with one that uses (z-transformed) correlation matrices directly

• when new data come in, simply compute the correlation matrix and (if needed) blur it out to the indicated locations
• update: multiply each object by n_subs, then divided by the sum of n_subs. if locations are different, first blur both matrices out to the union of the locations
• get_model: compute the inverse of the fisher z-transform (_z2r)
• sub: for a - b, multiply b's correlation matrix by -1 and run a.update(b)

if numerator/denominator approach reconstructs synthetic data well, we can stick with it. otherwise we should switch to the "pyFR" approach (just average the correlation matrices across subjects)

• change rbf --> log_rbf
• save log of numerator and denominator
• full model = np.exp(numerator - denominator); also need to change everywhere there's an np.divide-- I think this has now been consolidated to one function
• add and update can be supported via np.logsumexp of the numerators and denominators
• subtract can be supported by np.logsumexp but multiplying numerator and denominator of the "to-be-added" model by -1

use existing tests, then check in a checkpoint

the handling of locations isn't correct-- need to work through the logic and update. i added some comments directly to the code.

Questions:

• Why is there a separate _fit and _predict function?
• We can see where old and new locations match by looking where the weights == 0
• Need clearer variable names to make it clear which are the old vs. new locations

Given m by m "old" correlation matrix and n by n "new" correlation matrix:

When the old and new locations match, assign the corresponding entries of new to those values in old. This can be done efficiently by creating a subset of matching (old_match_inds, new_match_inds) and not matching (new_inds) locations, and then:

• setting new[new_match_inds, ][new_match_inds] = old[old_match_inds, ][old_match_inds]. Log numerator: values from logZ. Log denominator: all zeros.
• for new[new_inds, ][new_inds] we need to blur out old (considering all locations) to the corresponding locations in new. The weights are given by log_rbf_weights[~old_match_inds, ][~new_match_inds]. Log numerator: values + weights. Log denominator: weights.

can be replaced by np.round