Prep: weekly mtg 20160216 (Byron, Steve, Matt, me) - mobeets/nullSpaceControl GitHub Wiki

Main points

Null activity changes over time

If you just view the perturbation null space activity over time (even starting later in trial count to allow for learning to supposedly be complete), it's clear that things are changing over time:

As Matt puts it, this is a broken assumption of our hypotheses. Learning is still occurring. So how to move forward? He proposes two options:

  1. Can choose filtering to insist that there's not that much change

    • i.e., drop out kinematics angle entirely, or at least choose the subset of trials you use based on where the null activity is stable

  2. Or could come up with hypotheses that account for changes over time

    • but seems best to focus on where they're not changing, and then approach the learning problem

Volitional hypotheses obeying bounds

I tried two things:

  1. relax velocity-matching assumption (i.e., minimize how close to matching velocity we are) subject to being in the latents' bounds

  2. match velocity, stay in bounds, subject to being close to the precursor activity (i.e., the mean activity in the intuitive mapping--then we could get rid of precursor activity)

histograms of the norm of predictions. L to R: observed, habitual, vol w/ 2FA, vol w/ 2FA (s=5), method 1 above:

screen shot 2016-02-11 at 12 40 14 pm

same for method 2. and here's the errors in mean, cov orient, and cov shape

errors in mean =

1.7448    0.9782    1.4461    1.1554     1.7972

cov orient =

1.7984   63.0912    5.2903   25.8683  42.2239

cov shape =

6.4684  625.6778   92.8859  191.9548 181.5490

(method 2 is the 5th one here...what i did for method 2 was to take precursor plus a solution to the cursor kinematics as long as it was already in the observed bounds. if it wasn't, i searched for the solution closest to the precursor activity that solved the kinematics.)

So basically, the max-limited volitional is as good as the volitional w/ 2FA in mean, and slightly better in variance, but not in a way that's competitive with the habitual in its variance.