covRatio in DAP.pdf doesn't match what Pete's code makes

mobeets / nullSpaceControl

1 stars 0 forks source link

covRatio in DAP.pdf doesn't match what Pete's code makes #42

Closed mobeets closed 8 years ago

mobeets commented 8 years ago

Maybe a data issue? Did he use a different date?

mobeets commented 8 years ago

determinant = product of eigenvalues.

mobeets commented 8 years ago

well, it's true that the baseline and minimal hypothesis have very tiny eigenvalues, so the determinant just gets super tiny.

but just looking at the covariance matrices for a given kinematic condition, it's clear minimal and baseline are way off. from top left, clockwise: observed, habitual, minimal, baseline:

covhyps

mobeets commented 8 years ago

http://bmcevolbiol.biomedcentral.com/articles/10.1186/1471-2148-12-222

mobeets commented 8 years ago

http://www.itl.nist.gov/div898/software/dataplot/refman1/auxillar/ks2samp.htm

mobeets commented 8 years ago

project eigenvectors onto other data and see how variance is explained.

mobeets commented 8 years ago

From "A simple procedure for the comparison of covariance matrices": You get two scores, each describing the error in terms of mismatch in shape and mismatch in orientation.

This describes the ability of one cov's eigenvector to explain the data of the other, and vice versa, relative to how well the cov's eigenvector describes itself.

Big strength I think is in interpretation: just as we have a error metric for the mean, where 0 is the best, now we have an error metric for the covariance, where 0 is the best.

mobeets commented 8 years ago

Okay, I've gone back to Pete's original code as pulled from the repo. It took a few adjustments to get right:

adding scripts like simpleKalmanWhatever, subCellArray, computeAngle, etc. which I had on hand.
data_directory needed to be set globally so i added a few lines to load it in the right places
apparent error using subCellArray fixed by a normal indexing

So there was only the one bug, with subCellArray, which I fixed.

But alas! Minimal hypothesis still has a covError of around 10e-3.

I also noticed some code computing a volitionalBasis other than just the row space's intuitive mapping. And some minimal/baseline hypotheses fit in the latent space rather than neural space.

mobeets commented 8 years ago

Given that a better covError seems appropriate anyway, I'm just going to mark this closed for now and suggest I use the new one.