Prep: weekly mtg 20160330 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

New hypothesis

Sources of information about 8-d null space activity:

• Dependence on theta

• Dependence on row space activity

• Correlations amongst the [8x1] vectors

Note: We can actually drop all but three dimensions to maintain most of the hypotheses' scores (using the SVD from intuitive activity in the perturbation block). Only covErrorOrient seems even slightly affected, and then only minorly.

Here's my current approach:

1. Fit a model on the intuitive data in the perturbation null space that predicts the 1st column of null space activity given the two columns of row space activity.
2. Fit a model that now predicts the 2nd column, given the row activity and the 1st column. And so on...
3. Now just use this model to predict null space columns in the perturbation block.

Note that this doesn't use thetas at all, and yet it looks pretty good! Not only that, but there's no sampling either: The variability comes solely from the row space activity.

Also, I've clearly been ignoring the raw cloud hypothesis for some reason? (This one also doesn't use thetas.)

Errors by cursor-to-target angle

(Note: reveals difference between otherwise-similar hab and cloud-hab hypotheses.)

Learning and Error

• implemented Patrick's learning metrics: Lbest, Lmax, lrn (not yet validated)
• for cloud-hab: more learning --> less mean error

(axes flipped so that more positive always means more learning)

rsq =

    0.5075    6.1823    0.0474    0.5056
    0.0312    0.1934    0.6755    1.0650
    0.0595    0.3796    0.5605    0.0838
    0.1102    0.7428    0.4219    0.6043
    0.5782    8.2258    0.0285    0.0796

where 1st col = rsq, 3rd col = p-val of F-test

and the trends are the same for habitual and cloud-hab

IME plots

• compare to normal decoder: for two days, very similar observed null activity

First shown them hypsByTrials and how the ones that increase with time are the same ones that in the right plot below (the non-IME model scores) have higher mean errors.

Now show them hypsByTrials_ime and how the better score also results in errors not going up as much over time.

DataHigh

• comparing distributions

Behavior, error, and CCA over time

• plots